《Iconiq Capital:2025年人工智能状况报告:AI产品构建者策略手册(英文版)(67页).pdf》由会员分享,可在线阅读,更多相关《Iconiq Capital:2025年人工智能状况报告:AI产品构建者策略手册(英文版)(67页).pdf(67页珍藏版)》请在三个皮匠报告上搜索。
1、June 20252025 State of AI ReportPrivate and Strictly ConfidentialCopyright 2025 ICONIQ Capital,LLC.All Rights ReservedThe Builders PlaybookThe Builders PlaybookFor Professional Clients Only.ICONIQ Partners(UK)LLP(973080)is an appointed representative of Kroll Securities Ltd(466458)which is authorize
2、d and regulated by the Financial Conduct AuthorityPrivate&Strictly Confidential2Follow our researchFollow our researchSUBSCRIBENavigating Todays Public MarketsNavigating Todays Public MarketsThe metrics that matter and the market realities of 2025 and beyondThe SaaS GlossaryThe SaaS GlossaryA guide
3、to understanding and tracking key SaaS metricsThe ICONIQ Enterprise FiveThe ICONIQ Enterprise FiveKey performance indicators of Enterprise SaaS companiesGrowth&EfficiencyGrowth&EfficiencyExplore our research on best-in-class SaaS growth and efficiencyEngineering SeriesEngineering SeriesDefinitive gu
4、ides to engineering excellenceGoGo-ToTo-Market SeriesMarket SeriesGuides to sales,customer success,marketing compensation and morePrivate&Strictly Confidential3IntroductionIntroductionExplore Our AI PerspectivesExplore Our AI PerspectivesWe believe that building and operationalizing AI products is t
5、he new frontier of competitive advantage new frontier of competitive advantage and that the voices of the architects,engineers,and product leaders driving this work deserve their own spotlight.While last years State of AI report centered on the buying journey and enterprise adoption dynamics,our 202
6、5 report pivots squarely to the“how“how-to”:what it takes to conceive,to”:what it takes to conceive,deliver,and scale AIdeliver,and scale AI-powered offerings endpowered offerings end-to to-endend.This years report unpacks core dimensions of the builders playbook:1.1.Product Roadmap&ArchitectureProd
7、uct Roadmap&Architecture:The emerging best practices for balancing experimentation,speed to market,and performance at each stage of model evolution2.2.GoGo-to to-Market StrategyMarket Strategy:How teams are aligning pricing models and go-to-market strategies to reflect AIs unique value drivers3.3.Pe
8、ople&TalentPeople&Talent:Building the right team to harness AI expertise,foster cross-functional collaboration,and sustain long-term innovation4.4.Cost Management&ROICost Management&ROI:Strategies and benchmarks for spend associated with building and launching AI products5.5.Internal Productivity&Op
9、erationsInternal Productivity&Operations:How companies are embedding AI into everyday workflows and the biggest drivers of productivity unlockDrawing on our proprietary survey results alongside in-depth interviews with AI leaders across the ICONIQ community,the 2025 State of AI report offers a bluep
10、rint for anyone tasked with turning generative intelligence from a promising concept into a dependable,revenue-driving asset.Table of ContentsTable of Contents4Types of AI Products9Model Usage and Key Purchasing Considerations11Top Models Providers13Model Training Techniques14AI Infrastructure15Mode
11、l Deployment Challenges16AI Performance Monitoring17Agentic Workflows18Building Building Generative AI Generative AI ProductsProductsAI Product Roadmap20Pricing21AI Explainability and Transparency24AI Compliance and Governance25GoGo-to to-Market Market Strategy&Strategy&ComplianceComplianceDedicated
12、 AI/ML Leadership27AI-Specific Roles and Hiring28Pace of Hiring29%of Engineering team Focused on AI30Organization Organization StructureStructureAI Development Spend32Budget Allocation33Infrastructure Costs34Model Training Costs36Inference Costs37Data Storage&Processing Costs38AI CostsAI CostsIntern
13、al Internal ProductivityProductivityInternal Productivity Budget40Budget Sources41AI Access and Usage42Key Purchasing Considerations43Deployment Challenges44Number of Use Cases45Top Use Cases46Attitude Towards Internal AI Adoption48Tracking ROI49Top AI ToolsTop AI ToolsLLM&AI Application Development
14、51Model Training&Finetuning52Monitoring&Observability53Inference Optimization54Model Hosting55Model Evaluation56Data Processing&Feature Engineering57Vector Databases58Synthetic Data&Data Augmentation59Coding Assistance60DevOps&MLOps61Product&Design62Other Internal Productivity Use Cases63Private&Str
15、ictly Confidential5DataDataSourcesSources&Methodology&MethodologyThis study summarizes data from an April 2025 survey of 300 April 2025 survey of 300 executives at software companies executives at software companies building AI productsbuilding AI products,including CEOs,Heads of Engineering,Heads o
16、f AI,and Heads of Product.Throughout this report,we also weave in perspectives,insights,and what we believe to be best practices from AI leaders from the ICONIQ community.All industry perspectives shared in this report have been anonymized to protect company-level information.Respondent Firmographic
17、sRespondent FirmographicsNotes:(1)This data was collected anonymously by an external survey.Survey responses include some but not all ICONIQ Venture and Growth portfolio companies as well as companies not part of ICONIQ Venture and Growths portfolio.(2)Certain questions in the survey were optional.A
18、ccordingly,some N-Size numbers in this presentation are less than 30013%10%9%7%13%11%8%4%26%Revenue RangeRevenue Range88%12%North AmericaEuropeHeadquartersHeadquartersIn this report,select companies are referred to as“high growth companies”because they meet the following criteria AI Product Traction
19、AI Product Traction:AI product is in General Availability or ScalingRevenueRevenue:At least$10M in annual revenueTopline GrowthTopline Growth:100%+YoY revenue growth if$25M Revenue,50%+YoY revenue growth if$25M-250M Revenue,30%+YoY revenue growth if$250M+Revenue13%High Growth High Growth CompaniesCo
20、mpanies%of respondents20%25%55%Less than$100M$100-$200M$200M+Revenue RangeRevenue Range%of Respondents%of Respondents%of High-Growth RespondentsMost SaaS companies have evolved to add new AI capabilities and products;the following pages will dive into how AIMost SaaS companies have evolved to add ne
21、w AI capabilities and products;the following pages will dive into how AI-enabled and AIenabled and AI-native companies are approaching product developmentnative companies are approaching product developmentPrivate&Strictly Confidential6AI MaturityNotes:Representative Examples provided for illustrati
22、ve purposes only.Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our
23、community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkAIAI-Enabled:Creating a new(nonEnabled:Creating a new(non-core)AI productcore)AI productAIAI-Enabled:Adding AI Capabilities Enabled:Adding AI Capabilities to Existing Productsto
24、Existing ProductsTraditional SoftwareTraditional Software-asas-a a-ServiceServiceTraditional SaaSTraditional SaaSGenerative AI ProductsGenerative AI ProductsRepresentative Representative ExamplesExamplesAIAI-Native:Core product or business Native:Core product or business model is AImodel is AI-drive
25、ndrivenFocus of this reportFocus of this report31%of survey respondents31%of survey respondents37%of survey respondents37%of survey respondents32%of survey respondents32%of survey respondentsEmbedded AI-powered features into flagship offerings to boost automation,personalization,and end-user product
26、ivitywhile leaving underlying business model and UX largely intactStandalone AI-driven product or services alongside core product portfolio to explore adjacent use cases and revenue streamsEntire value proposition is architected around generative intelligence where model training,inference,and conti
27、nuous learning are the fundamental drivers of customer value and growthDelivery of subscription-based applications built around core business workflowsPrivate and Strictly ConfidentialCopyright 2024 ICONIQ Capital,LLC.All Rights ReservedBuilding GenAIBuilding GenAIProductsProductsAIAI-native compani
28、es are further along in the development cycle compared to AInative companies are further along in the development cycle compared to AI-enabled peers,with around 47%of products enabled peers,with around 47%of products analyzed having reached critical scale and proven market fitanalyzed having reached
29、 critical scale and proven market fitPrivate&Strictly Confidential8Stage of Primary AI ProductSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,C
30、TOs,our Technical Advisory Board,and others in our networkStage of Primary AI ProductStage of Primary AI Product%of Respondents,N=29111%1%34%10%42%42%13%47%AI-EnabledAI-NativePrePre-LaunchLaunchThe product is still in development and not officially available to external usersScalingScalingThe produc
31、t has proven market fit and is now focused on growing its user base and infrastructure to handle higher demandBetaBetaThe product is sufficiently developed to be tested by a limited group of external users for feedback and bug identificationGeneral AvailabilityGeneral AvailabilityThe product is form
32、ally released with the stability and support expected for broad adoptionOnly 1%of AI1%of AI-native companiesnative companies are still in pre-launch,compared to 11%of AI11%of AI-enabled companiesenabled companies.Meanwhile,while not surprising to see that 47%of AI47%of AI-native products are already
33、 scalingnative products are already scaling,this may imply AI-native companies are moving faster through the product lifecycle and achieving traction earlier.This begs the question whether AI-native orgs may be structurally better equipped-through team composition,infrastructure,or funding models-to
34、 validate product-market fit and scale effectively,and perhaps leapfrogging the leapfrogging the trialtrial-andand-error phaseserror phases that slow down AI-enabled companies retrofitting AI into existing workflows.Agentic workflows and the application layer are the most common types of products be
35、ing built across AIAgentic workflows and the application layer are the most common types of products being built across AI-native and AInative and AI-enabled companies;notably,around 80%of AIenabled companies;notably,around 80%of AI-native companies are currently building agentic workflowsnative com
36、panies are currently building agentic workflowsPrivate&Strictly Confidential9Types of AI ProductsSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprise
37、s,CTOs,our Technical Advisory Board,and others in our networkWhat type of AI products are you building?What type of AI products are you building?%of Respondents,Select All That Apply,N=291AI-NativeAI-Enabled79%65%56%55%48%62%57%49%40%27%Agentic workflowsVertical AI applicationsHorizontal AI applicat
38、ionsAI platforms/infrastructureCore AI models/technologiesi.e.focused on specific industry or functionMost companies building AI applications are relying on thirdMost companies building AI applications are relying on third-party AI APIs;however,a larger proportion of highparty AI APIs;however,a larg
39、er proportion of high-growth growth companies are also finetuning existing foundation models and developing proprietary models from scratchcompanies are also finetuning existing foundation models and developing proprietary models from scratchPrivate&Strictly Confidential10Model UsageSource:Perspecti
40、ves from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkHow does your company use AI models?How does your co
41、mpany use AI models?%of Respondents,N=26580%61%32%71%77%54%Rely on third-party AI APIsFine-tune existing foundation modelsDevelop proprietary models from scratchHigh Growth CompanyOther RespondentsA greater percentage of later stage companies later stage companies($100M+revenue)tend to develop propr
42、ietary($100M+revenue)tend to develop proprietary models or finemodels or fine-tune existing foundation modelstune existing foundation models,likely due to greater resources and need for enterprise customizationWhen choosing foundational models for customerWhen choosing foundational models for custom
43、er-facing use cases,companies prioritize model accuracy above all facing use cases,companies prioritize model accuracy above all other factorsother factorsPrivate&Strictly Confidential11Top Considerations for Foundational Models:Product DevelopmentSource:Perspectives from the ICONIQ GenAI Survey(Apr
44、il 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkTop Considerations When Choosing a Foundational ModelTop Considerations When Choosing a
45、Foundational Model%of Respondents who ranked each aspect in Top 3,N=265In last years State of AI report,cost,cost ranked as the lowest key purchasing ranked as the lowest key purchasing considerationconsideration in comparison to other factors like performance,security,customizability,and control.No
46、tably,cost is much higher in this years data much higher in this years data perhaps echoing the commoditization of perhaps echoing the commoditization of the model layer with the rise of more the model layer with the rise of more costcost-efficient models like DeepSeek.efficient models like DeepSeek
47、.74%74%Ability to fineAbility to fine-tune/customizetune/customizePrivacyPrivacyLatencyLatencyModel transparency/explainabilityModel transparency/explainabilityInference efficiency/compute requirementsInference efficiency/compute requirementsSOC2/Enterprise SLAsSOC2/Enterprise SLAsOpen SourceOpen So
48、urceVendor lockVendor lock-in/portabilityin/portability5757%41%41%34%34%25%25%19%19%18%18%14%14%9%9%6%6%CostCostAccuracyAccuracyOpenAIs GPT models continue to be the most popular model;however,many companies OpenAIs GPT models continue to be the most popular model;however,many companies are increasi
49、ngly adopting a multiare increasingly adopting a multi-model approach to AI products across use casesmodel approach to AI products across use casesPrivate&Strictly Confidential12Top Model ProvidersSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and n
50、etwork of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkTop Model ProvidersTop Model Providers%of Respondents,Select All That Apply,N=240Full Stack1Horizontal ApplicationVertical ApplicationCompa
51、nies are increasingly adopting a multimulti-model approach to AI productsmodel approach to AI products,leveraging different providers and models based on use on use case,performance,cost,and customer case,performance,cost,and customer requirementsrequirements.This flexibility enables them to optimiz
52、e for diverse applications like cybersecurity,sales automation,and customer service while ensuring compliance and superior user ensuring compliance and superior user experience across regionsexperience across regions.Architectures are being built to support quick Architectures are being built to sup
53、port quick model swapsmodel swaps,with some leaning toward open-source models for cost and inference speed advantages.Generally,most respondents are using a combination of OpenAI models and 1-2 other models from the other providers.We use different proprietary and 3rd party models because our custom
54、ers have diverse needs.Specialized models allow us to better tailor the Specialized models allow us to better tailor the experiences for our customers experiences for our customers and their use case-sales automation,agents for customer service and internal tools.In addition,we can offer our offer o
55、ur customers more flexible price points and optionscustomers more flexible price points and options,as well as be constantly experimenting with new models and business opportunities.VP Product,$1B+Revenue,Full Stack AI Company95%54%54%50%26%23%17%10%9%78%55%29%43%8%14%10%12%2%81%55%42%34%13%7%7%8%4%
56、OpenAI/GPTAnthropic/ClaudeGoogle/GeminiMeta/LLamaMistral AIDeepSeekCohereOtherxAIAvg number of models per respondent=2.82.8Notes:(1)Companies building both end user applications and core AI models/technologiesRetrieval augmented generation(RAG)and fineRetrieval augmented generation(RAG)and fine-tuni
57、ng are the most common model training techniques;hightuning are the most common model training techniques;high-growth growth companies tend to use a greater variety of promptcompanies tend to use a greater variety of prompt-based techniquesbased techniquesPrivate&Strictly Confidential13Model Trainin
58、g TechniquesSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkModel Training/Adaptatio
59、n TechniquesModel Training/Adaptation Techniques%of Respondents,N=273High Growth CompanyOther RespondentsCompared to last years State of AI report,a greater percentage of a greater percentage of respondents in this years survey respondents in this years survey are actively using RAG and are actively
60、 using RAG and finetuning techniquesfinetuning techniques.We expected finetuning to be a lower expected finetuning to be a lower percentage percentage given the investment required and how quickly base models are improving but it remains an area of focus 66%68%32%69%67%31%RAGFine-tuningPretraining49
61、%25%67%36%Few-Shot LearningZero-Shot LearningTraining TechniquesTraining TechniquesPromptPrompt-Based TechniquesBased TechniquesMost companies are using cloudMost companies are using cloud-based solutions and AI API providers for training and based solutions and AI API providers for training and inf
62、erenceinferencePrivate&Strictly Confidential14AI InfrastructureSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Boar
63、d,and others in our networkAI Infrastructure for Training and InferenceAI Infrastructure for Training and Inference%of Respondents,Select All That Apply,N=27368%64%23%10%8%Fully cloud-basedExternal AI APIprovidersHybridDedicated inferenceprovidersFully on-preminfrastructure(e.g.,cloud+on-prem GPU cl
64、usters)Most organizations are clearly leaning into fully managed AI solutions fully managed AI solutions-68%operate entirely in the cloud and 64%rely on external AI API providers-because this model minimizes upfront minimizes upfront capital outlay and operational capital outlay and operational comp
65、lexity,while maximizing speedcomplexity,while maximizing speed-toto-marketmarket.However,this reliance also means vendor selection,SLA negotiation,and cost-per-call management have become strategic priorities rather than just technical considerations.Meanwhile,only 23%of teams use a only 23%of teams
66、 use a hybrid approach and fewer than 1 in 10 hybrid approach and fewer than 1 in 10 maintain onmaintain on-prem or dedicated inference prem or dedicated inference infrastructureinfrastructure,underscoring that these models remain niche,adopted primarily in scenarios where control,compliance,or spec
67、ialized performance justify the extra overhead.As real-time AI use cases grow,theres an emerging opportunity emerging opportunity for turnkey inference platforms to for turnkey inference platforms to capture more sharecapture more share,but any move away from fully managed services will hinge on a c
68、lear business case or regulatory imperative.(e.g.,Fireworks,Together.ai,Baseten)Top challenges noted by companies when deploying models include hallucinations,explainability/trust,and proving ROITop challenges noted by companies when deploying models include hallucinations,explainability/trust,and p
69、roving ROIPrivate&Strictly Confidential15Model Deployment Challenges:Product DevelopmentSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,ou
70、r Technical Advisory Board,and others in our network3939%Proving ROIProving ROICompute costCompute costSecuritySecurityFinding right use casesFinding right use casesEase of integration with existing systemsEase of integration with existing systemsRegulatory and ethical considerationsRegulatory and e
71、thical considerationsTalentTalentLatencyLatency3838%34%34%32%32%26%26%25%25%24%24%20%20%16%16%15%15%Challenges in Model DeploymentChallenges in Model Deployment%of Respondents who ranked each aspect in Top 3,N=273Explainability&trustExplainability&trustHallucinationsHallucinationsMonitoringMonitorin
72、g10%10%Model drift over timeModel drift over time9%9%Accessing GPUsAccessing GPUs5%5%Explainability and trust ranked higher for companies building vertical AI applications,who may deal with additional compliance and legal restrictions in regulated industries like healthcareAs AI products scale,perfo
73、rmance monitoring becomes more important with many scaled AI products offering some kind of As AI products scale,performance monitoring becomes more important with many scaled AI products offering some kind of advanced performance monitoringadvanced performance monitoringPrivate&Strictly Confidentia
74、l16AI Performance MonitoringSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkApproach
75、 to AI Performance MonitoringApproach to AI Performance Monitoring%of Respondents,N=27019%14%3%4%75%66%59%40%6%16%31%44%4%7%12%Pre-LaunchBetaGeneral AvailabilityScalingAI Product MaturityNo formal monitoring in placeNo formal monitoring in placeBasic monitoring(tracking model Basic monitoring(tracki
76、ng model accuracy and performance)accuracy and performance)Advanced monitoring(drift Advanced monitoring(drift detection,realdetection,real-time feedback loops)time feedback loops)Fully automated model monitoring Fully automated model monitoring and retraining pipelinesand retraining pipelinesA sign
77、ificant number of companies are evaluating agentic workflows,with high growth AI companies more actively A significant number of companies are evaluating agentic workflows,with high growth AI companies more actively deploying AI agents in productiondeploying AI agents in productionPrivate&Strictly C
78、onfidentialAgentic WorkflowsSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkAgentic
79、WorkflowsAgentic Workflows%of Respondents,N=268No,and we have no current plans to invest in AI agentsNo,and we have no current plans to invest in AI agentsNo,but we plan to explore AI agents within No,but we plan to explore AI agents within the next 12 monthsthe next 12 monthsYes,but we are in early
80、 research and Yes,but we are in early research and exploration stagesexploration stagesYes,we are actively deploying AI agents Yes,we are actively deploying AI agents in productionin productionYes,we are experimenting with AI agents in Yes,we are experimenting with AI agents in pilots or internal us
81、e casespilots or internal use casesMany of our users like the insights and analytics we are surfacing but are unwilling to commit the time to fully explore the information housed in the product.We are looking to build out AI agents that effectively use the product for the effectively use the product
82、 for the endend-users to surface worthwhile users to surface worthwhile useruser-journeysjourneys and bring the end-user along for them.VP Product,$10-25M Revenue,Full Stack AI Company173%11%8%23%3%32%42%32%47%All Other CompaniesHigh GrowthPrivate and Strictly ConfidentialCopyright 2024 ICONIQ Capit
83、al,LLC.All Rights ReservedGoGo-toto-Market Market Strategy&Strategy&ComplianceComplianceFor AIFor AI-enabled companies,around 20enabled companies,around 20-35%of their product roadmap has been focused on AI35%of their product roadmap has been focused on AI-driven features with highdriven features wi
84、th high-growth companies dedicating closer to 30growth companies dedicating closer to 30-45%of their roadmap to AI45%of their roadmap to AI-driven featuresdriven featuresPrivate&Strictly Confidential19AI Product RoadmapSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from
85、 the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkWhat%of your product roadmap is focused on AIWhat%of your product roadmap is focused on AI-driven features?driven fea
86、tures?AI-Enabled Companies Only,Median,N=26822%31%36%43%All Other CompaniesHigh GrowthBy End of 2025(Estimated)By End of 2024Many companies are using a hybrid pricing model which includes a combination of subscription/planMany companies are using a hybrid pricing model which includes a combination o
87、f subscription/plan-based pricing along based pricing along with either usagewith either usage-based or outcomebased or outcome-based pricingbased pricingPrivate&Strictly Confidential20Primary Pricing ModelSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ t
88、eam and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkPrimary Pricing Model(Including AI Products/Features and Software)Primary Pricing Model(Including AI Products/Features and Softwar
89、e)%of Respondents,N=26638%36%19%6%HybridSubscription/Seat-basedUsage-basedOutcome-basedMost AI-enabled SaaS vendors seem to see AI as a tiebreakertiebreaker or upsell hook upsell hook-not yet as its own profit center.While bundling AI into premium tiers or including at no extra cost is the fastest w
90、ay to drive adoption and defend against competitors,we expect this approach to shift in the coming years as expect this approach to shift in the coming years as companies start to build telemetry on AI usage and companies start to build telemetry on AI usage and ROIROI,likely necessitating the shift
91、 to a usage-based model to avoid margin compression.Currently,most AICurrently,most AI-enabled companies are either including AI features as part of a premiumenabled companies are either including AI features as part of a premium-tier product or including them at tier product or including them at no
92、 extra costno extra costPrivate&Strictly Confidential21Pricing Models for AI FeaturesPrimary Pricing Model for AI Features/ProductsPrimary Pricing Model for AI Features/ProductsAI-Enabled Companies Only,%of Respondents,N=17440%33%21%5%2%AI features are part ofa premium-tierproductAI features areincl
93、uded at no extracostAI features have aseparate usage-basedpricing modelAI features have aseparate seat-basedpricing modelAI features have aseparate outcome-based pricing modelSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders c
94、onsisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkICONIQ CrossICONIQ Cross-Functional InsightFunctional InsightIn our 2025 State of GTM report,we asked this same we asked this same question to GTM leaders,and th
95、eir responsesquestion to GTM leaders,and their responseslargely aligned with R&D leaders largely aligned with R&D leaders further reinforcing the consistency of this trend across the market.Included in a premium-tier productIncluded at no extra costUsage-based modelSeat-based modelOutcome-based mode
96、l38%32%19%9%1%40%of companies have no plans to change pricing,but 37%of respondents are exploring new pricing models based on 40%of companies have no plans to change pricing,but 37%of respondents are exploring new pricing models based on consumption,ROI,and usage tiersconsumption,ROI,and usage tiers
97、Private&Strictly Confidential22Pricing ChangesPlans to Change AI Pricing in Next Twelve MonthsPlans to Change AI Pricing in Next Twelve Months%of Respondents,N=27323%40%37%YesYesNoNoI dont knowI dont know“We would like to integrate willingness to integrate willingness to pay and clear connection to
98、ROIpay and clear connection to ROI outcomes into our pricing model”VP Product,$100-150M Revenue,Full Stack AI Company“We are observing if AI capabilities deliver extra value to customer.Once we have critical adoption and proof of added value,we might segment the current tiers of our segment the curr
99、ent tiers of our platformplatform(i.e.create a top tier with the full AI/agents,a limit on the basic,and enterprise tiers)”VP Product,$100-150M Revenue,Full Stack AI Company“We will complement premium tier model complement premium tier model pricing with pricing models centered around pricing with p
100、ricing models centered around consumptionconsumption.I expect we will also experiment with outcome-based pricing but it is unclear how we will structure pricing in such a way that it allows customers to allows customers to accurately budget for these costsaccurately budget for these costs.”VP Produc
101、t,$100-150M Revenue,Full Stack AI Company“The subscription model is not working for us.Power users tend to use a lot Power users tend to use a lot resulting in negative margins resulting in negative margins considering LLM API costs,while users who arent users who arent using are at risk of churnusi
102、ng are at risk of churn.Considering the high variable cost we are planning to move planning to move to usage basedto usage based but bundle usage as a subscription e.g.,10M token per year package”VP Product,$100-150M Revenue,Full Stack AI CompanyFactoring in ROIFactoring in ROIConsumption and Outcom
103、eConsumption and Outcome-Based PricingBased PricingSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others
104、 in our networkAs AI products scale,providing detailed model transparency reports or basic insights on how AI influences outcomes As AI products scale,providing detailed model transparency reports or basic insights on how AI influences outcomes becomes more criticalbecomes more criticalPrivate&Stric
105、tly Confidential23AI Explainability and Transparency13%1%3%31%26%24%25%50%64%58%47%6%10%17%25%Pre-LaunchBetaGeneral AvailabilityScalingAI Product MaturityOtherOtherWe dont provide AIWe dont provide AI-specific specific explanations to customersexplanations to customersWe offer basic insights on how
106、We offer basic insights on how AI influences outcomesAI influences outcomesWe provide detailed model We provide detailed model transparency reportstransparency reportsStrategy for AI Explainability and Transparency to CustomersStrategy for AI Explainability and Transparency to Customers%of Responden
107、ts,N=266Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkMost companies have guardrai
108、ls around AI ethics and governance policies,with the majority of respondents using humanMost companies have guardrails around AI ethics and governance policies,with the majority of respondents using human-inin-thethe-loop oversight to ensure AI fairness and safetyloop oversight to ensure AI fairness
109、 and safetyPrivate&Strictly Confidential24AI Compliance and GovernanceBasic compliance with Basic compliance with data privacy laws(e.g.,data privacy laws(e.g.,GDPR,CCPA)GDPR,CCPA)Formal AI ethics and Formal AI ethics and governance policies in placegovernance policies in placeDedicated AI complianc
110、e and Dedicated AI compliance and governance teamgovernance teamHow does your company handle AI How does your company handle AI compliance and governance?compliance and governance?%of Respondents,N=29166%42%38%21%21%14%1%Human-in-the-loopoversightExplainability andtransparencymeasuresBias detection
111、andmitigationtechniquesAdversarial testingfor robustnessAI model red teamtestingNo formalsafeguards in placeOtherWhat safeguards does your company use to ensure AI fairness and safety?What safeguards does your company use to ensure AI fairness and safety?%of Respondents,N=29111%47%29%13%No formal AI
112、 compliance No formal AI compliance strategystrategySource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and other
113、s in our networkPrivate and Strictly ConfidentialCopyright 2024 ICONIQ Capital,LLC.All Rights ReservedOrganization Organization StructureStructureMany companies have dedicated AI leadership by the time they reach$100M in revenue likely due to increasing operational Many companies have dedicated AI l
114、eadership by the time they reach$100M in revenue likely due to increasing operational complexity and the need to have a centralized owner for AI strategycomplexity and the need to have a centralized owner for AI strategyPrivate&Strictly Confidential26Dedicated AI/ML Leadership33%50%48%51%61%4%3%6%3%
115、5%59%47%42%40%31%5%3%6%3%$100M$100M-$200M$200M-$500M$500M-$1B$1B+2024 RevenueNo,but AI is part of our No,but AI is part of our broader R&D strategybroader R&D strategyYes,we have dedicated AI Yes,we have dedicated AI leadershipleadershipNo,but we are planning to hire No,but we are planning to hire d
116、edicated AI/ML leadershipdedicated AI/ML leadershipNo,we rely on external No,we rely on external AI providersAI providersDoes your company have dedicated AI/ML leadership(e.g.,Chief AI Officers,Head of ML,AI Research Lead)?Does your company have dedicated AI/ML leadership(e.g.,Chief AI Officers,Head
117、 of ML,AI Research Lead)?%of Respondents,N=290Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in o
118、ur networkMost companies currently have dedicated AI/ML engineers,data scientists,and AI product managers,with AI/ML engineers Most companies currently have dedicated AI/ML engineers,data scientists,and AI product managers,with AI/ML engineers taking the longest time on average to hiretaking the lon
119、gest time on average to hirePrivate&Strictly Confidential27AI-Specific RolesAIAI-Specific Roles and Hiring PlanSpecific Roles and Hiring Plan%of Respondents,N=29088%72%54%38%22%20%17%2%67%45%46%24%12%21%26%4%AI/ML engineersData scientistsAI productmanagersData architectsData visualizationspecialists
120、Prompt engineers AI design specialistsOther70686766446261N/AAvg Lead Time to Avg Lead Time to Hire(#Days)Hire(#Days)Currently havePlanning to hireSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of
121、 CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our network46%54%Across respondents,there was a relatively even split in sentiment towards the pace of hiring,with those who felt like they Across respondents,there was a relatively even split in senti
122、ment towards the pace of hiring,with those who felt like they were not hiring fast enough primarily citing lack of qualified candidates as the main constraintwere not hiring fast enough primarily citing lack of qualified candidates as the main constraintPrivate&Strictly Confidential28Pace of HiringP
123、ace of HiringPace of Hiring%of Respondents,N=291Yes,we are hiring Yes,we are hiring fast enoughfast enoughNo,we are not hiring No,we are not hiring fast enoughfast enough60%49%35%25%4%Hiring is slow due tolack of qualifiedcandidatesHiring is slow due tocost constraintsHiring is slow due tocompetitio
124、nHiring is slow due tointernal processchallengesOtherReasons for Slow HiringReasons for Slow Hiring%of Respondents,N=134Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI in
125、itiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkOn average,companies plan to have 20On average,companies plan to have 20-30%of their engineering team focused on AI,with high growth companies having a 30%of their engineering team focused on AI,with high growth comp
126、anies having a higher proportion of their engineering team focused on AIhigher proportion of their engineering team focused on AIPrivate&Strictly Confidential29%of Engineering Team Focused on AIEstimated%of Engineering Team Focused on AIEstimated%of Engineering Team Focused on AI%of Respondents,N=29
127、018%28%28%37%All Other CompaniesHigh Growth2025%of Eng Team2026%of Eng TeamSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical A
128、dvisory Board,and others in our networkPrivate and Strictly ConfidentialCopyright 2024 ICONIQ Capital,LLC.All Rights ReservedAI CostsAI CostsOn average,companies are allocating 10On average,companies are allocating 10-20%of their R&D budget to AI development,with most companies planning to 20%of the
129、ir R&D budget to AI development,with most companies planning to increase spend on AI in 2025increase spend on AI in 2025Private&Strictly Confidential31AI Development SpendWhat percentage of your total R&D budget is allocated to AI development?What percentage of your total R&D budget is allocated to
130、AI development?AI-Enabled Companies Only,%of Respondents,N=1402024 Budget2025 BudgetSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Te
131、chnical Advisory Board,and others in our network14%10%10%10%15%25%15%15%20%25%$100M$100M-$200M$200M-$500M$500M-$1B$1B+2024 RevenueAs AI products scale,the cost of talent tends to go down as a total proportion of spend;conversely,infrastructure and As AI products scale,the cost of talent tends to go
132、down as a total proportion of spend;conversely,infrastructure and compute costs tend to increase as products start to see market tractioncompute costs tend to increase as products start to see market tractionPrivate&Strictly Confidential32Budget AllocationWhat percentage of your AI budget is allocat
133、ed across the following categories?What percentage of your AI budget is allocated across the following categories?%of Respondents,N=2915%6%6%7%13%24%20%22%10%12%12%13%4%9%11%10%8%11%10%12%57%38%40%36%3%1%Pre-LaunchBetaGAScalingAI Product MaturityOther AI related costsOther AI related costsAI talent(
134、salaries,hiring,AI talent(salaries,hiring,upskilling)upskilling)AI model trainingAI model trainingAI model inferenceAI model inferenceData storage&processingData storage&processingAI infrastructure&cloud costsAI infrastructure&cloud costsAI governance,compliance,and strategyAI governance,compliance,
135、and strategySource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkOf the various infrastru
136、cture costs,respondents cited API usage fees as the cost most challenging to control,suggesting Of the various infrastructure costs,respondents cited API usage fees as the cost most challenging to control,suggesting companies face the most unpredictability around variable costs tied to external API
137、consumptioncompanies face the most unpredictability around variable costs tied to external API consumptionPrivate&Strictly Confidential33Infrastructure CostsWhich Infrastructure Costs are Most Challenging to Control?Which Infrastructure Costs are Most Challenging to Control?%of Respondents who ranke
138、d each aspect in Top 3,N=29170%49%48%47%42%API usage feesInference costsModel retraining and updatesTraining costsStorage costsSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseein
139、g AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkTo cut AI infrastructure costs,organizations are exploring openTo cut AI infrastructure costs,organizations are exploring open-source models and ways to optimize inference efficiencysource models and ways to o
140、ptimize inference efficiencyPrivate&Strictly Confidential34Cost OptimizationHow are you optimizing AI infrastructure costs?How are you optimizing AI infrastructure costs?%of Respondents,N=29141%37%32%28%26%3%Moving to open-sourcemodelsOptimizing inferenceefficiencyNo significant costoptimization eff
141、ortsLeveraging modeldistillation orquantizationSwitching to more cost-efficient hardwareOtherSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CT
142、Os,our Technical Advisory Board,and others in our networkMost respondents are training or finetuning models at least monthly,with estimated monthly model training costs ranging Most respondents are training or finetuning models at least monthly,with estimated monthly model training costs ranging fro
143、m$160Kfrom$160K-$1.5M depending on the product maturity$1.5M depending on the product maturityPrivate&Strictly Confidential35Model TrainingHow often do you retrain or fineHow often do you retrain or fine-tune your AI models?tune your AI models?%of Respondents,N=29120%12%31%19%13%5%Multiple times per
144、 weekMultiple times per weekOnce a weekOnce a weekMonthlyMonthlyEvery 3Every 3-6 months6 monthsRarelyRarelyNeverNever$163K$249K$1.1M$1.5MPre-LaunchBetaGAScalingAI Product MaturityEstimated Monthly Model Training CostsEstimated Monthly Model Training CostsAverage USD,N=229$38M$125M$225M$500MMedian An
145、nual RevenueSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkInference costs surge po
146、stInference costs surge post-launch with highlaunch with high-growth AI companies spending up to 2x more at GA and scale than their peersgrowth AI companies spending up to 2x more at GA and scale than their peersPrivate&Strictly Confidential36Deployment Costs:InferenceMonthly Spend for InferenceMont
147、hly Spend for Inference%of Respondents,N=221$100K$286K$1.0M$1.1M$1.6M$2.3MPre-LaunchBetaGAScalingAI Product MaturityN/AN/AOther CompaniesHigh Growth Companies$38M$125M$225M$500MMedian Annual RevenueSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and
148、network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkData storage&processing costs also climb steeply from GA stage onward,with highData storage&processing costs also climb steeply from GA st
149、age onward,with high-growth AI builders spending more on growth AI builders spending more on data storage and processing than their peersdata storage and processing than their peersPrivate&Strictly Confidential37Deployment Costs:Data Storage&ProcessingMonthly Spend for Data StorageMonthly Spend for
150、Data Storage%of Respondents,N=221$188K$554K$1.2M$1.9M$1.6M$2.6MPre-LaunchBetaGAScalingAI Product MaturityAI Product MaturityOther CompaniesHigh Growth CompaniesMonthly Spend for Data ProcessingMonthly Spend for Data Processing%of Respondents,N=226$107K$594K$0.7M$1.8M$1.6M$2.0MPre-LaunchBetaGAScaling
151、AI Product MaturityAI Product MaturityN/AN/AN/AN/A$38M$125M$225M$500M$38M$125M$225M$500MMedian Annual RevenueSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives i
152、n enterprises,CTOs,our Technical Advisory Board,and others in our networkPrivate and Strictly ConfidentialCopyright 2024 ICONIQ Capital,LLC.All Rights ReservedInternal Internal ProductivityProductivityInternal AI productivity budgets are set to nearly double in 2025 across all revenue tiers,with com
153、panies spending anywhere Internal AI productivity budgets are set to nearly double in 2025 across all revenue tiers,with companies spending anywhere from 1from 1-8%of total revenue8%of total revenuePrivate&Strictly Confidential39Annual Internal Productivity Budget$0.3$0.6$1.0$1.0$1.4$2.0$6.9$34.2$0.
154、4$1.0$1.7$1.8$2.3$3.2$14.5$60.4$10M$10M-$24M$25M-$49M$50M-$99M$100-$200M$200-$500M$500M-$1B$1B+2024 Revenue2024 RevenueApproximately what is your organizations annual generative AI spend for internal productivity?Approximately what is your organizations annual generative AI spend for internal produc
155、tivity?Average($M USD)by Revenue Range2024 Spend2025 Spend(Estimated)5%8%3%6%3%4%1%2%1%2%1%1%1%2%1%2%Approximate%of RevenueSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI
156、 initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkR&D budgets still remain the most common source of AI internal productivity budgets for enterprise companies;however,R&D budgets still remain the most common source of AI internal productivity budgets for enterpri
157、se companies;however,we are also starting to see headcount budgets being used for internal productivity spendwe are also starting to see headcount budgets being used for internal productivity spendPrivate&Strictly Confidential40Internal Productivity Budget Sources for EnterprisesWhere is the budget
158、for internal productivity coming from?Where is the budget for internal productivity coming from?%of Respondents,$500M+Revenue Respondents Only$500M+Revenue Respondents Only59%44%47%57%48%39%23%22%27%Coming from R&D budgetComing from businessunit(non-R&D)initiativesComing from innovationbudget(non-R&
159、D)Coming from headcountbudgetNet new budget beingcreatedN/A2024 State of AI Survey(N=126)2025 State of AI Survey(N=99)Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI init
160、iatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkWhile around 70%of employees have access to various AI tools for internal productivity,only 50%of employees are using While around 70%of employees have access to various AI tools for internal productivity,only 50%of em
161、ployees are using AI tools on an ongoing basis with adoption more difficult in mature Enterprises($1B+revenue)AI tools on an ongoing basis with adoption more difficult in mature Enterprises($1B+revenue)Private&Strictly Confidential41AI Access and UsageAI Tools for Internal Productivity:Access and Us
162、ageAI Tools for Internal Productivity:Access and UsageAverage%of Employees,N=25870%66%69%68%62%57%50%49%51%44%$100M$100M-$200M$200M-$500M$500M-$1B$1B+2024 Revenue%of Employees with Access to AI Tools%of Employees Using AI Tools on Ongoing BasisJust deploying tools is a recipe for disappointment,part
163、icularly for large enterprises.To truly empower employees,you need to pair pair availability with scaffolding availability with scaffolding that includes training,spotlighting champions,and most importantly relentless executive support.Don VuDon VuSVP,Chief Data&Analytics Officer,New York LifeSource
164、:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkWhen choosing foundational models for inte
165、rnal use cases,cost is the most important consideration followed by accuracy When choosing foundational models for internal use cases,cost is the most important consideration followed by accuracy and privacyand privacyPrivate&Strictly Confidential42Top Considerations for Foundational Models:Internal
166、 Use CasesTop Considerations When Choosing a Foundational Model for Top Considerations When Choosing a Foundational Model for Internal Use CasesInternal Use Cases%of Respondents who ranked each aspect in Top 3,N=26574%74%PrivacyPrivacyAbility to finetune/customizeAbility to finetune/customizeSOC2/En
167、terprise SLAsSOC2/Enterprise SLAsOpen SourceOpen SourceLatencyLatency7272%50%50%38%38%26%26%16%16%13%13%AccuracyAccuracyCostCostWhereas accuracy ranked as the most important factor when deploying external AI products,cost is the most cost is the most important consideration when important considerat
168、ion when choosing models for internal AI choosing models for internal AI use casesuse cases.Privacy also becomes a more important consideration for internal use cases compared to external.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of
169、 AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkThe biggest challenges facing organizations deploying AI for internal use cases are often strategic(i.e.finding the right uThe biggest challenges fa
170、cing organizations deploying AI for internal use cases are often strategic(i.e.finding the right use se cases and proving ROI)vs technicalcases and proving ROI)vs technicalPrivate&Strictly Confidential43Model Deployment Challenges:Internal Use CasesTop Challenges in Model Deployment for Top Challeng
171、es in Model Deployment for Internal Use CasesInternal Use Cases%of Respondents who ranked each aspect in Top 3,N=2734646%Explainability&trustExplainability&trustHallucinationsHallucinationsSecuritySecurityCompute costCompute costTalentTalentRegulatory and ethical considerationsRegulatory and ethical
172、 considerationsMonitoringMonitoringLatencyLatency4242%32%32%31%31%29%29%28%28%21%21%15%15%12%12%9%9%Proving ROIProving ROIFinding right use casesFinding right use casesModel drift over timeModel drift over time9%9%Accessing GPUsAccessing GPUs6%6%Source:Perspectives from the ICONIQ GenAI Survey(April
173、 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkCompanies are typically exploring multiple GenAI use cases across functions,with companies
174、 that have high employee Companies are typically exploring multiple GenAI use cases across functions,with companies that have high employee adoption using GenAI across 7+use casesadoption using GenAI across 7+use casesPrivate&Strictly Confidential44Number of Use Cases4.66.07.1LowMediumHighAverage Nu
175、mber of Use Cases by Strength of Internal AI AdoptionAverage Number of Use Cases by Strength of Internal AI Adoption%of Respondents,N=258Greater than 50%of employees actively using AI tools20-50%of employees actively using AI toolsLess than 20%of employees actively using AI toolsStrength of Internal
176、 Strength of Internal AI AdoptionAI AdoptionSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our
177、 networkR&D and S&M use cases lead in popularity,while G&A use cases still lag in comparisonR&D and S&M use cases lead in popularity,while G&A use cases still lag in comparisonPrivate&Strictly Confidential45Top Use Cases:By PopularityTop Use CasesTop Use Cases%of Respondents,Select All That Apply,N=
178、25877%65%57%56%48%45%42%42%41%40%38%33%26%13%R&DS&MG&ASource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and oth
179、ers in our networkTop use cases by impact mirror usage trends with coding assistance by far outpacing other use cases in terms of tangible Top use cases by impact mirror usage trends with coding assistance by far outpacing other use cases in terms of tangible impact on productivityimpact on producti
180、vityPrivate&Strictly Confidential46Top Use Cases:By ImpactTop Use Cases by Biggest Impact on ProductivityTop Use Cases by Biggest Impact on Productivity%of Respondents who ranked each aspect in Top 3,N=25865%37%30%28%22%21%18%16%14%13%10%5%5%4%Coding assistanceContent generation/writing assistantsDo
181、cumentation and knowledge retrievalProduct and DesignCustomer engagement/serviceSales productivityData analytics and business intelligenceQA and TestingDevOps/MLOpsMarketing automationIT&SecurityLegal and contract reviewHR and recruiting toolsFP&A automationHigh growth companies tend to see an avera
182、ge 33%of their total code 33%of their total code being written with AI being written with AI compared to 27%for all other companiesRespondents cited an average average productivity gain of 15productivity gain of 15-30%30%across these GenAI use casesR&DS&MG&ASource:Perspectives from the ICONIQ GenAI
183、Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkHigh growth companies tend to more actively experiment with and adopt new AI t
184、ools,suggesting that leading companies High growth companies tend to more actively experiment with and adopt new AI tools,suggesting that leading companies view AI as a strategic lever and are moving faster to integrate it across internal workflowsview AI as a strategic lever and are moving faster t
185、o integrate it across internal workflowsPrivate&Strictly Confidential47Attitude Towards Internal AI AdoptionAttitude Towards Internal AI AdoptionAttitude Towards Internal AI Adoption%of Respondents,N=25880%92%19%8%2%All Other CompaniesHigh Growth CompanyWe actively experiment with and We actively ex
186、periment with and adopt new AI toolsadopt new AI toolsWe are cautious and selectively We are cautious and selectively integrate AI where its proven valuableintegrate AI where its proven valuableWe are skeptical and havent adopted We are skeptical and havent adopted many AImany AI-powered internal to
187、ols yetpowered internal tools yetSource:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkMos
188、t companies are measuring productivity improvements and cost savings from internal AI useMost companies are measuring productivity improvements and cost savings from internal AI usePrivate&Strictly Confidential48Tracking ROITracking ROITracking ROI%of Respondents,N=25817%23%16%14%30%No,we have not s
189、tarted No,we have not started measuring AIs impactmeasuring AIs impactNo,but we are currently working on No,but we are currently working on ways to measure AI impactways to measure AI impactYes,we track only qualitative gains Yes,we track only qualitative gains(e.g.,surveys,employee feedback)(e.g.,s
190、urveys,employee feedback)Yes,we track only quantitative gains(e.g.,Yes,we track only quantitative gains(e.g.,time savings,task completion rates)time savings,task completion rates)Yes,we track both quantitative and Yes,we track both quantitative and qualitative AIqualitative AI-driven efficiency gain
191、sdriven efficiency gainsHow are you measuring the impact of using AI for internal use on your business?How are you measuring the impact of using AI for internal use on your business?%of Respondents,N=25875%51%20%20%Productivity gainsCost savingsRevenue upliftCustomer retention&engagementimprovements
192、Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkPrivate and Strictly ConfidentialCop
193、yright 2024 ICONIQ Capital,LLC.All Rights ReservedAI Builder AI Builder Tech StackTech StackFrameworks vs Managed Platforms Frameworks vs Managed Platforms Core deepCore deep-learning frameworks remain popular learning frameworks remain popular with PyTorch and TensorFlow accounting for over half of
194、 all usage across respondents But theyre nearly matched by fully managed or APImatched by fully managed or API-driven offerings driven offerings-prevalence of AWS SageMaker and OpenAIs fine-tuning service show that teams are split between“build your own”split between“build your own”and“let someone e
195、lse run it”approachesand“let someone else run it”approachesEcosystem Players Gaining TractionEcosystem Players Gaining Traction The Hugging Face ecosystem and Databricks Mosaic AI Training are carving out meaningful niches,providing higherproviding higher-level abstractions over raw frameworkslevel
196、abstractions over raw frameworks Meanwhile,more specialized or emerging tools(AnyScale,Fast.ai,Modal,JAX,Lamini)landed in the single-digit percentages,suggesting experimentation is underway but broad adoption remains experimentation is underway but broad adoption remains nascentnascentEnterpriseEnte
197、rprise-Grade NeedsGrade Needs Later-stage companies typically have larger data teams,more complex pipelines,and stricter larger data teams,more complex pipelines,and stricter requirements around security,governance,and compliancerequirements around security,governance,and compliance Databricks unifi
198、ed“lakehouse”architecture(which blends data engineering,analytics,and ML)and AnyScales managed Ray clusters(which simplify distributed training and hyperparameter tuning)both speak directly to those enterprise needs with more respondents in the$500M+more respondents in the$500M+revenue range using t
199、hese solutionsrevenue range using these solutionsMost Used Tools:Model Training&FinetuningMost Used Tools:Model Training&FinetuningPrivate&Strictly Confidential50Notes:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services o
200、f ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkMost Widely Used ToolsMost
201、Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysOrchestration Frameworks Reign SupremeOrchestration Frameworks Reign Supreme Top frameworks used include LangChain and Hugging Faces toolset which signals that teams clearly value highvalue high-level libraries that simplify
202、prompt chaining,batching,and interfacing with either level libraries that simplify prompt chaining,batching,and interfacing with either public or selfpublic or self-hosted modelshosted models Around 70%of respondents also specified that they use private or custom LLM APIs use private or custom LLM A
203、PIs Safety and HigherSafety and Higher-Level SDKs Gaining TractionLevel SDKs Gaining Traction Roughly 3 in 10 respondents use Guardrails to enforce safety checks,and almost a quarter leverage Vercels AI SDK(23%)for rapid deployment which shows growing awareness that production LLM growing awareness
204、that production LLM apps need both guardrails and streamlined integration layersapps need both guardrails and streamlined integration layersLongLong-Tail ExperimentationTail Experimentation Emerging players like CrewAI,Modal Labs,Instructor,DSPy,and DotTXT had weaker usage,indicating that while expe
205、rimentation is widespread,broad standardization has yet to settle beyond while experimentation is widespread,broad standardization has yet to settle beyond the big playersthe big playersMost Used Tools:LLM&AI Application DevelopmentMost Used Tools:LLM&AI Application DevelopmentPrivate&Strictly Confi
206、dential51Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI
207、Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkIncumbent Infrastructure Still RulesIncumbent Infrastructure Still Rules Nearl
208、y half of teams lean on their existing APM/logging stacks existing APM/logging stacks(Datadog,Honeycomb,New Relic,etc.)rather than adopting MLrather than adopting ML-specific tools specific tools-underscoring that ease of integration and organizational standardization often outweigh the benefits of
209、bespoke AI monitoringEarly Traction for MLEarly Traction for ML-Native PlatformsNative Platforms Both LangSmith and Weights&Biases have broken through to reach 17%adoption,showing real real appetite for turnkey solutions that instrument prompt chains,track embeddings,and surface drift appetite for t
210、urnkey solutions that instrument prompt chains,track embeddings,and surface drift without boltwithout bolt-onsons to legacy systemsto legacy systemsFragmented Long Tail&Knowledge GapsFragmented Long Tail&Knowledge Gaps Beyond the top two ML-native names,usage quickly fragments across playersusage qu
211、ickly fragments across players like Arize,Fiddler,Helicone,Arthur,etc,and 10%of respondents didnt know which tool they used;this points to both a nascent ecosystem and confusion around what“observability”even means for generative AIa nascent ecosystem and confusion around what“observability”even mea
212、ns for generative AIMost Used Tools:Monitoring and ObservabilityMost Used Tools:Monitoring and ObservabilityPrivate&Strictly Confidential52Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective owner
213、s.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,ou
214、r Technical Advisory Board,and others in our networkNVIDIAs Grip on ProductionNVIDIAs Grip on Production TensorRT and Triton Inference Server together command over 60%adoption,underscoring how how dominant NVIDIAs stack remains for squeezing latency and throughput out of GPUdominant NVIDIAs stack re
215、mains for squeezing latency and throughput out of GPU-based based deploymentsdeploymentsCrossCross-Platform Alternatives Gaining SharePlatform Alternatives Gaining Share The ONNX Runtime(18%)is the top non-NVIDIA solution,reflecting teams desire for hardwaredesire for hardware-agnostic acceleration
216、across CPUs,GPUs,and acceleratorsagnostic acceleration across CPUs,GPUs,and accelerators TorchServe(15%)likewise shows that pure-PyTorch serving still has a foothold,especially for CPU-only workloads or simpler containerized setupsKnowledge Gaps&Untapped PotentialKnowledge Gaps&Untapped Potential Wi
217、th 17%respondents they didnt know which optimization they use and 14%reporting“None,”theres clear confusion or inexperience around inference tuningconfusion or inexperience around inference tuning,suggesting an opportunity for education(and tooling)around quantization,pruning,and efficient runtimes-
218、especially for teams running at scaleMost Used Tools:Inference OptimizationMost Used Tools:Inference OptimizationPrivate&Strictly Confidential53Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective
219、owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CT
220、Os,our Technical Advisory Board,and others in our networkDirectDirect-fromfrom-Provider Is KingProvider Is King The majority of teams hit model hosts directly via OpenAI,Anthropic,etc.underscoring that the path of least resistance remains calling the vendors own inference APIs rather than building o
221、r calling the vendors own inference APIs rather than building or integrating through a middle layerintegrating through a middle layerHyperscalersHyperscalers Close BehindClose Behind AWS Bedrock and Google Vertex AI have carved out substantial share,reflecting strong demand strong demand for unified
222、,enterprisefor unified,enterprise-grade ML platforms that bundle hosting with governance,security,and grade ML platforms that bundle hosting with governance,security,and billing in a single panebilling in a single pane In particular,a greater number of latergreater number of later-stage companies($5
223、00M+revenue)reported using stage companies($500M+revenue)reported using hyperscalerhyperscaler solutionssolutionsFragmented Alternatives&Emerging PlayersFragmented Alternatives&Emerging Players Beyond the big three,usage quickly fragments across players like Fireworks,Modal,Together.ai,usage quickly
224、 fragments across players like Fireworks,Modal,Together.ai,AnyScaleAnyScale,BasetenBaseten,Replicate,Deep Infra,etc.,Replicate,Deep Infra,etc.This long tail suggests teams are still exploring specialty hosts,often driven by unique pricing,still exploring specialty hosts,often driven by unique pricin
225、g,performance SLAs,or feature sets performance SLAs,or feature sets(e.g.,custom runtimes,on-prem options)Most Used Tools:Model HostingMost Used Tools:Model HostingPrivate&Strictly Confidential54Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNote
226、s:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CI
227、O/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkNo Clear StandNo Clear Stand-alone Leaderalone Leader Nearly 1 in 4 teams use mostly builtNearly 1 in 4 teams use mostly built-in evaluation features from platforms like Vertex,Weights&in evalu
228、ation features from platforms like Vertex,Weights&Biases,or GalileoBiases,or Galileo while 20%of respondents simply“didnt know”which tool they use,signaling many organizations are still leaning on the evaluation capabilities baked into their existing ML stacks rather than adopting a dedicated framew
229、orkEmerging Specialized FrameworksEmerging Specialized Frameworks LangSmithLangSmith and and LangfuseLangfuse lead the pack of purposelead the pack of purpose-built evaluation toolsbuilt evaluation tools,with HumanLoop and Braintrust also showing traction;these platforms are winning mindshare by off
230、ering richer prompt-level metrics,customizable test suites,and drift detection out of the boxKnowledge Gaps and DIYKnowledge Gaps and DIY Almost a quarter of respondents did not know which evaluation tool they used or did not have an evaluation tool in place,signaling both confusion around what“eval
231、uation”entails for generative AI both confusion around what“evaluation”entails for generative AI and the risk of unmonitored model regressionsand the risk of unmonitored model regressions Meanwhile,some respondents are also rolling their own evaluation pipelines,suggesting offrolling their own evalu
232、ation pipelines,suggesting off-thethe-shelf tooling hasnt yet covered all use casesshelf tooling hasnt yet covered all use casesMost Used Tools:Model EvaluationMost Used Tools:Model EvaluationPrivate&Strictly Confidential55Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphab
233、etical orderKey TakeawaysNotes:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders cons
234、isting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkClassic Big Data Tools Still DominateClassic Big Data Tools Still Dominate Apache Spark(44%of respondents)and Kafka(42%of respondents)lead the pack,underscoring tha
235、t at scale,teams default to battleat scale,teams default to battle-tested,distributed batchtested,distributed batch-andand-stream frameworks for ETL and stream frameworks for ETL and realreal-time data ingestiontime data ingestionPython Power BasePython Power Base Despite heavy big-data footprints,4
236、1%of respondents still lean on Panda-showing that for for smaller datasets,prototyping,or edge cases,the simplicity and flexibility of insmaller datasets,prototyping,or edge cases,the simplicity and flexibility of in-memory Python memory Python tooling remain indispensabletooling remain indispensabl
237、eFeature Stores on the HorizonFeature Stores on the Horizon Only 17%are using a dedicated feature store,indicating that while the concept of“build once,serve while the concept of“build once,serve everywhere”for features is gaining visibility,most organizations havent yet operationalized it at everyw
238、here”for features is gaining visibility,most organizations havent yet operationalized it at scalescale As maturity grows,well likely see feature stores and lightweight orchestrators As maturity grows,well likely see feature stores and lightweight orchestrators(Dask,Airflow,etc.)climb the ranks-but f
239、or now the Apache ecosystem rulesMost Used Tools:Data Processing&Feature EngineeringMost Used Tools:Data Processing&Feature EngineeringPrivate&Strictly Confidential56Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the propert
240、y of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiat
241、ives in enterprises,CTOs,our Technical Advisory Board,and others in our networkSearch Engines Evolve Into Vector StoresSearch Engines Evolve Into Vector Stores Elastic and Pinecone lead adoption,reflecting how teams either retrofit existing fullteams either retrofit existing full-text search text se
242、arch platforms for embeddings or adopt purposeplatforms for embeddings or adopt purpose-built,managed vector enginesbuilt,managed vector enginesRedis&the“Long Tail”Redis&the“Long Tail”Redis shows the appeal of leveraging inRedis shows the appeal of leveraging in-memory data stores you already runmem
243、ory data stores you already run,while other solutions like Clickhouse,AlloyDB,Milvus,PGVector,etc,underscores that many organizations are experimenting with different backends to balance cost,latency,and feature needsexperimenting with different backends to balance cost,latency,and feature needsRise
244、 of OpenRise of Open-Source SolutionsSource Solutions Specialist open-source tools like Chroma,Weaviate,Faiss,Qdrant,and Supabases vector addon are chipping away at the early leaders,signaling a competitive battleground for easea competitive battleground for ease-ofof-use,scaling,use,scaling,and clo
245、udand cloud-native integrationsnative integrationsMost Used Tools:Vector DatabasesMost Used Tools:Vector DatabasesPrivate&Strictly Confidential57Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective
246、 owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,C
247、TOs,our Technical Advisory Board,and others in our networkInIn-House Reigns SupremeHouse Reigns Supreme Over half of teams(52%)build their own toolinghalf of teams(52%)build their own tooling,suggesting that off-the-shelf providers still struggle to cover every use case or integrate with existing pi
248、pelinesScale AI is the clear vendor leaderScale AI is the clear vendor leader At 21%adoption,Scale AI is the goScale AI is the go-to thirdto third-party syntheticparty synthetic-data platform data platform-but even it only reaches one in five organizationsEarly Traction for Programmatic FrameworksEa
249、rly Traction for Programmatic Frameworks Snorkel AI and Mostly AI show that programmatic labeling and generation tools are gaining programmatic labeling and generation tools are gaining mindshare,but remain far behind custom solutionsmindshare,but remain far behind custom solutionsMost Used Tools:Sy
250、nthetic Data&Data AugmentationMost Used Tools:Synthetic Data&Data AugmentationPrivate&Strictly Confidential58Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective owners.None of the companies illust
251、rated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and
252、 others in our networkDominance of First MoversDominance of First Movers GitHub Copilot is used by nearly threeGitHub Copilot is used by nearly three-quarters of development teamsquarters of development teams,thanks to its tight VS Code integration,multi-language support,and backing by GitHubs massi
253、ve user base Copilots network effects and product-market fit make it hard to dislodge,but the strong secondstrong second-place showing for Cursor(used by 50%of respondents)signals appetite for diverse IDE integrationsplace showing for Cursor(used by 50%of respondents)signals appetite for diverse IDE
254、 integrationsLong Tail of Offerings LagLong Tail of Offerings Lag After the top two,adoption drops off sharply with a fractured long tail of solutions,suggesting that fractured long tail of solutions,suggesting that while most teams have trialed at least one assistant,very few have standardized on a
255、lternativeswhile most teams have trialed at least one assistant,very few have standardized on alternatives Low-code or no-code solutions like Retool,Lovable,Bolt,and Replit also had honorable mentions indicating that there is increasing appetite in the market for ideaincreasing appetite in the marke
256、t for idea-toto-application solutionsapplication solutionsMost Used Tools:Coding AssistanceMost Used Tools:Coding AssistancePrivate&Strictly Confidential59Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their
257、respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in ent
258、erprises,CTOs,our Technical Advisory Board,and others in our networkMLflowMLflow LeadsLeadsbut No Monopolybut No Monopoly MLflow was used by 36%of respondents and the clear frontrunner for experiment tracking,model clear frontrunner for experiment tracking,model registry,and basic pipeline orchestra
259、tionregistry,and basic pipeline orchestration this is only just over one-third of teams,indicating plenty of room for alternatives Weights&Biases also holds strong share with 20%of respondents using,reflecting its appeal as a its appeal as a managed SaaS for tracking,visualization,and collaborationm
260、anaged SaaS for tracking,visualization,and collaboration Beyond the top two,usage quickly fragments Beyond the top two,usage quickly fragments 16%“dont know”which tools power their MLOps and other tool mentions include Resolve.ai,Cleric,PlayerZero,Braintrust,etc.This points to both confusion around
261、responsibilities(DevOps vs.MLOps)and a market still sorting itself outGap between Tracking and FullGap between Tracking and Full-Scale OpsScale Ops The dominance of tracking-first platforms like MLflow and W&B suggests that many teams havent many teams havent yet adopted endyet adopted end-toto-end
262、end MLOpsMLOps suites suites-continuous delivery,drift monitoring,or automated rollback continuous delivery,drift monitoring,or automated rollback remain work in progress for mostremain work in progress for mostMost Used Tools:DevOps and Most Used Tools:DevOps and MLOpsMLOpsPrivate&Strictly Confiden
263、tial60Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey TakeawaysNotes:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Sur
264、vey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkFigmas NearFigmas Near-Universal ReachUniversal Reach With 87%adoption,Figma is e
265、ffectively the deFigma is effectively the de-facto standard for UI/UX and product design facto standard for UI/UX and product design-teams overwhelmingly stick with its real-time collaboration,component libraries,and plugin ecosystem rather than seeking out AI-specific design toolsMiro for HigherMir
266、o for Higher-Level CollaborationLevel Collaboration With 37%adoption,Miro remains the goMiro remains the go-to for wireframing,userto for wireframing,user-journey mapping,and crossjourney mapping,and cross-functional brainstormingfunctional brainstorming;its whiteboard-style interface complements Fi
267、gmas pixel-perfect canvases,especially in early ideation phasesRise of AIRise of AI-Enabled Product WireframesEnabled Product Wireframes Design teams arent yet feeling the urgent need for AI-native product/design platforms,however many are using low/nomany are using low/no-code solutions to Bolt,Lov
268、able,and code solutions to Bolt,Lovable,and VercelVercel V0 for rapid V0 for rapid protoypingprotoypingMost Used Tools:Product and DesignMost Used Tools:Product and DesignPrivate&Strictly Confidential61Most Widely Used ToolsMost Widely Used ToolsFrom survey respondents;By alphabetical orderKey Takea
269、waysNotes:Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services of ICONIQ.Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our communi
270、ty of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkInternal Productivity Use Cases(Part 1 of 2)Internal Productivity Use Cases(Part 1 of 2)Private&Strictly Confidential62Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and pe
271、rspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkSales ProductivitySales ProductivityMarketing Marketing Automation&Automation&Content GenerationConten
272、t GenerationCustomer Customer EngagementEngagementDocumentation and Documentation and Knowledge Knowledge RetrievalRetrievalFor more information on specific tools in each category,please reach out to ICONIQ InsightsUse CaseUse CaseKey TrendsKey TrendsMany teams are getting their AI-powered sales fea
273、tures straight out of Salesforce-indicating that an easy path is to lean on lean on your existing CRMs builtyour existing CRMs built-in recommendations,forecasting,and opportunityin recommendations,forecasting,and opportunity-scoring rather than bolt on a separate servicescoring rather than bolt on
274、a separate serviceOther respondents are also using salessales-engagement platforms engagement platforms like Apollo,Salesloft,Gong,etc,while others are also leaning into AI driven prospecting tools like Clay and People.aiAI driven prospecting tools like Clay and People.aiAs embedded capabilities mat
275、ure,we will likely see consolidation around a handful of platforms or clearer differentiation from consolidation around a handful of platforms or clearer differentiation from the pointthe point-solution upstartssolution upstartsMarketers overwhelmingly turn to Canvas generative features for onCanvas
276、 generative features for on-brand visuals and quick content iterationsbrand visuals and quick content iterations,making it by far the most common“AI”touchpoint in the marketing stackMany respondents are also using solutions like n8n or homegrown solutions,indicating that marketing use cases sometime
277、s marketing use cases sometimes require a high degree of inrequire a high degree of in-house customizationhouse customizationMany respondents are also using specialized AI writing tools like Writer and Jasper,with adoption higher for later stage specialized AI writing tools like Writer and Jasper,wi
278、th adoption higher for later stage companies companies($100M+revenue)Teams overwhelmingly rely on Zendesk or Salesforces embedded AI features for customer interactions,signaling that ease of ease of plugging into existing ticketing and CRM workflows still beats adopting a standalone conversational A
279、I platformplugging into existing ticketing and CRM workflows still beats adopting a standalone conversational AI platformA sizable minority lean on specialist tools like Pylon,Forethought,Grano.la,or Intercom when they need deeper bot specialist tools like Pylon,Forethought,Grano.la,or Intercom when
280、 they need deeper bot customizations,selfcustomizations,self-service wizards,or tight inservice wizards,or tight in-app support widgets app support widgets-suggesting that best-of-breed still has a role when out-of-the-box AI falls shortMost teams either build on existing wikis and note-taking tools
281、 or standardize on Notion;this shows that organizations often default to whatevers already in place for knowledge capture before experimenting with AIdefault to whatevers already in place for knowledge capture before experimenting with AI-powered overlayspowered overlaysHowever,a sizable proportion
282、of respondents are also leaning into purposepurpose-built AI tools like Glean and Writer for indexing built AI tools like Glean and Writer for indexing and semantic searchand semantic searchInternal Productivity Use Cases(Part 2 of 2)Internal Productivity Use Cases(Part 2 of 2)Private&Strictly Confi
283、dential63Source:Perspectives from the ICONIQ GenAI Survey(April 2025)and perspectives from the ICONIQ team and network of AI leaders consisting of our community of CIO/CDOs overseeing AI initiatives in enterprises,CTOs,our Technical Advisory Board,and others in our networkIT&SecurityIT&SecurityLegal
284、LegalHR&RecruitingHR&RecruitingFP&A AutomationFP&A AutomationFor more information on specific tools in each category,please reach out to ICONIQ InsightsUse CaseUse CaseKey TrendsKey TrendsServiceNow(used by 33%of respondents)and Snyk(used by 30%of respondents)lead the pack,showing that large large o
285、rganizations are still defaulting to their existing ITSM and securityorganizations are still defaulting to their existing ITSM and security-scanning platforms rather than standing up brandscanning platforms rather than standing up brand-new AI new AI toolstoolsZapier and Workato were also commonly m
286、entioned,underlining how much teams value lowvalue low-code orchestration for stitching code orchestration for stitching together alerts,ticket creation,and remediation scripts across disparate toolstogether alerts,ticket creation,and remediation scripts across disparate toolsLegal departments are d
287、ipping toes into AI primarily through ChatGPT and ad hoc scripts,but purposedipping toes into AI primarily through ChatGPT and ad hoc scripts,but purpose-built legal assistant built legal assistant platforms are starting to gain tractionplatforms are starting to gain tractionAs regulation and securi
288、ty concerns mount,well likely see a bifurcation:mainstream LLMs for informal research and mainstream LLMs for informal research and compliancecompliance-focused suites for missionfocused suites for mission-critical contract workflowscritical contract workflowsNearly half of teams rely on LinkedIns b
289、uilt-in AI features-profile suggestions,candidate matching,and outreach sequencing-underscoring that recruiters lean on platforms they already use daily rather than integrating standalone solutionsrecruiters lean on platforms they already use daily rather than integrating standalone solutionsHowever
290、,niche platforms like HireVue for AI-driven video interviews and Mercor for candidate engagement are starting to see modest uptakeMany teams are using Ramp for FP&A automation,likely leveraging its spend management and data sync features in an allRamp for FP&A automation,likely leveraging its spend
291、management and data sync features in an all-inin-one platformone platformSpecialized suites like Pigment,Basis,and Tabs are also starting to pick up traction,showing growing interest in drivergrowing interest in driver-based based planning and multiplanning and multi-scenario modeling platformsscena
292、rio modeling platformsAround oneAround one-third of respondents are also using homegrown solutionsthird of respondents are also using homegrown solutions,reflecting investment in custom scripts,Excel macros,and bespoke pipelines to glue together ERP,billing,and BI systemsA global portfolio of catego
293、ryA global portfolio of category-defining businessesdefining businessesPrivate&Strictly Confidential64These companies represent the full list of companies that ICONIQ Venture and Growth has invested in since inception through ICONIQ Strategic Partners funds as of the date these materials were publis
294、hed(except those subject to confidentiality obligations or companies for which the issuer has not provided permission for ICONIQ Venture and Growth to disclose publicly).Trademarks are the property of their respective owners.None of the companies illustrated have endorsed or recommended the services
295、 of ICONIQ.DisclosuresUnless otherwise indicated,the views expressed in this presentation are those of ICONIQ(“ICONIQ”or the“Firm”),are the result of proprietary research,may be subjective,and may not be relied upon in making an investment decision.Information used in this presentation was obtained
296、from numerous sources.Certain of these companies are portfolio companies of ICONIQ.ICONIQ does not make any representations or warranties as to the accuracy of the information obtained from these sources.This presentation is for educational purposes only and does not constitute investment advice or
297、an offer to sell or a solicitation of an offer to buy any securities in connection with any investment fund or investment product that ICONIQ sponsors.Any such offer or solicitation will only be made pursuant to definitive offering documents and subscription agreements.Any reproduction or distributi
298、on of this presentation in whole or in part,or the disclosure of any of its contents,without the prior consent of ICONIQ,is prohibited.This presentation may contain forward-looking statements based on current plans,estimates and projections.The recipient of this presentation(“you”)is cautioned that
299、a number of important factors could cause actual results or outcomes to differ materially from those expressed in,or implied by,the forward-looking statements.The numbers,figures and case studies contained in this presentation have been included for purposes of illustration only,and no assurance can
300、 be given that the actual results of any ICONIQ portfolio company will correspond with the information contained in this presentation.No information is included herein with respect to conflicts of interest,which may be significant.The portfolio companies and other parties mentioned herein may reflec
301、t a selective list of the prior investments made by ICONIQ.Certain of the economic and market information contained herein may have been obtained from published sources and/or prepared by other parties.While such sources are believed to be reliable,none of ICONIQ or any of its affiliates and partner
302、s,employees and representatives assume any responsibility for the accuracy of such information.All of the information herein is presented as of the date made available to you(except as otherwise specified),and is subject to change without notice,and may not be current or may have changed(possibly ma
303、terially)between the date made available to you and the date actually received or reviewed by you.ICONIQ assumes no obligation to update or otherwise revise any information,projections,forecasts or estimates contained in this presentation,including any revisions to reflect changes in economic or mar
304、ket conditions or other circumstances arising after the date the items were made available to you or to reflect the occurrence of unanticipated events.Numbers or amounts herein may increase or decrease as a result of currency fluctuations.For avoidance of doubt,ICONIQ is not acting as an adviser or
305、fiduciary in any respect in connection with providing this presentation and no relationship shall arise between you and ICONIQ as a result of this presentation being made available to you.ICONIQ is a trading name of ICONIQ Partners(UK)LLP.ICONIQ Partners(UK)LLP(Registration Number:973080)is an appoi
306、nted representative of Kroll Securities Ltd.(Registration Number:466588)which is authorised and regulated by the Financial Conduct Authority.ICONIQ Partners(UK)LLP is a limited liability partnership whose members are ICONIQ Capital(UK)Ltd,Seth Pierrepont and Lou Thorne,and it is registered in Englan
307、d and Wales and has its registered office at 27 Soho Square,London W1D 3QR.ICONIQ Partners(UK)LLP acts as an adviser to ICONIQ Capital LLC.These materials are provided for general information and discussion purposes only and may not be relied upon.This material may be distributed to,or directed at,o
308、nly the following persons:(i)persons who have professional experience in matters relating to investments falling within article 19(5)of the Financial Services and Markets Act 2000(Financial Promotion)Order 2005(the“FP Order”),(ii)high-net-worth entities falling within Article 49(2)of the FP Order,an
309、d(iii)any other persons to whom it may otherwise lawfully be communicated(all such persons together being referred to as“FPO Relevant Persons”).Persons who are not FPO Relevant Persons must not act on or rely on this material or any of its contents.Any investment or investment activity to which this
310、 material relates is available only to FPO Relevant Persons and will be engaged in only with FPO Relevant Persons.Recipients must not distribute,publish,reproduce,or disclose this material,in whole or in part,to any other person.Copyright 2025 ICONIQ Capital,LLC.All rights reserved.Private&Strictly Confidential65Private&Strictly ConfidentialTechnology matters.Strategy matters.People matter most.Meet the ICONIQ Venture and Growth team66Private&Strictly ConfidentialSan Francisco|Palo Alto|New York|LondonJoin our communityJoin our communityPrivate&Strictly Confidential