《花旗银行:2025代理型人工智能(Agentic AI):探索DIFM (Do It For Me)经济下的金融未来研究报告(英文版)(44页).pdf》由会员分享,可在线阅读,更多相关《花旗银行:2025代理型人工智能(Agentic AI):探索DIFM (Do It For Me)经济下的金融未来研究报告(英文版)(44页).pdf(44页珍藏版)》请在三个皮匠报告上搜索。
1、Citi is one of the worlds largest financial institutions,operating in all major established and emerging markets.Across these world markets,our employees conduct an ongoing multi-disciplinary conversation accessing information,analyzing data,developing insights,and formulating advice.As our premier
2、thought leadership product,Citi GPS is designed to help our readers navigate the global economys most demanding challenges and to anticipate future themes and trends in a fast-changing and interconnected world.Citi GPS accesses the best elements of our global conversation and harvests the thought le
3、adership of a wide range of senior professionals across our firm.This is not a research report and does not constitute advice on investments or a solicitations to buy or sell any financial instruments.For more information on Citi GPS,please visit our website at 2025Agentic AIFinance&the Do It For Me
4、 Economy Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 Ronit Ghose,CFA Head of Future of Finance Citi Global Insights Sophia Bantanidis Future of Finance Analyst Citi Global Insights Kaiwan Master Future of Finance Analyst Citi Global Insights Ronak S Shah Future of Finance Analyst Ci
5、ti Global Insights Savina Chahal Citi Global Insights Prag Sharma Global Head Artificial Intelligence Centre of Excellence Citi Catherine Zhai Strategic Investments Markets Citi With thanks to:Aaron Powers Hunit Ajit Tripathi Hadron Founders Club Alex Sion Blend360 Alexandra Mousavizadeh Evident Ant
6、hony Diamandakis Citi Arvind Jain Glean Biser Dimitrov Citi Brian Benedict Arcee AI Bryan Zhang Cambridge Centre for Alternative Finance Charles Zvibleman AlphaSense Chris Smith Arcee AI Christian Baker Arcee AI Christian Frahm United FinTech David Birch Consult Hyperion Dor Eligula Bridgewise Drew
7、Propson World Economic Forum Erin McCune Bain&Company Erkin Adylov Behavox Greg Ulrich Mastercard Huy Nguyen Trieu CFTE Jason Warner Poolside John Nay Norm AI Jrg Wiemer TIS Joseph Bonanno Citi Jrgen Eckel BCG X Kevin Chung Writer Kevin Levitt NVIDIA Kirsty Rutter Lloyds Banking Group Manuel Grenach
8、er Unique FinanceGPT Mark Bennewith Arteria AI Martin Goodson Evolution AI Martin Moeller Microsoft Mason Allen Reality Defender Nar Vardanyan Ntropy Nitya Sharma Simpl Nuno Sebastio Feedzai Peter Lancos eXate Philip Drury Citi Ren Ito Sakana AI Roger Zhu Bain&Company Ron Chakravarti Citi Simon Tayl
9、or Sardine Sina Wulfmeyer Unique FinanceGPT Sirisha Kadamalakalva Citi Sonal Rattan eXate Steve Blanchet Pictet Group Suresh Ramamurthi NetXD Vibhor Rastogi Citi January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 3 AGENTIC AI Finance&the Do It For Me Economy An artificial in
10、telligence that can make autonomous decisions without human intervention.It needs no prompting.And it can solve complex problems independently and proactively.Meet agentic AI.AI,and agentic AI could have a bigger impact on the economy and finance than the internet era.This paradigm shift in AI is po
11、wered by a combination of technological breakthroughs in contextual understanding,memory,and multi-tasking abilities.References to agentic AI by BigTech in corporate documents and press articles increased 17x in 2024 and we expect them to go parabolic in 2025.Agentic AI effectively turbocharges the
12、Do It For Me(DIFM)economy.In financial services,users will have their own bots or AI agents helping them choose products and execute transactions.Competition will tick-up as starts-ups grow.The nature of work could change.Those tasks that are outsourced today to contractors or third parties will be
13、increasingly done by agentic AI.In 2024,37%of VC funding was to AI startups.Thats likely going higher in 2025.Autonomous agents and digital co-workers saw the biggest growth in VC deal activity in 2024,followed by GenAI for customer support operations.The trend will likely continue as BigTech and th
14、e VC community make big moves in agentic AI in 2025 and beyond.The financial services sector is the second largest consumer of GenAI after the telecom and media sector.Historically,banks are the biggest spender on technology(outside of the tech industry).This trend will likely continue with GenAI an
15、d agentic AI too.Agentic AI is largely in an experimental phase and this report offers up insights gleaned as we looked under the hood on what is being built now for future production.We spoke to 30+AI startup founders,BigTech executives and other external experts in preparing this GPS report.It hig
16、hlights some of the key use cases of agentic AI in financial services,ranging from compliance,deepfake and fraud prevention,onboarding and KYC,wealth,credit,treasury workflows and much more.Agentic AI 2025 CitigroupBigTech References to Agentic AI Up 17x in 12 months1,8001,5001,20090060030004Q231Q24
17、2Q243Q244Q24The AI Flywheel is in Motion 4Q2023 values indexed to 100AI is driving a powerful flywheel of technology,with advanced systems creating better AI.This in turn accelerates breakthroughs in computing and capabilities at unprecedented pace.2025 is the year of Agentic AI.We see AI and Agenti
18、c AI having a bigger impact on the economy and finance than the internet era.Agentic AI turbocharges the Do It For Me Economy.Finance&the Do It For Me EconomyAI/MLGenAIComputerSystemsAgentic AINotes:1.Based on keyword searches in company documents,transcripts,and press articles for“agentic AI”,“AI a
19、gents”and“AI bots”;2.Non-financial sector reference above refers to all sectors excluding the financials and the technology sector.Source:AlphaSense,Citi Global InsightsSource:Citi Global InsightsNon-FinancialTechnologyFinancial 2025 CitigroupAgentic AI Use Cases Across Financial Services Agentic AI
20、 Use CasesWealth Management/Retail BankingCorporate BankingInstitutional InvestorInsurancePersonalized Offers Adaptive financial advice Real-time savings goal optimization Custom lending offers Optimized loan structures Dynamic pricing Dynamic investment portfolios Bespoke investment plans Tailored
21、insurance policies Dynamic loyalty offersCustomer Engagement Virtual financial assistants Tax and retirement planning agents Financial planning agents Adaptive tax planning Custom research insights Real-time market alerts AI-driven query handling Proactive policy updates,renewalsOperational Efficien
22、cy Automate routine tasks with context-aware workflows Streamlined complex operations Invoice processing,reconciliations Automated fund review reports Adaptive claim managementRisk and Underwriting Real-time risk profiling Predictive default modelling Real-time risk assessment Diversification risk m
23、anagement Real-time hedging strategies Real-time underwriting modelsFinancial Forecasting Future savings,expense insights Live cashflow forecasts Dynamic investment timing strategies Investment performance insights Predicting claim reserves Automated premium calculationsKYC/Onboarding Adaptive ident
24、ity verification Real-time AML compliance Adaptive onboarding workflows Real-time sanctions monitoring Real-time due diligence Investor suitability analysis Adaptive identity verificationFraud Prevention Contextual-based suspicious activity detection Coprorate fraud detection Insider trading detecti
25、on Automated regulatory compliance Claim fraud detection with contextual analysisInternet Activity:Humans vs BotsSource:Imperva(a Thales Company),Citi Global InsightsBots already account for half of global internet traffic,and a large part of them are malicious.Source:Citi Global Insights100%80%60%4
26、0%20%0%201820192020202120222023Good BotsHumansBad Bots62%63%59%58%53%50%18%13%15%15%17%18%20%24%26%28%30%32%Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 6 Contents Agentic AI The New New Thing 7 Imitating Humans&More.8 2025:The Year of Agentic AI.10 Your Own Digital Ja
27、rvis.12 Bad Bots&Beyond:Commerce 2.0.14 AI x Crypto.16 Competitive Destruction&New Technologies.16 Agentic AI Could Supercharge Knowledge Workers.18 The End of Physical Roles.19 Governance&Cybersecurity is Critical for AI Agents.20 Agentic AI Use Cases in Finance 23 Time to Build Expert View.24 Inve
28、sting in AI Implementation Expert View.24 Compliance&Regulatory Agents.26 Deepfakes&Frauds.27 Onboarding&KYC.28 Agentic Contracts.29 Corporate Treasury&Payments.30 Wealth&Agentic Advisers.31 Hyper-Personalization.33 Loan Documentation&Credit Workflow.34 Insurance&Agentic Claim Management.35 Investme
29、nt Research.36 What are Agents?37 How Do AI Agents Work?.37 Building Blocks for Agentic Workflows.38 January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 7 Agentic AI The New New Thing In the words of Andrew Ng,computer scientist and technology entrepreneur,agentic Artificial
30、Intelligence(AI)is one of the most exciting technology trends right now.1 Agentic AI is the enabler of the next generation Do It For Me(DIFM)economy,which could lead to material changes on how users access services and markets are structured.The internet era reshaped finance and the economy.AI,inclu
31、ding agentic AI,might generate even greater change.The internet era DIFM promised a user-centric economy.Agentic AI holds potential to turbo charge this.We could all have our own Digital Jarvis,Iron-man style.Cloud computing increased operational leverage by converting capex into opex.Agentic AI cou
32、ld hyperscale opex,enabling a potentially golden age for start-ups.Repetitive tasks,which have often been outsourced to contractors or overseas,will likely instead be done by agentic AI.Many people could go from managing other humans to managing agents.Tasks that we have not outsourced are unlikely
33、to be automated.Fully autonomous agents in finance,in the near term,are likely to be limited.In 2025,we may see the first AI agents join the workforce and materially change the output of companies.2 SAM ALTMAN,CEO,OPENAI General purpose technologies reshape how we live and work.The steam engine and
34、later electricity transformed the world of the late 1700s to early 1900s.The industrial revolution,starting in the UK in the 1700s,swept across the world during the 1800s and beyond.Each new era creates new ways of living and working,and with it,new firms,and institutions.We are in an era now where
35、we are moving way faster than Moores Law These new AIs are helping us create new computer systems which helps us create even better AI That flywheel is really flying now the progress with AI agents over the next year or two will be spectacular and surprising.3 JENSEN HUANG,FOUNDER&CEO,NVIDIA The int
36、ernet and the digital economy have been key in shaping the start of the 21st century,creating new e-commerce,social media,and mobile payment companies.Many aspects of our lives,from how we manage our money to meeting our future spouses,have been changed by the internet.The unfolding of the AI era ho
37、lds the potential to create new winners and losers.Generative AI(GenAI)has driven the AI narrative since 2023,but there is a new kid on the block:agentic AI,which utilizes GenAI to imitate human agency in autonomous decision-making and executing tasks.1 Andrew Ng,The Rise of Agentic Workflows in AI,
38、19 November 2024.2 Sam Altman,Reflections blog post,06 January 2025.3 Dreamforce 2024,Trailblazing the Future of AI,17-19 September 2024.What is Agentic AI?AI systems capable of autonomous decision-making and actions.In finance and business,many such systems will likely be semi-autonomous in practic
39、e.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 8 Agentic AI will power a startup golden age.Cloud(computing)moved the economics of entrepreneurship from CAPEX to OPEX.Agentic AI moves them from employee payroll to software subscription.Digital banks and regulated FinTe
40、chs will be able to leverage their tech infrastructure and licenses to grow even faster than before.HUY NGUYEN TRIEU,CO-FOUNDER,CENTRE FOR FINANCE,TECHNOLOGY AND ENTREPRENEURSHIP(CFTE)Agentic AI is still a work-in-progress.To get a better understanding of what is coming,and to have a peek at what is
41、 being built right now,we spoke to over 30 founders and executives from AI startups,BigTechs,consultants,and policy institutes.Figure 1.Financial Firms and Eras ERA TRENDS FINANCIAL FIRMS RENAISSANCE(1400s-1500s)Trade&Commerce in Europe Medici Bank(1397)EARLY MODERN(1600s-1700s)Maritime Trade,State
42、Finance Bank of Amsterdam(1609)Bank of England(1694)INDUSTRALIZATION(1700s-1800s)Industry,Infrastructure,International Trade Rothschild(late 1700s)Deutsche Bank(1870)COMPUTERS(1950s-1980s)Mainframe Computers,Proprietary Electric Networks Visa(1958)Mastercard(1966)INTERNET(1995 onwards)Digital Econom
43、y,e-Commerce PayPal(1998)Alipay(2004)SMARTPHONE(2009 onwards)Mobile Internet,App Economy,API Economy Nubank(2013)Revolut(2015)AI(2020s onwards)Productivity,Personalization,Agentic AI Workflows AI Powered FinTechs Agile Incumbents Source:Citi Global Insights Imitating Humans&More AI is software that
44、can imitate human cognitive functions such as reasoning and learning.AI could have an even bigger impact than the internet on society and the economy.AI,especially generative and agentic AI,will move us to the Do It for Me Economy from the Show It to Me Economy.AI,like previous technology revolution
45、s,should lower costs,raise productivity,and boost overall economic activity.The internet revolutionized information,communication,and related activities.AI will aid in reshaping cognitive work and how tasks are executed across nearly all activity.AI does not just imitate humans.It can produce emerge
46、nt answers,and in some cases novel solutions.For example,AlphaZero became the worlds best chess computer program in 2017 based purely on AI training.Previous programs were based on a library of human actions.AlphaZero created chess moves which humans copied.4 AI can also now create content developed
47、 from its training.GenAI can create code,text,pictures,or video.OpenAI demonstrated its GPT-3(generative pre-trained transformer)model in 2020.ChatGPT,initially based on the GPT3.5 model,was launched to the public in November 2022,transforming awareness and interest in AI.4 Henry Kissinger,Eric Schm
48、idt,and Daniel Huttenlocher,The Age of AI,16 Nov.2021.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 9 The next frontier is agentic AI.While GenAI relies on user prompts,agentic AI can act autonomously to make plans and decisions and execute specific tasks.OpenAI CEO Sam
49、 Altman has described AI agents as“super competent colleagues”that will be able to complete tasks5.The IT department of every company is going to be the HR department of AI agents in the futureToday they manage and maintain a bunch of software from the IT industry;in the future they will maintain,nu
50、rture,onboard,and improve a whole bunch of digital agents.6 JENSEN HUANG,FOUNDER&CEO,NVIDIA AI agents as a co-worker are expected to become increasingly common.Today,AI agents tend to be focused on specific tasks.In the future,AI agents will combine to manage multiple tasks.United Fintech was one of
51、 the first finance leadership teams to talk to us about AI as“new workers”.We expect this to be common language in 2025.Figure 2.Comparing Artificial Intelligence,Generative AI,and Agentic AI PARAMETER ARTIFICIAL INTELLIGENCE(AI)GENERATIVE AI(GenAI)AGENTIC AI Definition Broad field of computer syste
52、ms that simulate human intelligence Subset of AI focused on creating new content,such as text,images,and code AI systems capable of autonomous decision-making and actions Primary function Performs tasks like analysis,prediction,and automation Generates creative outputs based on patterns learned from
53、 data Acts autonomously to make decisions and execute tasks without human intervention Examples/applications Chatbots,recommendation systems,fraud detection,and diagnostics Art,audio,video,music,text,software generation Autonomous trading,robotics,complex simulations,and digital personal assistants
54、Impact Improves efficiency,reduces costs,and enhances productivity Revolutionizes creativity,automates content creation Redefines autonomy in industries,driving efficiency and innovation Scope Broad,spanning predictive analytics,automation,and decision-making Focused on creativity and content genera
55、tion Broader than GenAI,with the ability to take actions and interact with the environment Data dependency Relies on structured or unstructured data for training and operation Requires vast datasets for training(e.g.,text,images)Uses real-time and historical data,including data it has generated itse
56、lf,for context-aware decision-making Autonomy Limited;often requires human input or oversight Non-autonomous;outputs depend on user prompts Autonomous,capable of initiating actions without human prompts Complexity Varies from simple rule-based systems to advanced learning algorithms Advanced,requiri
57、ng deep learning architectures for training Highly complex,integrating AI decision-making with execution Key technologies Machine learning,deep learning,neural networks Transformer models(e.g.,GPT,DALL-E),generative networks Combines AI with sensors,actuators,and dynamic decision-making frameworks T
58、echnology limitations Performance dependant on model training,data and features Resource-intensive for training and deployment Complex to evaluate and deploy Ethical concerns Bias,transparency,data privacy Copyright issues,misinformation,ethical content use Accountability for actions,risk of uninten
59、ded consequences,ethical decision-making Source:Citi Global Insights In our workspaces,we have become used to our colleagues being in different time zones.In the AI era,software will become our colleagues.Agentic AI is likely to become part of daily workflows in our personal and professional lives.I
60、ts applications could range from self-driving assistants to self-driving automobiles and beyond.We do a deep dive into the use cases in finance in later in this report.5 MIT Technology Review,May 2024 6 CES 2025,Jensen Huang Keynote Speech,06 January 2025.Citi GPS:Citi GPS:Global Perspectives&Soluti
61、ons January 2025 2025 Citigroup 10 What is clear is that things appear set to change.Previously,only the privileged had personalized offerings such as assistants to take care of every facet of their lives:A chef to cook.A butler to serve.A chauffeur to drive.A tutor to teach.A private banker to mana
62、ge their finances.In my 20 years of experience in the software industry and data analytics industry and the last decade focusing on AI,I have never been so excited to see a general-purpose technology that can be helpful to humanity in so many ways.The technology will get better at a faster pace.It i
63、s hard to extrapolate what these technologies can do with each iteration.SIRISHA KADAMALAKALVA,GLOBAL HEAD OF AI/ML INVESTMENT BANKING,CITI In tomorrows AI-powered world,access to personalized offerings will become more widely accessible.Everyone could have a private banker in their wallet,smartphon
64、e,or wearable device and they wouldnt need to be a high-net-worth individual.But we do not necessarily want to outsource all jobs to third-party software or a human.For example,you may ask an agent about flights,routes,and car hire options for your family holiday,but would you want to automate every
65、thing?Part of the joy of a family holiday lies in the planning you do together as a family.You might not want the agent or a bot to decide the destination and where to stay.2025:The Year of Agentic AI Agentic AI is number one in IT consultancy Gartners top 10 strategic technology trends for 2025.Sta
66、rtup accelerator Y Combinator has predicted that vertical AI agents specialized tools designed to automate entire workflows could disrupt software in the same way software as service(SaaS)did over the past two decades.7 The year 2025 will be a“wow”year for agentic AI in finance,similar to the year 2
67、023 being a“wow”year for GenAI and 2024 for large language models.REN ITO,CO-FOUNDER&COO,SAKANA AI So why now?Agentic AI development is powered by a combination of technological breakthroughs.These include model improvements including contextual understanding,memory,and multi-tasking abilities,and o
68、n the demand side process-oriented industries looking for improved solutions beyond search and retrieval functions.Agents that will change our lives and our workflows are a good marketing theme for BigTech firms.But the real driving force has been rapid changes in the underlying technology during 20
69、23-24.LLMs capabilities have dramatically improved,and Agentic AI builds on top of this improvement.7 Y Combinator Vertical AI Agents Could Be 10X Bigger than SaaS,22 November 2024.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 11 As businesses ramp up AI spends and hunt
70、 for ROI,agentic AI is the new frontier that orchestrates a whole zoo of AI models of all sizes and modalities to automate processes and materialize productivity gains.HEAD OF AI EMEA FOR FINANCIAL SERVICES&TELECOM,US BIGTECH FIRM GenAI had a“wow”moment in 2023.BigTech references to the term in corp
71、orate statements and press articles surged 55 times in the 12 months from 4Q22 to 4Q23(Figure 3).References to agentic AI increased 17x in 2024 and we expect them to go parabolic in 2025(Figure 4).Figure 3.BigTech References to GenAI up 55x in 12 months Figure 4.BigTech References to Agentic AI Up 1
72、7x in 12 months Notes:1 Based on keyword searches in company documents,transcripts,and press articles for generative AI OR GenAI OR LLM OR GPT and“Agentic AI”,“AI agents”and“AI bots”;2 Non-financial sector reference above refers to all sectors excluding the financials and the technology sector.Sourc
73、e:AlphaSense,Citi Global Insights Agentic AI is not just about talk;investors are also walking the talk.In 2024,37%of the VC funding and 17%of the deal activity was to AI startups both all-time highs according to CB Insights.Autonomous agents and digital co-workers had the biggest growth in VC deal
74、activity in 2024,followed by GenAI for customer support operations.Artificial Intelligence is driving strategic investments as corporates globally are looking to transform their business using AI.The amount of investment going into AI related data centers and the amount of funding that will need to
75、take place over the next 5-10 years is transformational.8 PHILIP DRURY,GLOBAL HEAD OF TECHNOLOGY AND COMMUNICATIONS IN BANKING,CITI Based on the recent encouragement by BigTech leaders,startup entrepreneurs,and the VC community,we expect a continued growth in investments and corporate announcements
76、around agentic AI during 2025.And all of this is before agentic AI has had a ChatGPT-like splash of a launch.8 Bloomberg,AI Will Drive M&A Pipeline,Says Citigroups Drury,12 June 2024.Total Mentions 1,142Total Mentions 9,081Total Mentions 13,42701,0002,0003,0004,0005,0006,0004Q221Q232Q233Q234Q23Finan
77、cialsNon-FinancialTechnology4Q2022 values indexed to 100Total Mentions 100Total Mentions 945Total Mentions 3,13503006009001,2001,5001,8004Q231Q242Q243Q244Q24FinancialsNon-FinancialTechnology4Q2023 values indexed to 100 Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 12 Fi
78、gure 5.Percentage of VC Funding and Deals to AI Startups9 Figure 6.VC Deal Activity in Agentic AI Startups Grow10 Source:CB Insights,Citi Global Insights Source:CB Insights,Citi Global Insights We think the hottest new area of venture investment will be autonomous AI agents,software agents that can
79、simulate human behavior and plan,make decisions,and execute tasks in complex environments without human intervention or supervision.Autonomous AI agents are just emerging from research and development,and as of now,their funding pales in comparison to the funding for GenAIs core infrastructure.11,12
80、 VIBHOR RASTOGI,GLOBAL HEAD OF AI/ML INVESTMENTS,CITI VENTURES Your Own Digital Jarvis In the Iron Man movies,Tony Stark has an AI assistant integrated into his home,and his suits,called J.A.R.V.I.S(Just A Rather Very Intelligent System).In the original Marvel comics,Jarvis is a human butler.As with
81、 much fiction,reality may be following in its footsteps.We would not need to be a billionaire to have a digital Jarvis.In Citi GPS:AI in Finance:Bot,Bank&Beyond(June 2024),we argued that the next frontier for AI innovation is likely to be autonomous agents and people using AI bots to manage their li
82、ves.Instead of prompting LLMs with simple one-line instructions,we could embed higher AI capabilities in existing and new digital assistants(hello,Siri!).These bots,equipped with sophisticated algorithms and access to vast amounts of data,will negotiate with counterparties to secure the best possibl
83、e deals for their users.This shift will not only streamline services but also ensure that decisions are made with a level of precision and foresight that human users may not have.9 CB Insights,State of Venture 2024 Report,07 January 2025( CB Insights,State of Venture 2024 Report,07 January 2025( For
84、tune,AI Investors Need Focus,But Their Exuberance Isnt Entirely Unjustified,20 September 2024.12 Evident AI Findings:Citi Leads 2023 AI Venture Activity Amongst Major Banks,April 2024.0%2%4%6%8%10%12%14%16%18%20%0%4%8%12%16%20%24%28%32%36%40%2015201620172018201920202021202220232024VC Funding to AI a
85、s%of Total VC FundingVC Deals in AI as%of Total VC Deals(RHS)150%150%113%82%60%52%50%45%44%44%Autonomous agents&digital co-workersGenAI for customer support operationsWater treatment systemsEdge AI processorsEarth observation satellitesIndustrial humanoid robotsSecurity operations centre AI agents&c
86、opilotsAI-derived biologics-proteinsAutonomous driving systemsSustainable cement&concrete developersDeal growth YoY(Enterprise tech)(Enterprise tech)(Industrials)(Enterprise tech)(Industrials)(Industrials)(Enterprise tech)(Healthcare and life sciences)(Industrials)(Industrials)2023 deals868111025811
87、1892024 deals20151720163812162613January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 13 With wider adoption of autonomous AI agents,consumers will be enabled to make informed decisions without being directly involved in data gathering,comparison shopping,and carrying out manu
88、al tasks to execute the transaction.Consumers can focus on the yes/no/switch decisions.Whose Bot Is It Anyway?In this new paradigm,the critical decision for consumers will be selecting the right bot or agent.Choosing bot-powered advisers,much like choosing human personal finance advisers,will become
89、 a key task.But who will the bots work for BigTech firms,trusted institutions such as banks,or new startups?Leading banks and firms will most likely opt to provide their own AI-powered services.But BigTech firms may have a competitive advantage in terms of being digitally native and have faster go-t
90、o-market speed.In some markets,they may also have stronger consumer brands.Smaller firms and startups may see their growth turbo charged by the growth of agents.AI may dramatically improve their reach.But will consumers want to spend too much time thinking about which autonomous agent to use?We are
91、likely to default to known and trusted brands.As David Birch noted in Citi GPS:AI in Finance:Bot,Bank&Beyond(June 2024)The concept of value will be critical.Customers are likely to deal with bots that reflect their values and interests.This could lead to the creation of a marketplace for bots where
92、individuals select a bot based on desired preferences.In retail activities,traditional marketing strategies will not work on bots.For instance,a bot is not influenced by brand loyalty or advertisements;it operates within the model parameters on APIs.Products designed for bots,which may be making the
93、 selection and or even buying decision,will also need to be translatable to humans.Lloyds Banking Groups Kirsty Rutter asks in the same Citi GPS:AI in Finance:Bot,Bank&Beyond:if AI-driven client interface technology exists,why should humans waste time on petty administrative tasks?In a financial man
94、agement context,an AI bot can directly engage with the bank on behalf of the customer.For example,if a mortgage is up for renewal,the bot can notify the customer of the end date,conduct research to find the best deals,determine the right product,and fulfill formalities for refixing the mortgage rate
95、.Bots can achieve this based on parameters(retirement age,preference for debt tenure,maximum acceptable price,etc.)pre-determined by the customer.In the initial phase,the bot could rely on human confirmation before executing tasks;but as trust in the bot grows,tasks can be handled autonomously.Finan
96、cial institutions and large organizations are likely to proceed with due caution,due to the weight of regulations,internal processes,and governance requirements.However,they may find their speed of adoption being accelerated by client behavior.Individuals are adopting AI fast,and agentic AI may be n
97、o different.My 70+-year-old mother took 10 odd years to adopt basic internet use-cases but has been using Perplexity since 2023.NITYA SHARMA,FOUNDER,SIMPL Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 14 Digitization has shortened the speed of adoption,as we have seen w
98、ith the spread of ChatGPT.AI agents are part of an ongoing technological revolution that could fundamentally change commerce(how individuals and businesses buy and sell),capital and competition(scale and reach of companies),and labor(the nature of work).The implications for society and governance co
99、uld be far-reaching.The clients bots will likely arrive first,while bank and big company agents may be playing catch up.Our agents will be out there,working on our behalf to find the best travel possibilities,assessing mortgages,and generating views on investments in AI stocks.Who will provide these
100、 agents?It may be incumbents.Or could it be brand new entrants?Perhaps the two person and infinite capacity AI agent army start up?Bad Bots&Beyond:Commerce 2.0 A world with AI agents poses many risks.Bots,or task-specific automated programs,already account for half of global internet traffic,and mos
101、t are malicious actors.13 Social media and cryptocurrencies bear witness to the prevalence of bots,and the risks of them creating negative results.If bots can be bad,malicious agentic AI has the potential to be even more dangerous.Bots are simple rule-based programs.AI agents are much more advanced
102、systems that can learn,adapt,and make decisions based on their context and objectives.They are smart bots,with an ability to evolve and act.Agentic AI acts as a delivery mechanism and could enable the dissemination of fraud at scale.Half of all fraud today already involves some form of AI14,which is
103、 likely to increase.Illicit actors are usually early adopters of new technology as they are incentivized by the availability of large profit margins.Figure 7.Internet Traffic:Humans vs Bots15 Figure 8.Bad Bot Traffic by Industry16 Source:Imperva(a Thales Company),Citi Global Insights Source:Imperva(
104、A Thales Company),Citi Global Insights 13 Imperva(A Thales Company),Bad Bot Report,16 April 2024.14 Signicat,Battle Against AI-driven Identity Fraud,30 May 2024.15 Imperva(a Thales Company),Bad Bot Report,16 April 2024.16 Imperva(a Thales Company),Bad Bot Report,16 April 2024.62%63%59%58%53%50%18%13
105、%15%15%17%18%20%24%26%28%30%32%0%10%20%30%40%50%60%70%80%90%100%201820192020202120222023HumansGood BotsBad Bots57%49%46%45%42%41%33%32%31%31%27%26%24%23%23%21%21%19%18%3%15%14%4%7%8%7%8%55%2%27%20%10%7%9%7%14%36%2%40%36%41%51%51%51%60%60%14%68%46%54%67%70%68%72%65%45%80%GamingTelecom&ISPsComputing&I
106、TTravelCommunity&SocietyBusiness ServicesHealthcareNewsEntertainmeentGamblingFinancial ServicesRetailEducationLaw&GovernmentLifestyleSportsAutomotiveFood&GroceriesMarketingBad BotGood BotHumanJanuary 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 15 Agents may work technological
107、ly and undertake commerce in 2025.But as Erin McCune and her colleagues at Bain&Company ask17,how do we ensure adequate controls and guardrails?They highlight two potential approaches to enabling agents with sufficient authority to act without compromising security:AI Wallets:AI agents have a prefun
108、ded wallet that enables it to act autonomously without needing approval for every transaction,but the total budget is capped.This is analogous to a B2B card today that an employer can use under specific rules(time,place,amount).Automated Authentication Layer:AI agents are asked for contextual authen
109、tication,for example chat history with the original ask from the human user.A parallel layer of authentication might include past agent purchase history.This may be combined with other markets(e.g.IP addresses or other use data).Figure 9.Important Considerations of AI Agents on Commerce and Finance1
110、8 Source:Erin McCune&Roger Zhu,Bain&Company 17 Erin McCune,Roger Zhu,et al.,Bain&Company,Implications of AI Agent Commerce,12 December 2024.18 Erin McCune,Roger Zhu,et al.,Bain&Company,Implications of AI Agent Commerce,12 December 2024.What is the appropriate business model for intelligent agent inf
111、rastructure?What will you be monetizing?How do you orchestrate collaboration and engender trust to ensure agentic transactions are made consistently and safely?PAYMENT INFRA How consolidated will the agent landscape become and in what timeframe?Who will build the winning agents?Major LLMs or several
112、 fragmented/specialized providers?INVESTORS Who will own customer relationships?Banks or the provider of the agent purchasing on behalf of banks customer?How will you prevent fraud and manage risk in the context of AI agent buyers?BANKS Are you ready to accept payments from an intelligent agent?How
113、will you know the agent is legitimate?Are any new capabilities necessary to ensure against purchase within the delegated spend authority?PAYMENT SERVICE PROVIDERS What role(if any)will there be for verticalized agents?How can agents augment your offering?If so,do you possess enough proprietary data
114、to train your own model?INDEPENDENT SOFTWARE VENDORS Can current legal mechanisms to delegate authority work for AI agents?Is there an opportunity for supervisory agents to monitor financial institutions/PSPs activity and that of agents themselves?REGULATORSErin McCune,Partner,Bain&Company Roger Zhu
115、,Partner,Bain&Company Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 16 AI x Crypto Cryptocurrencies and blockchain could play an important role in the agentic AI world to help ensure openness,transparency,and competition.GenAI and agentic AI solutions are embedded into
116、the services of a small number of BigTech firms,and many users by default would use the services provided by these digital leviathans.But how can we ensure that the data used,model construction or the agent behavior is not biased or corrupted?Crypto incentives provide one kind of mechanism to ensure
117、 that you get paid to validate data and model quality.On-chain participants can be rewarded for data verification and model validation via so-called staking systems.On-chain staking is a mechanism where participants contribute,or stake,their cryptocurrency tokens.The tokens are locked in a smart con
118、tract and act as collateral.In the AI example,participants stake tokens as collateral which they can lose,creating an economic incentive model.AI agents can create,issue,and distribute tokens,and they can hold private keys and control crypto wallets.Those crypto tokens can be funded as a limited par
119、tner(LP)to liquidity pools,participate in a Decentralized Autonomous Organization(DAO)governance,and many more models.Essentially,they can be sustainable and fund opportunities,pay invoices as they see fit to achieve their overall goals.A future world of AI agents,combined with IoT(Internet of Thing
120、s),may need to make autonomous micropayments for data or energy it uses as part of their ongoing activity.Cryptocurrencies,including stablecoins,may be one of the payment options used.The AI agents themselves may be decentralized and crypto would be the incentive mechanism for payment of services pu
121、rchased or to ensure collaboration.Just as there is an app for everything,there will be an AI agent for everything.Everything that is possible off-chain,will be possible on-chain.AJIT TRIPATHI,CORE CONTRIBUTOR,HADRON FOUNDERS CLUB Crypto is in many ways a natural partner for AI agents.Bots,the first
122、 generation of AI agents,have played an outsized role in crypto markets.Attention is meanwhile a very important commodity on-chain and online(Twitter and Telegram are in effect the house communication channels of crypto).Bots which interact with live data 24x7 help capture,control and direct attenti
123、on.Hence,bots become a key component of the crypto decentralized world.Evolving from bots to AI agents was a natural development,and we will need to embed stronger governance and higher cybersecurity in the world of bots and agents.Competitive Destruction&New Technologies General purpose technologie
124、s,like the steam engine,electricity,and the internet,have reshaped economies and societies.Will AI lead to another wave of creative destruction or a retrenchment of current oligarchic capitalism?Will there be more consolidation at the infrastructure layer,and increased competition at the app layer?I
125、n previous cycles,technology revolutions led to the rise of new firms.The steam engine era saw large factories and industrial businesses replace artisanal firms.Electricity further transformed mass production and urbanization.And the first and second industrial revolutions ushered in a new era of ba
126、nks and financial firms.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 17 I have spent most of my waking hours over the past two weeks thinking about this exact topic:Small teams+AI Agents+Cheap Compute.Exciting is one way to describe it.Terrifying is another.HEAD OF AI
127、EMEA FOR FINANCIAL SERVICES&TELECOM,US BIGTECH FIRM Similarly,the internet era from the 1990s onwards has seen the creation of new e-commerce,media,and payments and FinTech firms.Traditional retailers and media companies saw significant disruption and loss of clients.Some managed to transform and su
128、rvive,such as independent bookshops in America.The internet era promised decentralization.Commentators have talked about how“more billion-dollar companies will be built by 4-5 people”(entrepreneur and investor Naval Ravikant in 2012).Crypto firms,such as Tether,have generated billions in profits wit
129、h a small number of employees,while individual influencers have built huge global followings(Mr.Beast has 342 million followers on YouTube,as of 08 January 2025).But at the same time,the largest companies have never been bigger in terms of size and influence.In recent years,the US stock market has b
130、ecome increasingly concentrated.The top-10 stocks now account for 38%of the S&P 500,a record over the last century.And the top-5 stocks represent 27%of the S&P 500,compared to less than 10%in 2014.The top of the capitalization table is dominated by technology stocks,and they are spending huge sums o
131、n AI investment.In January 2025,Microsoft announced plans to spend“approximately$80 billion to build out AI-enabled datacenters to train AI models and deploy AI and cloud-based applications”.19 Figure 10.US Stock Market Concentration is at its Highest20 Figure 11.Global Data Centre Market Source:Fin
132、aeon,S&P 500,Citi Global Insights Source:DLA Piper,Citi Global Insights The major cloud computing firms are all-in on agentic AI.At the recent Amazon Web Services(AWS)re:Invent 2024 event in December,key announcements included a new generation of foundation models(Amazon Nova)and tools for automatin
133、g complex tasks(Bedrock Agents).19 Microsoft,The Golden Opportunity for American AI,03 January 2025.20 Bryan Taylor,Finaeon,200 Years of Market Concentration,22 May 2024.27%22%28%26%19%20%28%26%28%28%22%20%23%19%38%1880 1890 1900 1910 1920 1930 1940 1950 1960 1970 1980 1990 2000 2010 2024Top 10 Stoc
134、ks as%of S&P50030048320242029$billion10%CAGR Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 18 Microsoft Azure rolled out its AI Agent Service in November 2024 designed to automate tasks such as customer support and workflow management.21 Oracle in September 2024 announc
135、ed AI agents that would automate workflows in finance,sales,supply chain and similar functions.22 Cloud computing firms illustrate a paradox of our era.They are increasingly gargantuan in size they have multi-billion-dollar capex plans,and their parent firms boast trillion-dollar equity market valua
136、tions.At the same time,the rise of cloud computing has been a key enabler of the growth of new entrants such as FinTechs and neobanks.Cloud service providers removed the high fixed technology costs of competing with large incumbents,while also offering access to the latest innovations.Private equity
137、 markets are in a new elevated normal in dealmaking,with activity expected to accelerate in the coming years.Deal flow remains focused on the technology sector,with AI-related investments concentrated in software and data centers.Infrastructure funds,in particular,view data centers as a significant
138、growth opportunity ahead.23 ANTHONY DIAMANDAKIS,HEAD OF GLOBAL ASSET MANAGERS,INVESTMENT BANKING,CITI Marc Andreessen of Andreessen Horowitz(aka a16z)noted in 2009 that“new technologies like cloud computing allows everyone to scale their companies.24 Cloud providers will likely play a similar role i
139、n the current unfolding AI era,enabling new entrants and smaller firms to scale fast.In fact,the combination of cloud,AI,and agents may lead to a further revolution in the delivery of digital services,including finance.Larger financial firms may move with caution due to cultural,regulatory,and techn
140、ological constraints,but a new generation of risk-embracing startups may attempt to race ahead.The minimum viable size for the next generation of new entrants,including in financial services,may be considerably smaller than in the past decade.The operational efficiency gains in workflow management f
141、rom AI agents,combined with GenAIs content generation,could lead to interesting outcomes.Agentic AI Could Supercharge Knowledge Workers Like the printing press,AI decentralizes access to knowledge and creativity,enabling individuals without specialized skills to produce competitive content.Add in AI
142、 agents,and you can scale the reach of a single professional or small team at an exciting(or terrifying)speed.While AI offers the promise of a princely life,with a surfeit of new digital personal assistants,it also raises significant social questions.How will these AI era princes and princesses earn
143、 a living?In previous economic revolutions,many new jobs were created.Will it be the same this time?21 Microsoft Community Hub,Introducing Azure AI Agent Service,19 November 2024.22 Oracle Press Release,Oracle AI Agents Help Organizations Achieve New Levels of Productivity,11 September 2024.23 Bloom
144、berg Brief,Citis Diamandakis on Private Equity Dealmaking,30 May 2024.24 The Charlie Rose Show,Marc Andreessen,09 February 2009.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 19 Agentic AI could lead to proliferation of new high-skill jobs in AI development,ethics,govern
145、ance,and oversight,consistent with our thesis in Citi GPS:AI in Finance:Bot,Bank&Beyond(June 2024).Entire industries may arise around training,fine-tuning,and auditing AI systems.Data will be key,and curation jobs are bound to grow exponentially.Many repetitive jobs are potentially at risk of automa
146、tion,intensifying job polarization and potentially economic inequality.The way electricity replaced repetitive manual labor,agentic AI could automate repetitive mental tasks and aid decision-making.The roles destroyed are likely to be the ones doing more routine tasks.The likes of software engineers
147、,business analysts and client support staff are not about to disappear.Rather,their roles may be augmented and enhanced using AI.There will be a new focus on the synthesis of professional human skills and judgement combined with AI capabilities.Agentic AI will not directly replace jobs in financial
148、services in the short term.However,it will supercharge employees,and more can be done with less people,leading to less headcount.New roles will of course emerge too.25 HUY NGUYEN TRIEU,CO-FOUNDER,CENTRE FOR FINANCE,TECHNOLOGY AND ENTREPRENEURSHIP(CFTE)However,the productivity of the software enginee
149、r,analyst or client adviser improving will raise an important social question:if we can do more output with less input,what will happen to the junior engineers,analysts,and client support staff?Where will people who did the entry level jobs in these roles go?Figure 12.The Future of the Workforce JOB
150、S CREATED AUGMENTED Development AI engineer/trainer Software engineer using AI tools Data AI data curator AI-powered business analysts Governance AI ethics&safety Lawyer aided by AI Client Engagement AI UX designer AI-enabled client support Logistics Autonomous systems/vehicle designer Logistics,man
151、agers aided by AI Source:Citi Global Insights The End of Physical Roles Just as offshoring and outsourcing of white-collar jobs in the past few decades raised questions about talent pipelines and entry level jobs,AI and agentic AI will reframe the conversation,but this time globally.The nature of ta
152、sks and roles in shared services hubs will change due to agentic AI.There will also be a revaluation of value ascribed to skills seen as human.Creativity,empathy,and judgement will be more in demand and could be valued at a premium.The growth of AI-powered deepfakes and an explosion of AI agents may
153、 create a desire for person-to-person contact,which will be revalued upwards.25 Huy Nguyen Trieu,The AI-fication of Jobs:Preparing Ourselves for the Future of Work,02 November 2024.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 20 Sam Altman,CEO of OpenAI,predicted in 20
154、21 that the price of work in front of a computer will drop faster than the price of work in the physical world.But many new or augmented roles will also emerge.In addition,look out for embodied AI.Agentic AI is likely to destroy many physical roles.Waymo,the autonomous vehicle firm with no human dri
155、vers,launched in August 2023.At the time the market share of Uber and Lyft were at 66%and 34%in San Francisco.Just 15 months later(November 2024),Waymo was at 22%the same as Lyft.26 However,the CEO of Lyft has disputed these numbers.27 In December 2024,Waymo published a study claiming that it had a
156、better safety performance than human-driven vehicles with an 88%reduction in property damage claims and 92%reduction in bodily injury claims.28 Taxicab and long-distance truck drivers,as well as routine computer or desk work could also be largely replaced by software agents.But many existing jobs wi
157、ll be turbo-charged by the incorporation of AI into their roles.GenAI is the fastest adopted consumer technology in history.Agents are becoming smarter and more effective,and we will probably expand our adoption of them in our personal lives before our professional lives,especially in jurisdictions
158、where large companies and banks are tightly regulated,which might place speed bumps on certain professional usage cases.Governance&Cybersecurity is Critical for AI Agents We believe regulators,policymakers,and supervisors will focus on two big areas when scrutinizing firms agentic AI projects.Both r
159、elate to the new element that agentic AI introduces,namely,1 agents acting autonomously and making decisions and 2 the risks that they present.Governance:The first point relates to how the agents will be overseen,in other words,the governance framework that is put in place.Firms should ensure their
160、current governance frameworks are fit for purpose for the age of agentic AI.Key points for consideration include:Agentic AI Use Cases:How has the firm decided where to deploy agentic AI?This is likely to be linked to a firms risk appetite.The more regulated firms are likely to deploy agentic AI in r
161、oles and functions that pose minimal risks.Stakeholders:Are all stakeholders included in the governance framework upfront,or are some,such as senior representatives from control functions,added as an afterthought?Accountability:Who will be accountable to the regulator if agentic AI gets it wrong and
162、 what mechanism exists for individuals to override the machines?26 The market share data only includes rides that start and end in a select geolocation in San Francisco.It also does not include airport rides.Waymo has operated 24x7 driverless for the past 15 months,but the first nine were limited to
163、 a private waitlist.27 Yahoo Finance,Lyft Gains as CEO Says San Francisco Share Losses Could Be Overblown,16 December 2024.28 Waymo Study:Waymo is Safer than Even the Most Advanced Human Driven Vehicles,19 December 2024.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 21 M
164、onitoring and Oversight:Has the firm considered all the possible scenarios where an agent could make a mistake,and the level of harm that could be caused(to clients,employees,and the market)?And has it rated the risks and put in place robust controls through continuous monitoring of agents?Data Priv
165、acy:Has the firm deployed enough controls so that agentic AI responsibly handles enterprise data and does not share any information that is not intended for sharing(externally or internally).Cyber Risks:The second area where regulatory attention is likely to focus relates to new cybersecurity risks
166、posed by agentic AI.For starters,the attack surface has increased because agentic AI integrates with multiple systems within a firm that all share multitudes of data.Further,because tasks are automated with agentic AI,there is a risk that bad actors will not be caught until after an AI agent has exe
167、cuted a decision.Human oversight to detect violations would be critical,as would real-time cybercrime detection mechanisms.A human-in-the-loop approach will be a critical governance component and an ultimate oversight function.They will continuously check if the AI agents are functioning as per poli
168、cy and catching any violations or breaches in good time.DREW PROPSON,HEAD OF TECHNOLOGY AND INNOVATION IN FINANCIAL SERVICES,WORLD ECONOMIC FORUM Traditional cybersecurity defenses are built on deterministic thinking where specific attack inputs would lead to specific outcomes.Such an approach is un
169、likely to work in the world of agentic AI because cybercriminals could deploy an army of hacker agents at a very low cost to penetrate enterprise firewalls.Implementing agentic AI safely will come at a cost,and we expect firms to ramp up cybersecurity spends and investments in parallel with infrastr
170、ucture implementation spends.The jobs of the chief information security officer and the human above the loop is likely to be further elevated as agentic AI grows.AI systems,while powerful,are still nascent in their development of ethical frameworks and decision-making processes.The lack of clarity a
171、round AIs moral compass and its internal workings amplifies the perceived risk for banks.Banks will be cautious in adopting AI(and agentic AI)at scale,as they are liable for errors made by humans or AI agents.ERKIN ADYLOV,FOUNDER&CEO,BEHAVOX Finally,firms need to pay attention to restrictions on the
172、 use of automated decision-making.The European Commission(EC)has weighed in on the topic and thinks that customers should not be subject to a legally binding decision that is based solely on automated processing.The EC argues that if a firm processes customer data through an online platform/channel
173、and uses computer code or algorithms to decide if a product or service should be offered to customers,the firm should review the automated Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 22 decision before communicating it to the client.The client must be given the right
174、to contest the decision,if they feel the outcome has been unfavorable.29 Spain is the first country to set up a regulatory sandbox as part of the EU AI Act.We anticipate that the next wave of sandboxes will see a lot of agentic AI applications,alongside consortia,working groups,and tech-sprints.In t
175、he next chapter,we investigate how agentic AI will be used in one of the most important parts of the economy:financial services.29 European Commission,Dealing with Citizens January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 23 Agentic AI Use Cases in Finance Banks and financ
176、ial firms may be proceeding with caution,but they are still doing more than many may have expected.Data from Microsoft highlights that the financial services sector is one of the leaders in AI adoption based on consumption revenue data for cloud services on the Azure platform using foundation models
177、.Figure 13.Financial Services is One of the Fastest Adopters of AI Note:1 Measured as Azure consumption revenue,i.e.,the monetary value of Azure services consumed by a customer for small and large language models;2 Telecom and media includes social media companies.Source:Microsoft Figure 14.Summary
178、of Agentic AI Use Cases Across Financial Services AGENTIC AI USE-CASES WEALTH MANAGEMENT /RETAIL BANKING CORPORATE BANKING INSTITUTIONAL INVESTORS INSURANCE PERSONALIZED OFFERS Tailored financial advice adapting to client behavior/market trends Customized savings goals that adjust real-time to chang
179、es in income/expense Custom lending offers Optimized loan structures based on corporate profiles Dynamic pricing for corporate accounts Dynamic investment portfolios Bespoke investment plans Tailored insurance policies Dynamic loyalty offers based on profiles CUSTOMER ENGAGEMENT Virtual financial as
180、sistants Tax planning agents Retirement guidance agents Financial planning agents Adaptive tax planning Custom research insights Real-time market alerts AI-driven customer support.Proactive policy updates,renewals OPERATIONAL EFFICIENCY Automate routine tasks with context-aware workflows Streamline
181、complex business operations Invoice processing and reconciliations Automated document review for fund performance Streamline claims with adaptive workflows RISK&UNDERWRITING Dynamic risk profiling for credit approvals and investments Predictive models for default probability with real-time update On
182、going risk assessment and liability coverage recommendations with live data Portfolio diversification risk management Real-time hedging strategies Real-time underwriting models using health,climate,and behavioral data FINANCIAL FORECASTING Insights for expenses,savings plans Predictive cashflows Liv
183、e cashflow forecasts Real-time insights for timing of entry/exit in investments Investment performance insights Actuarial predictions for claims reserves Automated premium calculations and loss forecasting KYC/ONBOARDING Adaptive identity verification Real-time AML compliance Adaptive onboarding wor
184、kflows Real-time sanctions monitoring Real-time due diligence Investor suitability analysis AI-assisted policy selection Adaptive identity verification FRAUD PREVENTION Detect unauthorized activity using contextual data Monitor corporate fraud Anomaly detection in trade finance and payment invoices
185、Insider trading detection Automated regulatory compliance Detecting claim anomalies Real-time fraud detection with contextual analysis Source:Citi Global Insights Retail&Consumer GoodsIndustrials&ManufacturingHealthcareFinancial ServicesTelco&MediaIndexed absolutes FY2024 Citi GPS:Citi GPS:Global Pe
186、rspectives&Solutions January 2025 2025 Citigroup 24 The financial services sector is one of the biggest users of Generative AI worldwide.MARTIN MOELLER,HEAD OF AI&GENERATIVE AI FOR FINANCIAL SERVICES,EMEA,MICROSOFT Time to Build Expert View Q:How much of current AI adoption is proof of concept vs.re
187、al-world application,and how does it compare to past digital transformation waves?Many AI use cases today are not new.Tasks like improving chatbots,automating documentation,or enhancing client interactions were long identified as valuable but never fully implemented.Now,we are seeing a shift where c
188、ompanies are finally delivering on these promises,moving from theory to practice.This mirrors patterns from past digital transformation waves,such as the early days of mobile apps.Initially,there was skepticism and experimentation,followed by widespread initiatives once the value became clear.2023 w
189、as a year of excitement and experimentation in GenAI,now the focus is on scaling and delivering value.Q:How do you approach the build vs.buy decision in implementing AI?The build vs.buy paradigm while implementing AI includes a middle ground compose,where firms leverage modular building blocks to cu
190、stomize solutions.This is enabled by tools like low-code/no-code platforms(e.g.,Copilot Studio),allowing companies to create tailored agents and workflows without requiring deep technical expertise.Vendors are also simplifying user interaction by abstracting complexity,such as embedding meta-prompti
191、ng in the background.Investing in AI Implementation Expert View Q:How can we identify leaders and laggards in AI adoption?Leaders in AI adoption distinguish themselves through:1 Talent:Recruiting and retaining skilled professionals;2 Strategic prioritization:Embedding AI as a core organizational foc
192、us;3 Data readiness:Ensuring data is accessible/optimized for AI use;and 4 Investments in AI factories:Building full-stack solutions,including hardware,software platforms,SDKs,and libraries,to enable scalable AI use.The financial services sector is uniquely suited to AI,due to its heavy reliance on
193、data to access risk,returns and customer needs.Firms leveraging the above factors are not only experimenting,but deploying AI-powered solutions at scale,creating competitive differentiation and winning market share.Financial services is one of the most exciting industries for the application of AI b
194、ecause the industry is not chasing one silver bullet use case.Banks,asset managers,payments firms,and insurers are using AI Factories to build hundreds of AI enabled applications and that number will scale into the thousands.KEVIN LEVITT,GLOBAL INDUSTRY BUSINESS DEVELOPMENT,FINANCIAL SERVICES,NVIDIA
195、30 30 Acrew Capital and Money 20/20 Report,All in on AI:Financial Services Adoption Index,2024.Martin Moeller Head of AI&Generative AI for Financial Services,EMEA Microsoft Kevin Levitt Global Industry Business Development,Financial Services NVIDIA January 2025 Citi GPS:Citi GPS:Global Perspectives&
196、Solutions 2025 Citigroup 25 Agentic AI&Finance Jrgen Eckel,Managing Director&Partner,BCG X Q:How is agentic AI being explored in financial services?Agentic AI is emerging as a prominent topic in financial services,though widespread implementation is yet to be seen.While a handful of enterprises are
197、moving agentic architecture workflows into production,the vast majority are still focused on catching up with core AI advancements.Key opportunities for agentic AI lie in compliance,client onboarding,and fraud detection,where it can enhance efficiency and augment human expertise,and we are likely to
198、 see more experimentation in 2025.The shift is moving beyond personal productivity tools,towards deeper process integration.We are seeing financial institutions explore the use of foundation models in investment banking and wealth management.Q:What are the key approaches to develop agentic AI system
199、s?How do they differ across major market players?BCG research suggests that agentic AI is set to revolutionize 2025,with 32%of surveyed companies planning to integrate it into their core strategies this year.In my view,the market is currently exploring two paths for agentic AI development.One involv
200、es creating agentic systems by combining smaller,specialized language models with additional agentic layers that help facilitate decision-making,multi-step reasoning,and task execution.These systems are designed to address specific functions or processes within industries and offers the flexibility
201、to tailor the AI system to meet the unique needs of business functions.The other approach involves major foundation model developers embedding agentic behavior directly into their large foundation models.Q:What are the key factors influencing the speed of AI adoption in financial services?Which area
202、s present the most opportunities?Transitioning from proof-of-concept to production requires new development and operations processes,addressing algorithmic bias,and ensuring explainability.Regulatory complexities,such as differing AI laws across different markets further slows adoption.Looking ahead
203、 at 2025 and beyond,the compliance and regulatory space offers the greatest potential for AI,as it can automate labor-intensive tasks like interpreting regulations,and subsequently drafting initial reports.Financial firms are using AI to improve their own operations and make workflows smarter.In add
204、ition,they employ AI not just for investment sourcing and dynamic due diligence,but also to get better returns from their portfolio companies.Private equity(PE)firms are creating AI centers of excellence for use by their portfolio companies,for operational efficiencies,customer insights and more.Pri
205、vate Equity Firms Investing in“AI Transformation”Alex Sion,Financial Services Vertical Lead,Blend360 Historically,private equity firms have primarily focused on financial and operational restructuring to enhance portfolio company value.Now,PE is focusing on and investing in AI transformation as a pr
206、imary lever for portfolio companies value creation.The rise of the digital revolution began to evolve this focus.The emergence of AI is a major accelerant.Recognizing the critical role of technology in driving business growth and competitiveness,PE firms are increasingly investing in technology oper
207、ating partners,technology development and delivery partnerships,and prioritizing AI transformations in their value creation strategies.As AI continues to rapidly evolve and scale,the ability of PE firms to effectively support their portfolio companies through their AI transformation journeys will be
208、come paramount for achieving successful investment outcomes.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 26 Compliance&Regulatory Agents The integration of agentic AI in financial services can help improve compliance and regulatory analysis.For example,in trade finance
209、,tasks such as evaluating transactions flagged for sanctions involve significant manual effort,requiring data gathering across multiple structured and unstructured sources like transaction records and international company registers.Agentic AI can help streamline these processes by codifying workflo
210、ws,conducting multi-agent comparisons,and providing summarized recommendations for human review,thereby saving time,and improving efficiency.Compliance Agents in KYC and AML Screening Simon Taylor,Head of Strategy&Content,Sardine In my view,scams and the criminals who use scams as their funding mech
211、anism,will become the biggest issue for banks and financial services in 2025.How can AI and agentic AI help us reduce or prevent scams and frauds?Compliance head count is a big cost for most banks.The scope and volume of sanctions alerts has rocketed.The industry standard is a 95%false positive rati
212、o.A big global bank needs enormous headcount to review a lot more alerts when 95%is wasted effort.And while theyre doing that,they might be missing true positives.Focusing on far better data collection,labelling,and then machine learning to identify true positives can save massive amounts of time an
213、d money,while maintaining or improving the true positive detection rate.As banks migrate to the cloud,theyre also getting a massive data dividend.This would have been unthinkable even a decade ago.The art is understanding the workflow well enough to build an AI-product or workflow that is consistent
214、,repeatable,and value add.Our Sanctions AI Agent will pull together all customer KYC and AML data,review rules and third-party data,and make a pre-decision about if a user should be added to a blocklist,allowlist or pushed for additional KYC screening.We cross-reference that with data from bank cons
215、ortia,government sources,telcos,email providers,watchlists,and sanctions lists.This feeds over 4,000 machine-learned features,ranging from as simple as what country is a user really in,or is this email fraudulent,to complex aggregations of multiple inputs.These complex calculations can then roll up
216、to a simple score and help make a decision(block/allow a transaction)or be made available to a banks fraud team to enhance what they do.Converting Regulations into Computer Programs John Nay,Founder&CEO,Norm Ai Growth of regulatory complexity has made it difficult for companies to comply with the ro
217、utine thousand-page regulation in a time-sensitive and cost-effective way.Regulatory AI makes it possible to convert these complicated regulations into AI agents that are capable of compliance determinations.What Can AI Agents Do?Regulatory AI agents enable compliance and business teams to assess wh
218、ether proposed content or decisions are likely compliant with relevant regulations and guidelines.They serve as a round of checks,highlighting potential areas that may be problematic and need human intervention.Additionally,regulatory AI agents can work alongside other AI agents,ensuring business ac
219、tions taken by other AI agents are also in compliance with regulations.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 27 Deepfakes&Frauds The Global Anti-Scam Alliance estimated that in 2023 the economy lost$1 trillion to scams.The UKs National Crime Agency says that sca
220、ms are now the#1 form of all crime,making up an estimated 55%of crimes,and 45%of reported crime.Nation states and large-scale criminals have become proficient at exploiting the system.In the US nearly 30%of consumers or households have been victim to a scam in the last five years.31 The GenAI capabi
221、lities to morph text,audio and video has exceeded exponentially in the recent times and the propensity to commit and the quantum of scams and frauds is likely to increase.Modern synthetic audio has become so sophisticated that it can make it difficult for human contact center agents and even biometr
222、ic solutions to distinguish between authentic and synthetic voices.Today,audio has already crossed that uncanny valley in terms of being able to be imperceptible to the human ear.Video is also quickly improving.Agentic AI will lead to the mass production and distribution of deepfakes.In finance,this
223、 could be used to manipulate transactions,create synthetic identity fraud,and automate scamming.Deepfake scams have increased more than 2000%over the last three years and banks/FinTechs are one of the most targeted victims of deepfake attacks.32 Detecting&Defeating Deepfakes Mason Allen,Head of Grow
224、th,Reality Defender In 2025,we expect audio and video-based impersonations to increase due to the democratization of Generative AI(GenAI)technologies enabling bad actors to impersonate known individuals such as colleagues,customers,family members,politicians,actors,and influencers.From banks perspec
225、tive,the risk and implications from threats of GenAI-based impersonations are tangible today and will accelerate this year.Financial institutions must protect themselves,their employees,and their customers from the next generation of fraud and social engineering attempts using these new tools that a
226、re more believable and scalable than previous methods.By enabling the infrastructure to analyze all calls and videos for synthetic deepfakes,enterprises will be able to mitigate against a risk they cannot detect today.One of the worlds largest financial institutions found itself grappling with this
227、emerging security challenge within its private banking division.As generative AI technologies underwent rapid advancement,the banks high-net-worth clients,who were prominent public figures whose voices and images are readily available online,became increasingly susceptible to sophisticated deepfake
228、impersonation attempts.How did Reality Defender analyze calls and identify deepfakes?When a call is placed to the contact center,the detection process happens seamlessly without introducing latency or disrupting the natural flow of conversation.As the audio stream flows through the clients existing
229、telephony infrastructure,its simultaneously routed to a Reality Defender secure audio processing pipeline within the banks environment.Real-time streaming ensures that the conversation is analyzed without creating delays or requiring post-call processing.AI models analyze the audio stream,examining
230、multiple aspects of the call for signs of synthetic manipulation or generation.A recent analysis shows 37 calls per week(0.17)were identified as potentially manipulated,demonstrating both the real presence of deepfake threats and the systems precision in detecting them.In 2025 and beyond,the way fin
231、ancial services manage fraud and scams will be fundamentally different because of the way fraudsters use AI to commit frauds.We will have an AI arms race between attackers and defenders.Some of the attackers have state-level resources and sophistication.31 PYMTS,The Impact of Financial Scams on Cons
232、umers Finances and Banking Habits,23 October 2024.32 Security.org,2024 Deepfakes Guide and Statistics,26 September 2024.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 28 Onboarding&KYC Agentic AI has the potential to revolutionize traditionally time-consuming and resourc
233、e-intensive onboarding and Know Your Customer(KYC)processes,turning them into seamless and efficient experiences,potentially increasing client satisfaction with a process often disliked by users.In financial services,onboarding individuals for banking,or small merchants for payment processing,tends
234、to require extensively manual and repetitive processes.New digital entrants,such as FinTechs,have improved onboarding.But AI-powered automation can significantly streamline incumbents workflows.For instance,automated document verification systems can help analyze and cross-reference identification d
235、ocuments with external databases to ensure authenticity,reducing the need for manual checks.In parallel,AI-powered tools like facial recognition and biometric systems can provide secure and efficient identity verification,enhancing customer experiences while minimizing fraud risk.Agentic AI also exc
236、els in processing structured and unstructured data,enabling intelligent data gathering and analysis.This allows AI agents to perform dynamic risk profiling,assessing customer risk levels and ensuring compliance with regulations on Anti-Money Laundering(AML)and Combating the Financing of Terrorism(CF
237、T).Research indicates that nine out of ten KYC or digital onboarding systems with liveness detection can be bypassed.33 However,agentic AI offers a more robust approach by analyzing contextual signals,such as how a device is being held,or whether unauthorized software is running.This deeper level of
238、 scrutiny enhances security and helps prevent sophisticated fraud attempts.Onboarding Agents Greg Ulrich,Chief AI&Data Officer,Mastercard Mastercard,in collaboration with Databricks,launched its first of a suite of GenAI based digital assistants designed to speed up the onboarding of Mastercards pro
239、ducts for customers and partners.Mastercards new in-house capability was built in collaboration with Databricks on its Data Intelligence Platform;this platform includes core functionalities like a chat-based assistant user interface,source management and learning with human feedback.The program ensu
240、res that new knowledge agents are continuously trained on Mastercards trusted,proprietary data sets while always operating under the companys AI and data governance principles and standards.The effort also removes the friction associated with vetting and procuring third-party solutions,further strea
241、mlining operations.The product onboarding assistant automates routine tasks and answers customers critical questions during onboarding by utilizing a large language model with Retrieval Augmented Generation(RAG)and fine-tuning.Using Mastercards existing onboarding documentation as its knowledge base
242、,RAG locates the accurate information needed to respond to user prompts.The tool also employs a human-in-the-loop approach to integrate feedback from subject matter experts,reinforcing continuous learning and ensuring accuracy in the agents responses.33 ID Tech,Nine of the Top 10 Liveness Detection
243、Systems are Vulnerable to Deepfakes,20 May 2022.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 29 Agentic Contracts Agentic contracts are programmable and enforceable contracts that can generate text and execute tasks autonomously based on how they have been programmed.A
244、gentic AI powers smart legal contracts by automating analog legal contracts.Agentic contracts take analog pain points out of the contract lifecycle system.Smart legal contracts(SLCs)are a subset of smart contracts and legal contracts.We discussed this in detail in Citi GPS:Money,Tokens&Games-Blockch
245、ains Next Billion Users and Trillions in Value(March 2023).Smart contracts can be used to define and perform the obligations of a legally binding contract.Agentic AI becomes transformative by executing the terms of the contract in the post-signature phase of the contract.This includes monitoring com
246、pliance,automating payments,and managing regulatory changes.The integration with distributed ledger technology ensures transparency and predictability.Executing the terms of the contract does not mean no human in the loop.Certain tasks still require human oversight due to regulatory requirements and
247、 the need to assign liability.Agentic AI handles most tasks but defers final approvals to humans.The anatomy of an agentic contract,its advantages,and the stakeholders it benefits are depicted in Figure 15 and Figure 16.Figure 15.Benefits of Agentic Maritime Private Credit Contract Figure 16.Agentic
248、 Contracts Benefit Source:Hunit Source:Hunit Agentic AI Use Case in Private Market Asset Finance Aaron Powers,Co-founder&CEO,Hunit Agentic contracts will move private market asset finance from the current world of PDF documents to one where stakeholders can interrogate if the terms of the legal agre
249、ement are being complied with in real-time and where valuations can be more standardized(which will become more important under Basel IV).A key benefit of investment instruments based on this technology is that they can be continually and algorithmically valued,and risk controlled.Risk mitigation st
250、rategies can be triggered based on upstream operational data,weeks or months before they impact cash flows or covenant compliance.Algorithmically monitoring the output from a wind power generation site,or kilometers driven by a fleet of leased commercial vehicles,allows for insight-driven risk manag
251、ement.In the maritime industry,data from the ship itself can be monitored to ensure compliance with environmental and other regulations(e.g.the ship is not going into sanctioned waters).AI-POWERED CAPABILITIES Auto-generate environmental and operational reporting.Monitor external regulatoryenvironme
252、nt and verify ongoing contract compliance against changing needs.Assist borrower with ongoing covenant compliance.Review and analyze covenant reporting on behalf of Lender Monitor vessel location and engage with owner if entering sanctioned waters.PROGRAMMATIC CAPABILITIES Create real-time transpare
253、ncyover cashflows by clearingpayments using contract embedded bank facilities.Remove manual processes by retrieving central bank interest rates,adjusting loan schedule and calculating borrower payments.Retrieve and store key operational data.Ensure KYC/AML verificationat onboarding and periodically
254、over contract term.ONBOARD DATA VAULTOPERATIONS Reduce costs,risks and manual work,even for bespoke deals.Eliminate information silos through real-time clarity into contracts in motion.Analyze legal portfolios for new insights.COMPLIANCE Legally certain records provide audit trail for environmental
255、reporting.Scalable access to data need for regulatory capital stress testing and reporting.Ensure counterparties have adequate KYC/AML.INVESTMENT MANAGEMENT Apply automated valuation and risk models continually and at scale(Basel IV).Manage risk better with real-time data on investment performance.F
256、ractionalize or transact beneficial ownership based on self-contained architecture.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 30 Corporate Treasury&Payments Digitization has transformed payments over the decade,and agentic AI on top of LLM architecture could supercha
257、rge payments.For example,Stripe enables AI agents and LLM providers to process payments to invoice customers,generate payment links,or issue virtual cards.34“Agentic treasurers”will employ machine learning,predictive analytics,and natural language processing to streamline processes like cash forecas
258、ting,fraud detection,and payment scheduling.Citi GPS:Treasury 2030 Modernize or Risk Irrelevance(December 2024)highlights that AI could bring one of the largest technological shifts in treasury.AI agents can automate the process of finding and reallocating corporate funds to the highest-yielding opp
259、ortunities within the firm,addressing market discrepancies and managing repetitive tasks such as cash forecasting,reconciliation,and compliance checks.This will free up human time to focus on liquidity risk management,and to create a scalable corporate treasury.35 Corporates operating in multiple co
260、untries usually bank with more than one bank.A global bank could leverage AI as a rule engine to generate recommendations to the treasury banker based on,e.g.,a clients data,balance sheet size,FX exposure,liquidity,and capital needs on how the company optimizes its working capital to improve return
261、on invested capital(ROIC).Local laws permitting,steps in an entire liquidity optimization service could be automated so that bankers and clients only get involved in the final approval/rejection stages.AI Driven Cash Flow Forecasting Jrg Wiemer Co-founder,TIS(Treasury Intelligence Solutions)AI agent
262、s can transform treasury functions and several other Office of the CFO functions by generating actionable insights from raw data,enabling treasurers to automate key decisions and collaboration within teams.AI improves accuracy in cashflow forecasts and generates insights in variance analysis(between
263、 forecasted and actual cashflows),addressing the“why”behind discrepancies.Agentic AI sitting on top of other AI/ML tools helps automate reconciliation statements for financial transactions reducing time and manual effort required.Identifying patterns manually can be time consuming and complex but an
264、 important precursor to accurately forecast cashflows.AI tools can help treasurers analyze their clients payment behavior and detect changes in specific behavior that can influence cashflow forecasts.Agentic AI could source data from multiple databases and models to execute on the output generated b
265、y other AI models.Figure 17.General Workflow Methodology in Building an Account Receivables Forecast Source:TIS(Treasury Intelligence Solutions)34 Stripe,Adding Payments to your LLM Agentic Workflows,14 November 2024 35 Deloitte,Global Treasury Survey,12 September 2024 4.VALIDATE FORECASTSCompare th
266、e forecast to the actuals and reiterate.3.FORECAST RULESe.g.,invoices are forecast on due date.2.IDENTIFY PATTERNSe.g.,a customer seems to always pay the Thursday after due date.1.ENRICH DATACounterparty,invoice,business unit info.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Cit
267、igroup 31 Wealth&Agentic Advisers AI agents can help transform operations in wealth management by streamlining client acquisition and onboarding.This can include automating parts of the due diligence processes such as document analysis or extraction.In client servicing,AI agents can stimulate human-
268、like conversations to answer client queries,explain investment options and provide 24x7 guidance.For high-net-worth clients,AI agents can be an important augmentation to the human adviser,making the latter smarter and better informed.AI agents can also monitor financial market trends,macroeconomic f
269、actors,and geopolitical events in real-time,generating insights and automating investment recommendations for clients.These insights can be provided directly to the client digitally or via a human adviser.Beyond investment advisory,agentic AI can streamline the management of customer relationship ma
270、nagement(CRM)systems.Agents can update CRM records with detailed notes,contact information and more,without manual intervention.Boosting Efficiency in Wealth and Asset Management Steve Blanchet,Head of Group Technology Strategy and Innovation,Pictet Group Manuel Grenacher,CEO,Unique FinanceGPT(www.u
271、nique.ch)An illustrative example for the use of AI in wealth management is One.Chat,an AI platform co-developed by Unique and Pictet,that allows easy access to internal information and finance-specific use cases.One.Chat is a platform that offers various GenAI capabilities.In addition to directly ac
272、cessing the latest Azure OpenAI models securely,it has an internal search chatbot for information retrieval and summarization generating specific responses to queries after reviewing relevant documents and summarizing multiple large documents to generate actionable insights.The platform is offered t
273、o 6,000+Pictet employees(4,000+monthly active users),generating 50,000+prompts per week,and is estimated to save 1.5-2 hours per employee per week.36 Within One.Chat,Relationship Managers(RMs)have also access to their own GenAI assistant,which enables streamlined access to information from various s
274、ources to better and more efficiently engage with clients.For instance,it can quickly filter out recommended products from the full investment product universe using natural language by mentioning any number of criteria such as bond maturity,yield,ratings,etc.While human RMs continue to be the core
275、of the client relationship,the GenAI assistant helps them be more efficient,which ultimately benefits clients.Employees can also use GPTs coding functions for software development,alongside other operational tools such as translation of information in multiple languages and drafting emails.36 Unique
276、 AG,Pictets One.Chat:Boosting Efficiency for Wealth and Asset Management,05 January 2025.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 32 Wealth management serves an incredibly diverse clientele,spanning young,self-directed investors exploring stocks and digital assets
277、to ultra-high-net-worth individuals and global change-makers shaping the worlds future.Advisors face the complex challenge of providing deeply personalized and high-touch services at scale.With the rapid evolution of AI technology,wealth managers can bridge these diverse needs by delivering tailored
278、 opportunities and insights,transforming wealth management into a scalable,inclusive,and accessible model for all clients,regardless of their financial journey.JOSEPH BONANNO,GLOBAL HEAD OF DATA ANALYTICS&INNOVATION,CITI WEALTH January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigr
279、oup 33 Hyper-Personalization While GenAI drives hyper-personalization,agentic AI drives the execution of hyper-personalization by analyzing vast amounts of data in real-time to offer tailored experiences and recommendations.These systems can adapt to individual preferences,behaviors,and needs,enabli
280、ng businesses to create personalized products and services,marketing campaigns or customer interactions.Agentic Marketing Automation to Deliver Hyper-personalized Consumer&Merchant Insight Alex Sion,Financial Services Vertical Lead,Blend360 We are witnessing substantial investment from financial ser
281、vices firms to achieve hyper-personalization across digital experiences and business intelligence.Furthermore,these hyper-personalization efforts are driving the development of novel platforms,products,and services.This enables financial services firms to pursue new growth as influential marketing a
282、dvisers and even establishing themselves as prominent media platforms.Figure 18.Hyper-personalization to Deliver Consumer and Merchant Insights Source:Blend360 Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 34 Loan Documentation&Credit Workflow AI agents can help streaml
283、ine processes,enhance accuracy,and reduce manual effort in loan documentation and credit workflows.AI-based automated document analysis can analyze and extract key information from complex loan documents,financial statements,contracts,and collateral agreements.Leveraging capabilities to analyze vast
284、 structured and unstructured datasets,AI agents can automate evaluation of borrower creditworthiness,adapting to new data in real-time.Agentic AI systems can help manage multi-step credit workflows,from application intake to approval.Agentic AI can identify specific tasks within a credit request and
285、 assign them to domain specific agents(models).These models perform detailed financial analysis,risk assessment,and compliance checks.Humans can provide real-time feedback e.g.,refining financial interpretations or adjusting language which the system incorporates dynamically to improve output.Small
286、Language Model(SLM)Powered Agentic AI Workflow Brian Benedict,Co-founder&CRO,Arcee AI Small language models(SLMs)combined with agentic AI offer a transformative solution that deliver an efficient,secure system tailored to specific tasks.An SLM-powered agentic AI platform takes the capabilities of SL
287、Ms to the next level by enabling enterprises to build custom workflows that handle multi-step tasks with precision and speed.These platforms eliminate repetitive work while delivering higher-quality output faster than traditional methods.This is achieved through a combination of SLMs,each specializi
288、ng in different areas such as coding,data summarization,content creation,vision-language processing,or reasoning.Intelligent routing mechanisms ensure that each task is assigned to the most suitable model based on its specialization.Once individual tasks are completed,their outputs are consolidated
289、into a unified response.This adaptive and modular approach makes agentic AI particularly well-suited for automating complex,data-driven workflows typically seen in financial services that require dynamic decision-making and continuous learning.Below is an example of how Arcee AIs systems can help au
290、tomate credit memo drafting,including data collection,financial analysis,risk assessment,and generation of a draft memo.Figure 19.Comparison of Traditional Process vs.Agentic AI Network for Credit Memo Generation Source:Arcee AI DATACOLLECTIONFINANCIALANALYSISRISKASSESSMENTMEMODRAFTINTERNAL REVIEWFI
291、NALAPPROVALManualGatheringSpreadsheetWorkManual DraftingCommitteePrepTRADITIONAL AS-IS PROCESSARCEE ORCHESTRA:AGENTIC NETWORKTASK DECOMPOSITION&ROUTING AGENTDATA COLLECTIONAGENTFINANCIALANALYSIS MODELMEMO DRAFTING MODELINTERNALREVIEWFINALAPPROVALRISK ASSESSMENTMODELCOMPLIANCEMODELClient DataExternal
292、 DataInternal SystemsData SourcesRevisionsDOMAIN SPECIALIZED SMALL LANGUAGE MODELS(SLMS)RevisionsJanuary 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 35 Insurance&Agentic Claim Management A standard medical or general insurance claim involves many small checkpoints that contri
293、bute to the whole claim management process.Agentic AI could interact with multiple models to make autonomous decisions within defined constraints.AI agents could handle initial claim reporting,gather necessary details,and help triage claims based on severity and urgency.Initial claim validation requ
294、iring the checking of policy details and moving on to next steps could be defined and automated as part of initial insurance claim management workflow.Agentic AI is a superpower for finance professionals.As an example,an organization that processes tens of thousands of claims daily can bring down th
295、e resolution time from months or weeks to just days with a handful of senior claim approvers to monitor the output and make decisions.KEVIN CHUNG,CHIEF STRATEGY OFFICER,WRITER For insurers,multi-agent systems can gather real-time data from internal and external sources and combine them to provide a
296、complete view of each customer and risk profile.AI systems can automate underwriting by assessing risk profiles and determining appropriate policy terms and pricing without human intervention.The proliferation of smartphones and wearables enable insurers to gather a variety of claim-related data,loc
297、al laws permitting.In Citi GPS:AI in Finance:Bot,Bank&Beyond(June 2024),we discuss how the availability of more data,often underpinned by AI,can allow insurers to segment risks more granularly.While this can help make pricing more risk-based,such micro-segmentation could also result in large number
298、of smaller risk pools.Agentic Claim Management Ren Ito,Co-founder&COO,Sakana AI Insurance processing involves an intensive repetitive manual process.Historically,everything was manual,until the introduction of robotic process automation(RPA).Now we have agentic middleware to improve the automations
299、and efficiencies achieved by RPA.AI agents could transform insurance claim management by automating key processes such as claim intake,damage assessment,fraud detection,and settlement.The workflows could be designed to gather and validate information,analyze photos or videos to estimate damages,and
300、cross-check claims against policy terms.For straightforward cases,AI can approve or reject claims instantly and trigger payouts,while flagging complex or suspicious claims for human review.Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 36 Investment Research In Citi GPS:
301、AI in Finance:Bot,Bank&Beyond(June 2024),we elaborate on how GenAI in investment research could enable analysts to spend more time interacting with clients AI tools help with repetitive tasks.Investment research consists of synthesizing a lot of information and data and generating differentiated ins
302、ights.GenAI tools are designed to ingest vast amounts of data and synthesize them.Agentic AI can potentially extract more value from these models without relying on more data,making existing models more actionable and application oriented.In the investment research and due diligence process,various
303、market data models and LLMs could be used to analyze and ingest a broad range of data into spreadsheets and summarize narratives from published text and transcripts.This could include searching,extracting,and summarizing information from news,research reports,financial statements,and other materials
304、.This can save researchers time and equip them with more information in a shorter time frame.The AI agents can populate documents or reports such as investment proposals for use in investment decision making.Agentic AI can also aid researchers and portfolio managers in helping narrow down a large fu
305、nnel of investment targets,through constructing queries and then analyzing the data to find the investments that best fit certain criteria.In a nutshell,a senior analyst would have a group of agent associates.In portfolio construction,agentic AI could be used to provide a list of investments ideas o
306、r suggestions to rebalance a portfolio based on market information.It could also aid personalization of portfolios by taking into account client preferences expressed in calls.Agentic AI could help summarize client portfolios,ensuring that portfolio managers are up to date.Agentic AI could also eval
307、uate the portfolio against client preferences and suggest rebalancing where needed,aiding the creation and maintenance of tailored portfolios.It can also contribute to customized client reporting.Agentic Investment Research Kevin Chung,Chief Strategy Officer,Writer Agentic AI is being leveraged in i
308、nvestments for many tasks such as creating markets research reports,press releases,and thought leadership while maintaining a companys brand,accuracy,and analysts specific writing style.This is further amplified by connecting to any type of structured and unstructured data,including historical repor
309、ts,market performance and internal resources to generate highly relevant and specific content in seconds.As an example,the AI associate(agent)could produce a quick note based on an industry breaking news,which the analyst can quickly review,fact check and get it published and distributed.The analyst
310、 could be on call interacting with clients within minutes of the news being announced rather than spending the hours getting the written note out.For periodicals and longer form research products,the analyst could use the AI agent to perform various tasks such as conducting all the analysis to gener
311、ate insights,updating models and spreadsheets,and even running different simulations by plugging in different assumptions.January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 37 What are Agents?The term agent refers to an entity or individual authorized to act on behalf of ano
312、ther,known as the principal,to carry out transactions or decisions.The concept of agency is deeply rooted in economic theory and practice,dating back to the emergence of trade and commerce.Historically,agents were human intermediaries,such as brokers,advisers,or managers,who facilitated transactions
313、 between buyers and sellers or managed assets on behalf of their clients.These agents played a pivotal role in building trust,enabling market efficiency,and reducing transaction costs.The advent of electronic trading in the late 20th century marked a transformation in the role of agents in finance.T
314、raditional human agents began to be supplemented(in some cases replaced)by digital systems and automated processes.The development of algorithmic trading in the early 2000s further shifted the landscape.Bots operating on simple rules have existed for some time.The next generation of agents are suppo
315、sedly smarter.They can reason,learn,and make decisions,but how do they work?How Do AI Agents Work?Agentic AI combines perception,reasoning,and action to autonomously interact with its data environment and achieve defined goals,making it capable of handling repetitive manual tasks.Agentic AI operates
316、 through a structured workflow involving several core components.Key steps are depicted in Figure 20.Figure 20.An Illustration of How Agentic AI Works37 Source:Nvidia,Citi Global Insights 37 NVIDIA,What is Agentic AI?,22 October 2024.UserAI AgentLLMData FlywheelModel CustomizationDatabaseVector Data
317、baseAction Citi GPS:Citi GPS:Global Perspectives&Solutions January 2025 2025 Citigroup 38 Building Blocks for Agentic Workflows The Augmented LLM:The foundational building blocks for an agentic AI system is an LLM enhanced with key augmentations including retrieval,tools,and memory.These augmentatio
318、ns allow the AI to move beyond static responses and access information,execute specific tasks,and support contextual continuity for more dynamic and personalized interactions(Figure 21).Workflow:This represents structured sequences of steps or actions that enable the system to achieve specific goals
319、 or solve a problem.Workflows can be found in different types such as prompt chaining,routing,parallelization,orchestrator workers etc.,depending on the varieties of task and complexities.Figure 21.The Augmented LLM 38 Figure 22.Orchestrator-Worker Workflows 39 Source:Anthropic,Citi Global Insights
320、Source:Anthropic,Citi Global Insights Figure 22 illustrates an orchestrator-worker workflow where a central LLM dynamically breaks down tasks,delegates them to worker LLMs,and synthesizes their results.Orchestrator-worker workflows are typically suitable for complex tasks that may involve gathering
321、and analyzing information from multiple sources.Agents:These refer to intelligent systems or programs designed to understand complex inputs,engage in reasoning,make decisions,or achieve goals on behalf of a user.Figure 23 illustrates an autonomous agent,which is typically a LLM using tools based on
322、environmental feedback in a loop that can take actions independently within predefined parameters.Figure 23.Autonomous Agent 40 Source:Anthropic,Citi Global Insights 38 Anthropic,Building Effective Agents,20 December 2024.39 Anthropic,Building Effective Agents,20 December 2024.40 Anthropic,Building
323、Effective Agents,20 December 2024.INOUTLLMTOOLSRETRIEVALMEMORYQuery/ResultsCall/ResponseRead/WriteLLM CALL2LLM CALL1LLM CALL3INOUTORCHESTRATORSYNTHESIZERHUMANENVIRONMENTLLM CallSTOPActionFeedbackAs our premier thought leadership product,Citi Global Perspectives&Solutions(Citi GPS)is designed to help
324、 readers navigate the most demanding challenges and greatest opportunities of the 21st century.We access the best elements of our global conversation with senior Citi professionals,academics,and corporate leaders to anticipate themes and trends in todays fast-changing and interconnected world.All Ci
325、ti GPS reports are available on our website The Rise of AI Robots Physical AI is Coming for You December 2024 Treasury 2030 Modernize or Risk Irrelevance December 2024 Nature Sustainabilitys Next Frontier November 2024 Global Trade in Flux Politics,Policy and the Reconfiguration of Supply Chains Nov
326、ember 2024 Obesity and Beyond The surprising reach of GLP-1s October 2024 Cross Border Payments 24x7 Faster,Simpler,Smarter October 2024 Global Giving Generosity and the Economics of Philanthropy October 2024 Sustainable Transitions:Unleashing the Power of Treasury Unleashing the Power of Treasury A
327、ugust 2024 Hard to Abate Sectors and Emissions II The Road to Decarbonization July 2024 Embracing Demographic Transition Health and Wealth in an Aging World July 2024 AI in Finance Bot,Bank&Beyond June 2024 What Machines Cant Master Human Skills to Thrive in the Age of AI May 2024 Disruptive Innovat
328、ions X Ten More Things to Stop and Think About May 2024 Future of Healthcare The Oxford Prescription May 2024 Closing the Racial Wealth Gap Restating the Economic Case April 2024 Supply Chain Financing Building Resilience January 2024 Unlocking Climate and Development Finance Creating Bankable Proje
329、cts December 2023 Food,Water,and Climate Change Solutions for Tackling this Critical Nexus November 2023 APEC Economic Outlook Navigating Shifting Dynamics November 2023 Supply Chains Lessons Learned Helped by New Technology November 2023 Food Security Solutions for a Perfect Storm November 2023 Tre
330、asury Leadership Does It Matter?November 2023 Philanthropy and the Global Economy v3.0 Perspectives on the Future of Giving October 2023 Supply Chain Technology Improving Efficiency,Resilience,and Transparency October 2023 Technology at Work v7.0 The Third Phase of Globalization October 2023 Africa
331、A New Growth Model October 2023 Skills That Pay The Returns from Specific Skills October 2023 Future of Cross-Border Payments Who Will Be Moving$250 Trillion in Next Five Years?September 2023 Unleashing AI The AI Arms Race September 2023 Hydrogen A Reality Check on the Hydrogen Craze August 2023 Qua
332、ntum Computing Moving Quickly From Theory to Reality July 2023 Future of Nuclear Energy in a Low-Carbon Environment Fission and Fusion Advanced Reactors to Prevail July 2023 Voluntary Carbon Market A Critical Piece of the Net Zero Puzzle July 2023 Economic and Social Mobility The Role of Business in
333、 Improving Outcomes June 2023 Sustainable Ocean Economy Charting a Prosperous Blue Future:Risk to Resilience June 2023 Asia as a Time Machine to the Future Seven Areas Where Asia Gives Insights Into the Future May 2023 January 2025 Citi GPS:Citi GPS:Global Perspectives&Solutions 2025 Citigroup 41 If you are visually impaired and would like to speak to a Citi representative regarding the details of