1、Future of the workforce:How AI agents will transform enterpriseworkflowsWe dive into what agents are,their limitations,key companies,andimplications for the future of work.The idea of autonomous AI agents LLM-powered bots that can independently reasonand execute tasks caught on like wildfire in 2023

2、,marking an important evolutionbeyond chatbots and copilots.OpenAI CEO Sam Altman has described agents as“AIs killer function”as recently asMay 2024.Now,the space is seeing a proliferation of startups:More than 50 companies haveemerged since 2022 focusing on agents,agentic workflows,and agent infras

3、tructure.While much of the tech remains limited in its ability to execute tasks reliably,use casesare gaining traction in horizontal enterprise applications like customer support,sales,and engineering.As AI agents evolve to tackle more complex workflows,the implications for enterprisesand their work

4、forces will be far-reaching,from replacing customer service reps andSDRs to augmenting compliance efforts and more.Below,we use CB Insights data to dig into:What AI agents are,current limitations,&the companies emerging in the spaceHorizontal applications gaining tractionIndustry applications on the

5、 horizonLooking aheadFuture of the Workforce|2ContentsThe state of AI agents4How AI agents work5Limitations include complex reasoning&planning and tool use6The AI agent landscape8Horizontal applications dominate8Agent infrastructure tools emerge9Investment surges10Big tech jumps into the fray11Horiz

6、ontal applications&impacts13Customer support reps are at high risk of AI automation13Sales“AI SDRs”is a crowded market15Software development agents look to go beyond copilots18Cybersecurity is an emerging category19General enterprise workflow automation20Emerging industry applications&opportunities2

7、2Financial services23Insurance24Healthcare25Industrials26Gaming27Looking ahead28Note:Companies profiled in this report are working on various levels of autonomy,fromagentic,LLM-powered workflows to fully autonomous agents.Future of the Workforce|3The state of AI agentsGenerative AI has revolutionize

8、d language processing,enabling tasks from creativewriting to code generation,with ChatGPTs 2022 release bringing the tech to hundredsof millions of users.Since then,a range of chatbots and copilots have emerged,leveraging LLMs to answeruser questions,summarize documents,complete software code,and mo

9、re.Agents mark an important evolution beyond chatbots and copilots:AI agents can tacklecomplex tasks on a users behalf.They can use websites,schedule meetings,planvacations,and more.While AI agents havent fully taken off since interest in them surged in 2023,the tech isevolving quickly and becoming

10、more capable with varying degrees of autonomy.Weexamine what agents are,their limitations,and the landscape below.Source:CB Insights news mentions searchFuture of the Workforce|4How AI agents workAutonomous agents are AI systems that can pursue a goal without needing humans totell them what to do an

11、d when.If given an objective,they can create tasks,completethose tasks,and re-prioritize them until reaching the objective.Integral components of these systems include foundation models(for reasoning andanalysis),tool use(to interact with the internet and other software or apps),andmemory access.Ear

12、ly projects released in 2023 highlight this framework,from the open-sourcedAutoGPT,which saw massive interest from the developer community,to its simplercousin BabyAGI.For example,BabyAGI(illustrated below),when given an objective by the user,will loopto execute on it,relying on“agents”(e.g.,OpenAIs

13、 GPT-4 model)to complete tasks,pulling in context from the LLM and storing completed tasks in memory(using a vectordatabase).Source:Yoheinakajima/BabyAGIFuture of the Workforce|5Limitations include complex reasoning&planning and tool useDespite compelling demos from booking flights to conducting mar

14、ket research fullyautonomous agents are still in the early adoption phase.Current agent projects are limited in their ability to execute tasks reliably,especiallythose aiming to work across the entire internet with its array of websites,error states,and unpredictable user interfaces.One of the most

15、advanced software engineer agents Cognitions Devin is able toresolve just 14%of issues unassisted,based on SWE-bench,a benchmark thatevaluates agents ability to resolve real GitHub issues in open-source projects.This stillplaces it far ahead of other models.Source:CognitionMeanwhile,OpenAI reportedl

16、y still considers its technology to be at the“chatbot”stage(AI with conversational language),though it claims it is approaching“reasoners”(problem-solving like a human).The company reckons that the third level full-fledgedagents that can handle complex actions for users has not arrived yet.Future of

17、 the Workforce|6One challenge is getting LLMs to interact accurately with webpages and outside toolsthrough APIs a key feature of agents expected to perform complex tasks acrosssoftware platforms.They also need to chain multiple functions logically,with no roomfor hallucination.For example,AI device

18、s like Rabbits R1 and the Humane Ai Pin,which rely on largemodels,have struggled to actually follow through on tasks,like ordering an Uber orqueuing up songs on Spotify when asked by the user.However,some startups are exploring workarounds,such as optimizing agents tohandle only a limited set of tas

19、ks,keeping humans in the loop to give feedback andsteer the agents approach,or deploying a secondary agent to check the first ones work.For instance,Sierra,which is focused on customer support applications,uses amulti-agent approach.One Sierra customer we interviewed described it as a“checksystem”:T

20、he other thing thats really unique about them is theyve got a checksystem.They dont have just one AI agent working on our project.Theyhave two.Before it actually responds to you,in a nanosecond,it bouncesthe answer off of another AI agent.That agent says,“Yes,this sounds likesomething that the brand

21、 would say,”and then it feeds it back and then itresponds to you.Theres this checks and balances system that allows itto make sure that its actually responding appropriately to thoseanswers.If its not,itll go back and work out the problem with itself,andthen itll go back to the consumer,or itll tran

22、sfer it to a human agent if itdoesnt feel like it can adequately answer it.Director of Marketing,Consumer goods companyFuture of the Workforce|7The AI agent landscapeHorizontal applications dominateAlongside the active community of open-source AI agent projects,the landscape ofVC-backed agent startu

23、ps is dominated by a focus on horizontal applications acrosssales,customer support,and other enterprise and general productivity workflows.Startups focusing on vertical-specific applications remain limited today based on ouranalysis.The vast majority of startups developing agents are targeting busin

24、esses,notconsumers.Theyre vying to replace the roles of SDRs,software developers,complianceanalysts,customer service reps,and more.The implications will be far-reaching from augmenting or replacing companiesemployees day-to-day to eventually becoming the outside face of companies to theircustomers.F

25、uture of the Workforce|8Agent infrastructure tools emergeCompanies in the AI infrastructure space are making it easier for developers to build AIagents.For example,LangChain offers an open-source framework for building LLMapplications,including agents.Meanwhile,Y Combinator-backed Zep AI,whichintegr

26、ates with LangChain,is building“l(fā)ong-term memory”for developers agentapplications.Emergence AI,a division of Merlyn Mind that came out of stealth in June 2024 withnearly$100M in funding,is building an“Orchestrator agent”that routes tasks to the bestLLM or agent.Source:EmergenceStartups are also emer

27、ging to make it easier for AI agents to interact with the web.For example,Browserbases platform allows developers to automate web interactionswith AI and headless browsers(web browsers without a graphical user interface thatcan be controlled programmatically),as highlighted by its CB Insights Scouti

28、ng Report.Future of the Workforce|9Source:CB Insights Browserbase Scouting ReportAnother company here,Anon,was founded to help solve authentication challenges thatAI agents face online,such as entering usernames and passwords securely.Thecompany offers developers a software development kit(SDK)that

29、enables agents toauthenticate on behalf of their users as they navigate websites.Investment surgesCompanies focusing on agents are seeing surging levels of investment and deals in2024 so far(this excludes OpenAI,which works across numerous AI areas).Top-funded companies include Adept($415M in total

30、funding),Imbue($232M),andCognition AI($196M).Future of the Workforce|10Notably,Adepts leadership was effectively acqui-hired by Amazon in June 2024,thoughthe remaining team plans to focus on agentic AI.Source:CB Insights advanced searchThe most active investors in the space are Y Combinator,having b

31、acked 12 startupshere,and Abstract Ventures,with 5.Corporate investors are also eyeing the space:Citi Ventures and New York Life Venturesboth backed compliance-focused Norm Ai,while Workday Ventures and AtlassianVentures invested in Adept.Big tech jumps into the frayMicrosoft,Google,and Amazon all h

32、ave agent projects in motion.Microsoft announced in May 2024 an extension of its Copilot tools as“teammembers,”with new capabilities to develop agents that automate workflows.Googles DeepMind debuted Project Astra at its developer conference in May2024.Project Astra is a“universal AI agent”that func

33、tions as a multimodal AIassistant,able to interact with users via video,image,speech,and text.AlphabetCEO Sundar Pichai described the project on the companys Q224 earnings call asa“glimpse of the future”:Future of the Workforce|11Source:CB Insights Alphabet earnings call transcriptsAmazon brought on

34、 Adepts co-founders and will license its tech as of June2024.In a memo to employees,Rohit Prasad,who heads up Amazons ArtificialGeneral Intelligence(AGI)team,said the move“will accelerate our roadmap forbuilding digital agents that can automate software workflows.”O(jiān)n the investment side,Google Ventu

35、res has backed Hebbia and Cognosys.TheAmazon Alexa Fund also invested in MultiOns January 2024 seed round and ImbuesSeries B.Future of the Workforce|12Horizontal applications&impactsThe vast majority of AI agent startups are focused on automating workflows that crosscompany departments.Leading categ

36、ories,which we dive into below,include customersupport,sales,and software development.Customer support reps are at high risk of AI automationAI has been a feature of the contact center and customer support landscape for years.For example,one buyer we spoke with replaced 135 full-time employees(FTE)i

37、nvolvedin call routing in 2019 by leveraging Googles Contact Center AI Dialogflow solution.The Google Dialogflow solution was part of our voicebot,and whatwas driving the voicebot was a broad,simple use case ofswitchboards.We would take calls into the stores,and then theswitchboard would understand

38、the caller intent and route the callerdirectly to the extension,managing the caller being serviced at thatparticular extension.There were around 135 FTE in this activity,andwe replaced them with a chatbot.As part of the chatbot,we neededan inferencing engine or machine learning/AI to understand theu

39、tterance and match it to a particular intent that we could thenmatch to a routing point to service that specific intent.Manager,British multinational merchandiserSource:CB Insights Google Cloud buyer interviewHowever,chatbots have historically been limited by rigid,rule-based responses and alack of

40、contextual understanding.Future of the Workforce|13Now,LLMs are enabling much deeper automation and cost savings by handling morecomplex conversations and workflows,in turn freeing up agents time(or replacingthem altogether).In February 2024,Klarna announced its OpenAI-powered AI assistant was doing

41、 thework of 700 full-time customer service agents after being live for 1 month.Theassistant is expected to deliver$40M in annual savings.In our conversations with customers,its clear that AI agents will increasingly handle thetasks of human support agents a role in which nearly 3M Americans are empl

42、oyed,asof 2022,per Bureau of Labor Statistics(BLS)data.Sierra,co-founded by former Salesforce CEO Bret Taylor,has raised$110M since itsfounding in 2023 to build customer service AI agents.This Sierra retailer customer isleveraging the companys AI agents to deal with the“need to flex temporary teamsd

43、uring high volume periods”:My stretch goal for everybody is 4.75/5 CSAT.Were consistently seeingthe AI agents and our human agents deliver around a 4.5.Its competingand its on par with what our human agents are doingThe second KPIwere looking at is actual cost per ticket handledI want to make sure t

44、hatthe cost of Sierra is equal to or lower than the human agent cost.Werealso looking at resolution rate the percentage of tickets that the AIagent is getting and resolving itself 100%without needing to escalate to ahuman agent.Right now,were seeing about a 60%resolution rate,whichis great because i

45、ts made me not have to hire the temporary staff on theticket side that Ive hired in previous summers.Director of Marketing,Consumer goods companySource:CB Insights Sierra buyer interviewFuture of the Workforce|14An older group of AI customer support companies,which emerged prior to LLMsmainstream ar

46、rival,are also pivoting to focus on AI agents.These include Ada,Forethought,and Intercom,as well as Cognigy,which raised$100M in Series C fundingin June 2024 to scale its contact center agents.Looking ahead,expect voice solutions to come more into focus.For instance,Fixie,which initially developed c

47、hat-based AI agents,has switched its focus to real-time voiceinteractions.Meanwhile,this Sierra buyer said,“I expect our AI usage to really increase as Sierra onboards voice,which theyrethinking about doing by the end of 2024.Im considering not turning it on full time,but potentially turning it on a

48、nd off hours when our agents arent working to be ableto solve those consumer inquiries when we dont have active phone supportavailable.”O(jiān)ther buyers are looking to leverage AI agents to bring new levels of personalization totheir customers outside of just conversational support interactions:“How can

49、 we leverage AI to predict preferences or lead to tailored productrecommendations or personalization in marketing,etc.?I see thats one potentialarea of opportunity that could be a part of a customer experience Sierra model.”Sales“AI SDRs”is a crowded marketOne of the most active categories in terms

50、of company activity is startups developingautonomous sales reps that can scale outbound automation.They research,route,andqualify leads,personalize messages,and book meetings.Clay($62M in total funding),for example,is developing Claygent,”its AI tool focusingon data enrichment and sales prospecting.

51、Other established companies focusing onbroader SDR workflows include Qualified($163M in total funding),which launchedPiper,its AI SDR,in April 2024.Future of the Workforce|15Source:CB Insights AI SDRs market reportFounded just last year,11xAI has scaled to$2M in ARR as of March 2024 with Alice,itsAI

52、 SDR.The company primarily offers a task-based pricing model charging forcompleted tasks like identifying and researching accounts,preparing outreach,andscheduling meetings.For example,one Sales Operation Manager whos a customer of 11xAI pointed to theusefulness of the tool to replace SDRs,but also

53、noted concerns about AI-generatedemail fatigue:Future of the Workforce|16Source:CB Insights 11xAI buyer interviewAnother common point of concern is the trustworthiness and quality of the messaging:Clay customer(CSAT:7/10):“Another way is that when were using the Claygentcapabilities,all the differen

54、t AI capabilities that you can use,even though itshows its work and it explains the logic or the rationale that its using,a lot oftimes,we just arent fully trustworthy of it yet.We have humansdouble-checking every single thing.It does the work for you,but we dont trust ityet.”11xAI customer(CSAT:7/1

55、0):“I think we lost a little faith in the AIs ability tocreate content that we,with peace in mind,could send out to potentialprospects,which I think has taken a little bit of the groundbreaking smartnessout of the tool for us because that actually means that we would just be able toset up campaigns

56、in any other cadence tools and run them through them.”11xAI customer(CSAT:5/10):“We were also selling more to the lower,middlemarkets or lower companies,100 to 150 employees,which again feels like wewould get more return.The challenge why Im giving it a 5 is we are actuallyshifting to work on enterp

57、rise companies and thereby the messaging,theFuture of the Workforce|17complexity of messaging is a lot more sophisticated.Therefore again,thereturns arent there compared to if you would message the lower/mid-market.”The prospect of automating many of the repetitive tasks SDRs do every day is driving

58、activity and interest in the market.However,more sophisticated messaging andpersonalization,and better incorporation of intent signals,will be key for buyersevaluating solutions in the market.Software development agents look to go beyond copilotsCoding copilots are becoming standard-issue tools at e

59、nterprises.Oil&gas major BP said on a recent earnings call that it needs“70%less coders from3rd parties”thanks to AI,as seen on the CBI platform below.Source:CB Insights BP earnings call transcriptsFuture of the Workforce|18The AI-powered coding assistant scene has exploded,with startups raising mas

60、sivefunding,alongside products from big tech like Microsofts GitHub Copilot,Metasopen-source Code Llama,and Amazons CodeWhisperer.Companies like Cognition and Magic are aiming for true coding agents that can replacethe work of human engineers as opposed to assistants that suggest code snippets.The t

61、ech is still early,and quality remains a major challenge.This Magic customer an AI cloud software developer lead at a Fortune 500 company highlighted 3 issues in a recent interview:1)lack of originality in the written code;2)privacy and data concerns;and 3)excess“fat”in the code.But the opportunity

62、is large.There are 1.8M software engineers in the US with a medianpay of$130K per year,per BLS data,driving a combined annual spend of over$230B.Given the high cost and extensive need for developers,we anticipate continuedinvestment and adoption in the space.Expect other companies in the copilot spa

63、ce tobegin developing agentic workflows as well.Cybersecurity is an emerging categoryOne emerging horizontal application of agents is in security operations for example,investigating alerts,generating reports,and eventually deploying fixes for vulnerabilities.Nullify is building an“AI security engin

64、eer,”while Dropzone AI is focused on creating AISOC(security operations center)analysts,as highlighted by CBIs Funding Insights forits recent Series A.Future of the Workforce|19Source:CB Insights Dropzone AI Funding InsightsGeneral enterprise workflow automationOne of the most robust categories of a

65、gent companies is those targeting a wide rangeof enterprise workflows.Applications span general productivity and research use cases.Many startups in the category also highlight applications across functions like sales,customer service,and HR.Future of the Workforce|20Select VC-backed enterprise work

66、flow AI agentsSource:CB Insights advanced searchEnterprise adoption is already starting to happen,based on company customer data andour buyer interviews.For example,Accel-backed Ema,a 2024 CB Insights AI 100 winner,advertises itself as a“universal AI employee”targeting a range of functions and indus

67、tries.The companyscustomers include TrueLayer and Moneyview.Another company here is early-stage Respell,which helps businesses build workflowsusing models from OpenAI,Cohere,Anthropic,and other providers.One Respellcustomer we spoke with a director of data engineering at a$50M+funded dataplatform le

68、verages Respell for a number of sales and marketing use cases:Future of the Workforce|21Source:CB Insights Respell buyer interviewA number of other early-stage companies in the space are working more specifically onweb research and data extraction,including Zeta Labs,Reworkd,and Lutra AI.Emerging in

69、dustry applications&opportunitiesWhile few agentic companies focus on single industries,we expect to see morecompanies“niching down”in the coming year,mirroring trends in the broader generativeAI space.Future of the Workforce|22Agents reliability is a bigger hurdle for sensitive,highly regulated ind

70、ustries like financialservices and healthcare.Fully autonomous workflows may be years away,butcorporations should monitor developments among companies in the space.Below,we highlight emerging applications or opportunity areas across 5 industries.Financial servicesWhile still early,two primary applic

71、ations of agents in financial services are on thehorizon:1)compliance and 2)investment research.In compliance,Norm Ai whose backers include notable finserv CVCs and incumbentsCiti Ventures,New York Life Ventures,and TIAA is building agents trained onregulatory filings to make compliance determinatio

72、ns.Another company here,Parcha AI,which was founded by former Coinbase and Brexemployees,emerged to build AI agents focused on compliance workflows in fintech andbanking.The company explained in a blog post the challenges of building“reliable agenticbehavior”for something like compliance workflow au

73、tomation:“While the concept of agentic behavior was promising,building reliable agenticbehavior with large language models(LLMs)was a massive endeavor.Creatinggeneral-purpose autonomous agents could have taken us years.Ourcustomers would much rather have a solution that was very accurate and reliabl

74、efor a subset of tasks than a fully autonomous solution that could automate aworkflow end-to-end but worked only 80%of the time.”Parcha has since pivoted to home in on LLM-powered workflows for KYC/KYB and duediligence rather than having a fully autonomous agent determine and take actions.With the r

75、eliability question still unsolved,more limited workflows will be in focus.Thatincludes investment research.Hebbia,which raised a$130M Series B led by a16z in May 2024,is building agents forfinancial services.For example,its Matrix product builds spreadsheets that mineinformation from files(in rows)

76、and deliver answers to questions(in columns).Future of the Workforce|23Source:Hebbia via a16zInsuranceApplications of agents in insurance include underwriting and claims processing.Tech providers are already eyeing genAI-powered underwriting workflows.Future of the Workforce|24Source:CB Insights Gui

77、dewire Earnings InsightsThe next step will be more agentic workflows.For example,Roots Automation whichoffers an insurance-focused LLM centered on unstructured data and an AI underwritingassistant raised a funding round from Erie Strategic Ventures in January 2024.HealthcareBased on our analysis of

78、the market,autonomous solutions in healthcare have yet toemerge though some agent companies like Ema advertise healthcare-specificapplications such as a pharmacy assistant.Monitor the medical scribes market for the development of more advanced workflows an already active generative AI market where c

79、opilots are automating tedious tasks likedocumentation.Future of the Workforce|25Source:CB Insights Virtual medical scribes&summarization tools marketIndustrialsCompanies are developing industrial autonomous agents that optimize processes andequipment including control systems,robots,and other indus

80、trial machines withoutneeding humans.For example,Composabl launched its autonomous agent in May 2024.It uses LLMs toautomatically control and optimize industrial equipment,with early adopters likeRockwell Automation and RoviSys already experiencing success.In the long term,as humanoid robot developm

81、ent progresses,humanoids are likely togain even more advanced decision-making capabilities(i.e.,agentic behaviors)toovercome unexpected challenges or adapt to their environments in ways thatdramatically increase their autonomy and effectiveness.Learn more about applications of humanoid robots here.F

82、uture of the Workforce|26GamingWhile a number of companies are developing genAI-powered NPCs(non-playercharacters)for games,some are going a step further.DeepMind published research in March 2024 on a“generalist AI agent for 3D virtualsettings.”The agent can understand and navigate different game en

83、vironments as wellas carry out tasks as a human could.Meanwhile,Altera is a research lab working on“digital human beings.”Its first product isan agent that can play alongside users in Minecraft.The company raised$9M in seedfunding in April 2024.Source:CB Insights Altera Funding InsightsFuture of the

84、 Workforce|27Expect to see far more activity and research related to gaming,given it offers apre-made digital playground already optimized for humans to engage with.Developments here may have implications for the broader AI agent space.As DeepMindhighlighted in March,“Video games are a key proving g

85、round for artificial intelligence(AI)systems.Like the real world,games are rich learning environments with responsive,real-time settings and ever-changing goals.”Looking aheadMore companies currently in the workflow automation and copilot/assistant space willmove toward becoming(or at least trying t

86、o become)fully autonomous solutions.At the same time,given that many of these agents are unproven and still largelyunreliable we remain years away from widespread enterprise adoption.Lack of trust will be the biggest barrier to adoption.In response,companies like Vijil which raised$6M in funding fro

87、m investors including Googles Gradient Ventures in July2024 are building security solutions to evaluate AI agents behavior andtrustworthiness.Eventually,LLMs may become sophisticated enough to solve the reasoning andplanning hurdles that agents currently face.Though another intelligence architecture

88、altogether may be necessary to solve the“non-linguistic aspects to thinking andreasoning,”as Sequoia Capitals Sonya Huang and Pat Grady write,to achieve higheraccuracy.In the meantime,expect some shaking out of the space the most notable recentexample being Amazons acqui-hire of Adepts leadership.Ot

89、her companies like Imbue,which raised over$200M in 2023,have yet to release a product.But the promise of an empowered digital workforce remains a trend that wouldreshape industries as we know them.As AI agents proliferate across enterprises,multi-agent architectures will emerge to manage them effect

90、ively,such as having a leadagent coordinate with specialized sub-agents.The implications would be enormous,changing how companies hire and scale,not to mention how they interact with eachother.Enterprises must pay attention to these developments or face the risk of beingoutcompeted by companies that do.Future of the Workforce|28