1、 2023 Boston Consulting Group1The release of ChatGPT in late 2022 created a groundswell of interest in generative AI.Within hours,users experimenting with this new technology had discovered and shared myriad productivity hacks.Inthe weeks and months since,organizations have scrambled to keep paceand

2、 to defend againstunforeseen complications.Some organizations have already adopted a more formal approach,creating dedicated teams to explore how generative AI can unlock hidden value and improve efficiency.The CEOs Guide to the Generative AIRevolutionMARCH 07,2023 By Franois Candelon,Abhishek Gupta

3、,Lisa Krayer,and Leonid ZhukovREADING TIME:15 MIN 2023 Boston Consulting Group2For CEOs,however,generative AI poses a much bigger challenge.Todays focus might be onproductivity gains and technical limitations,but a revolution in business-model innovation is coming.Much as Mosaic,the worlds first fre

4、e web browser,ushered in the internet era and upended the waywe work and live,generative AI has the potential to disrupt nearly every industrypromising bothcompetitive advantage and creative destruction.The implication for leaders is clear:todays breathlessactivity needs to evolve into a generative

5、AI strategy owned by the C-suite.This is no small task,and CEOswho are likely several steps removed from the technology itselfmay feel uncertain about their next move.But from our perspective,the priority for CEOs isnt to fullyimmerse themselves in the technology;instead,they should focus on how gen

6、erative AI will impacttheir organizations and their industries,and what strategic choices will enable them to exploitopportunities and manage challenges.These choices are centered on three key pillars:Each pillar raises an urgent question for CEOs.What innovations become possible when everyemployee

7、has access to the seemingly infinite memory generative AI offers?How will this technologychange how employees roles are defined and how they are managed?How do leaders contend withthe fact that generative AI models may produce false or biased output?Clearly,generative AI is a rapidly evolving space,

8、and each of the pillars above involves short-and long-term considerationsand many other unanswered questions.But CEOs need to prepare for themoment when their current business models become obsolete.Heres how to strategize for that future.Potential:Discover Your Strategic AdvantageAI has never been

9、so accessible.Tools such as ChatGPT,DALL-E 2,Midjourney,and Stable Diffusionallow anyone to create websites,generate advertising strategies,and produce videosthe possibilitiesare limitless.This“low-code,no-code”quality will also make it easier for organizations to adopt AIcapabilities at scale.(See“

10、The Functional Characteristics of Generative AI.”)THE FUNCTIONAL CHARACTERISTICS OF GENERATIVE AI 2023 Boston Consulting Group3The transformative potential of generative AI can be summed up by three keyfunctional characteristics.Seemingly“Infinite”Memory and Pattern Recognition.Because generative AI

11、 istrained on huge amounts of data,its memory can appear infinite.For example,ChatGPT has been trained on a massive portion of publicly available information onthe internet.To put this in context,as of 2018 the internet generated 2.5 quintillionbytes of new data daily,according to Domothe equivalent

12、 of 1.2 quintillion words.That number is likely much higher today.Generative AI can also create connections(orrecognize patterns)between distant concepts in an almost human-like manner.Low-Code,No-Code Properties.When describing the impact of ChatGPT,AndrejKarpathy,a founding member of OpenAI,said“t

13、he hottest new programminglanguage is English.”Thats because generative AIs natural-language-processinginterface allows nonexperts to create applications with little or no coding required.Bycontrast,coding assistant systems such as Github Copilot still require competentprogrammers to operate them.La

14、ck of a Credible Truth Function.Generative AIs“infinite”memory can become aninfinite hallucination.In reality,the level of error in todays generative AI systems isan expected characteristic that makes it useful for generating new ideas and content.But because generative AI does not use logic or inte

15、lligent thought,instead predictingthe most probable next word based on its training data,it should only be used togenerate first dras of content.Companies are working to make generative AIs output significantly more reliable byusing an approach known as reinforcement learning with human feedback;oth

16、erapproaches that combine generative AI with traditional AI and machine learning havealso been considered.Improvements to generative AI are expected soon,with somepredicting that it will be able to produce final-dra content by 2030.The immediate productivity gains can greatly reduce costs.Generative

17、 AI can summarize documents ina matter of seconds with impressive accuracy,for example,whereas a researcher might spend hourson the task(at an estimated$30 to$50 per hour).2023 Boston Consulting Group4But generative AIs democratizing power also means,by definition,that a companys competitors willhav

18、e the same access and capabilities.Many use cases that rely on existing large language model(LLM)applicationssuch as productivity improvements for programmers who use Github Copilotand for marketing content developers who use Jasper.aiwill be needed just to keep pace with otherorganizations.But they

19、 wont offer differentiation,because the only variability created will result fromusers ability to prompt the system.Identify the Right Use CasesFor the CEO,the key is to identify the companys“golden”use casesthose that bring true competitiveadvantage and create the largest impact relative to existin

20、g,best-in-class solutions.Such use cases can come from any point along the value chain.Some companies will be able to drivegrowth through improved offerings;Intercom,a provider of customer-service solutions,is runningpilots that integrate generative AI into its customer-engagement tool in a move tow

21、ard automation-firstservice.Growth can also be found in reduced time-to-market and cost savingsas well as in the abilityto stimulate the imagination and create new ideas.In biopharma,for example,much of todays 20-year patent time is consumed by R&D;accelerating this process can significantly increas

22、e a patentsvalue.In February 2021,biotech company Insilico Medicine announced that its AI-generatedantifibrotic drug had moved from conceptualization to Phase 1 clinical trials in less than 30 months,for around$2.6 millionseveral orders of magnitude faster and cheaper than traditional drugdiscovery.

23、Once leaders identify their golden use cases,they will need to work with their technology teams tomake strategic choices about whether to fine-tune existing LLMs or to train a custom model.(SeeExhibit 1.)1 2023 Boston Consulting Group5Fine-Tuning an Existing Model.Adapting existing open-source or pa

24、id models is cost effectivein a2022 experiment,Snorkel AI found that it cost between$1,915 and$7,418 to fine-tune a LLM model tocomplete a complex legal classification.Such an application could save hours of a lawyers time,whichcan cost up to$500 per hour.Fine-tuning can also jumpstart experimentati

25、on,whereas using in-house capabilities will siphon offtime,talent,and investment.And it will prepare companies for the future,when generative AI is likelyto evolve into a model like cloud services:a company purchases the solution with the expectation ofachieving quality at scale from the cloud provi

26、ders standardization and reliability.But there are downsides to this approach.Such models are completely dependent on the functionalityand domain knowledge of the core models training data;they are also restricted to availablemodalities,which today are comprised mostly of language models.And they of

27、fer limited options forprotecting proprietary datafor example,fine-tuning LLMs that are stored fully on premises.Training a New or Existing Model.Training a custom LLM will offer greater flexibility,but it comeswith high costs and capability requirements:an estimated$1.6 million to train a 1.5-billi

28、on-parametermodel with two configurations and 10 runs per configuration,according to AI21 Labs.To put thisinvestment in context,AI21 Labs estimated that Google spent approximately$10 million for trainingBERT and OpenAI spent$12 million on a single training run for GPT-3.(Note that it takes multipler

29、ounds of training for a successful LLM.)2 2023 Boston Consulting Group6These costsas well as data center,computing,and talent requirementsare significantly higher thanthose associated with other AI models,even when managed through a partnership.The bar to justifythis investment is high,but for a tru

30、ly differentiated use case,the value generated from the modelcould offset the cost.Plan Your InvestmentLeaders will need to carefully assess the timing of such an investment,weighing the potential costs ofmoving too soon on a complex project for which the talent and technology arent yet ready agains

31、t therisks of falling behind.Todays generative AI is still limited by its propensity for error and shouldprimarily be implemented for use cases with a high tolerance for variability.CEOs will also need toconsider new funding mechanisms for data and infrastructurewhether,for example,the budgetshould

32、come from IT,R&D,or another sourceif they determine that custom development is a criticaland time-sensitive need.The“fine-tune versus train”debate has other implications when it comes to long-term competitiveadvantage.Previously,most research on generative AI was public and models were provided thro

33、ughopen-source channels.Because this research is now moving behind closed doors,open-source modelsare already falling far behind state-of-the-art solutions.In other words,were on the brink of agenerative AI arms race.(See“The Future State of the LLM Market.”)Until recently,most generative AI researc

34、h has been publicly accessible.But manycompanies are choosing to stop or delay publishing their research findings and arekeeping model architectures as proprietary knowledge.(For example,GPT-2 was open-source but GPT-3 is proprietary.)The next improvements to generative models with vast number of us

35、ers will likelycome from logs of their user interaction,giving these models a significant competitiveadvantage over new entrants.This reality,combined with the heavy data,infrastructure,and talent costs required to train LLMs,means that the LLM markethas both economy and quality of scale.Advances in

36、 generative AI therefore might belimited to large companies,while the democratization of AI development for smalland medium-sized enterprises could be limited to nondifferentiated use cases.The jury is still out,but this dynamic appears comparable to the“search-engine wars.”Several large companies i

37、nvested heavily in search solutions,but Googles user-THE FUTURE OF THE LLM MARKET 2023 Boston Consulting Group7friendliness and accuracy helped set it apart from competitors.Once users preferredGoogle,other engines could not keep upbecause every search request Googlereceived made it better and smart

38、er.Soon,all other B2C solutions faded away.Asimilar winner-take-all situation could play out in the LLM market,with the big,earlyentrants eventually owning the models and having full control over access.A winner-take-all situation could play out in the LLM market.Its worth noting,however,that Google

39、 did not achieve the same level of success in theenterprise search market,which has unique requirements and challenges compared toB2C.At the enterprise level,search-engines lack the scale to build domain expertiseand lack the volume of user data to build that capability.Similarly,businesses will get

40、the most value out of LLMs that are trained on their proprietary data and that havemodalities that drive unique use cases.This could make it difficult for any single playerto dominate the B2B market.There is also the potential for companies and governments to fund open-sourcemodels to keep them stat

41、e of the artsimilar to how IBM funded Linux.These market dynamics have key implications for CEOs as they make customizationand implementation decisions:It is unlikely that any single LLM provider will dominate the B2B market;the keyfor companies is to find large models with the modality and function

42、ality thatmatch their golden use cases or use cases that require sensitive data.While training LLMs is an option for large businesses,the quality of scale couldmake purchasing solutions more reliable(similar to cloud).2023 Boston Consulting Group8As research accelerates and becomes more and more pro

43、prietary,and as the algorithms becomeincreasingly complex,it will be challenging to keep up with state-of-the-art models.Data scientists willneed special training,advanced skills,and deep expertise to understand how the models worktheircapabilities,limitations,and utility for new business use cases.

44、Large players that want to remainindependent while using the latest AI technology will need to build strong internal tech teams.People:Prepare Your Workforce Like existing forms of artificial intelligence,generative AI is a disruptive force for humans.In the nearterm,CEOs need to work with their lea

45、dership teams as well as HR leaders to determine how thistransformation should unfold within their organizationsredefining employees roles andresponsibilities and adjusting operating models accordingly.Redefine Roles and ResponsibilitiesSome AI-related shis have already occurred.Traditional AI and m

46、achine-learning algorithms(sometimes incorrectly referred to as analytical AI),which use powerful logic or statistics to analyzedata and automate or augment decision making,have enabled people to work more autonomouslyand managers to increasingly focus on team dynamics and goal setting.Now generativ

47、e AI,in its capacity as a first-dra content generator,will augment many roles byincreasing productivity,performance,and creativity.Employees in more clerical roles,such asparalegals and marketers,can use generative AI to create first dras,allowing them to spend more oftheir time refining content and

48、 identifying new solutions.Coders will be able to focus on activities suchas improving code quality on tight timelines and ensuring compliance with security requirements.Of course,these changes cannot(and should not)happen in a vacuum.CEOs need to be aware of theeffect that AI has on employees emoti

49、onal well-being and professional identity.Productivityimprovements are oen conflated with reduction in overall staff,and AI has already stoked concernamong employees;many college graduates believe AI will make their job irrelevant in a few years.Butits also possible that AI will create as many jobs

50、as it will displace.The impact of AI is thus a critical culture and workforce issue,and CEOs should work with HR tounderstand how roles will evolve.As AI initiatives roll out,regular pulse checks should be conducted to If choosing to train in-house,be wary of relying too much on individualresearcher

51、s.If only a small number of people have the expertise to advance andmaintain the model,this will cause a single point of failure if those researcherschoose to leave.2023 Boston Consulting Group9track employee sentiment;CEOs will also need to develop a transparent change-management initiativethat wil

52、l both help employees embrace their new AI coworkers and ensure employees retain autonomy.The message should be that humans arent going anywhereand in fact are needed to deploy AIeffectively and ethically.(See Exhibit 2.)As AI adoption accelerates,CEOs need to learn as they go and use those lessons

53、to develop a strategicworkforce planin fact,they should start creating this plan now and adapt it as the technologyevolves.This is about more than determining how certain job descriptions will changeits aboutensuring that the company has the right people and management in place to stay competitive a

54、ndmake the most out of their AI investments.Among the questions CEOs should ask as they assess theircompanys strengths,weaknesses,and priorities are:What competencies will project leaders need to ensure that individual contributors work is ofsufficient quality?How can CEOs create the optimal experie

55、nce curve to produce the right future talent pipelineensuring,for example,that employees at a more junior level are upskilled in AI augmentationand that supervisors are prepared to lead an AI-augmented workforce?How should training and recruiting be adjusted to build a high-performing workforce now

56、and inthe future?2023 Boston Consulting Group10Adjust Your Operating ModelWe expect that agile(or bionic)models will remain the most effective and scalable in the long term,butwith centralized IT and R&D departments staffed with experts who can train and customize LLMs.Thiscentralization should ensu

57、re that employees who work with similar types of data have access to thesame data sets.When data is siloed within individual departmentsan all-too-common occurrencecompanies will struggle to realize generative AIs true potential.But under the right conditions,generative AI has the power to eliminate

58、 the compromise between agility and scale.Because of the increased importance of data science and engineering,many companies will benefitfrom having a senior executive role(for example,a chief AI officer)oversee the business and technicalrequirements for AI initiatives.This executive should place sm

59、all data-science or engineering teamswithin each business unit to adapt models for specific tasks or applications.Technical teams will thushave the domain expertise and direct contact to support individual contributors,ideally limiting thedistance between the platform or tech leaders and individual

60、contributors to one layer.Structurally,this could involve department-focused teams with cross-functional members(for example,sales teams with sales reps and dedicated technical support)or,preferably,cross-departmental andcross-functional teams aligned to the business and technical platforms.Policies

61、:Protect Your BusinessGenerative AI lacks a credible truth function,meaning that it doesnt know when information isfactually incorrect.The implications of this characteristic,also referred to as“hallucination,”can rangefrom humorous foibles to damaging or dangerous errors.But generative AI also pres

62、ents other criticalrisks for companies,including copyright infringement;leaks of proprietary data;and unplannedfunctionality that is discovered aer a product release,also known as capability overhang.(See Exhibit3.)For example,Riffusion used a text-to-image model,Stable Diffusion,to create new music

63、 byconverting music data into spectrograms.2023 Boston Consulting Group11Prepare for RiskCompanies need policies that help employees use generative AI safely and that limit its use to cases forwhich its performance is within well-established guardrails.Experimentation should be encouraged;however,it

64、 is important to track all experiments across the organization and avoid“shadowexperiments”that risk exposing sensitive information.These policies should also guarantee clear dataownership,establish review processes to prevent incorrect or harmful content from being published,and protect the proprie

65、tary data of the company and its clients.Another near-term imperative is to train employees how to use generative AI within the scope of theirexpertise.Generative AIs low-code,no-code properties may make employees feel overconfident intheir ability to complete a task for which they lack the requisit

66、e background or skills;marketing staff,for example,might be tempted to bypass corporate IT rules and write code to build a new marketingtool.About 40%of code generated by AI is insecure,according to NYUs Center for Cybersecurityandbecause most employees are not qualified to assess code vulnerabiliti

67、es,this creates a significantsecurity risk.AI assistance in writing code also creates a quality risk,according to a Stanford Universitystudy,because coders can become overconfident in AIs ability to avoid vulnerabilities.Leaders therefore need to encourage all employees,especially coders,to retain a

68、 healthy skepticism ofAI-generated content.Company policy should dictate that employees only use data they fullyunderstand and that all content generated by AI is thoroughly reviewed by data owners.Generative AIapplications(such as Bing Chat)have already started implementing the ability to reference

69、 sourcedata,and this function can be expanded to identify data owners.2023 Boston Consulting Group12Ensure Quality and SecurityLeaders can adapt existing recommendations regarding responsible publication to guide releases ofgenerative AI content and code.They should mandate robust documentation and

70、set up aninstitutional review board to review a priori considerations of impact,akin to the processes forpublishing scientific research.Licensing for downstream uses,such as the Responsible AI License(RAIL),presents another mechanism for managing the generative AIs lack of a truth function.Finally,l

71、eaders should caution employees against using public chatbots for sensitive information.Allinformation typed into generative AI tools will be stored and used to continue training the model;evenMicroso,which has made significant investments in generative AI,has warned its employees not toshare sensit

72、ive data with ChatGPT.Today,companies have few ways to leverage LLMs without disclosing data.One option for dataprivacy is to store the full model on premises or on a dedicated server.(BLOOM,an open-sourcemodel from Hugging Faces BigScience group,is the size of GPT-3 but only requires roughly 512gig

73、abytes of storage.)This may limit the ability to use state-of-the-art solutions,however.Beyondsharing proprietary data,there are other data concerns when using LLMs,including protectingpersonally identifiable information.Leaders should consider leveraging cleaning techniques such asnamed entity reco

74、gnition to remove person,place,and organization names.As LLMs mature,solutions to protect sensitive information will also gain sophisticationand CEOs should regularlyupdate their security protocols and policies.Generative AI presents unprecedented opportunities.But it also forces CEOs to grapple wit

75、h toweringunknowns,and to do so in a space that may feel unfamiliar or uncomfortable.Craing an effectivestrategic approach to generative AI can help distinguish the signal from the noise.Leaders who areprepared to reimagine their business modelsidentifying the right opportunities,organizing theirwor

76、kforce and operating models to support generative AI innovation,and ensuring thatexperimentation doesnt come at the expense of security and ethicscan create long-term competitiveadvantage.ABOUT BOSTON CONSULTING GROUPBoston Consulting Group partners with leaders in business and society to tackle the

77、ir most importantchallenges and capture their greatest opportunities.BCG was the pioneer in business strategy when it wasfounded in 1963.Today,we work closely with clients to embrace a transformational approach aimed atbenefiting all stakeholdersempowering organizations to grow,build sustainable com

78、petitive advantage,and drive positive societal impact.Our diverse,global teams bring deep industry and functional expertise and a range of perspectives that 2023 Boston Consulting Group13question the status quo and spark change.BCG delivers solutions through leading-edge managementconsulting,technol

79、ogy and design,and corporate and digital ventures.We work in a uniquely collaborativemodel across the firm and throughout all levels of the client organization,fueled by the goal of helping ourclients thrive and enabling them to make the world a better place.Boston Consulting Group 2023.All rights r

80、eserved.For information or permission to reprint,please contact BCG at .To find the latestBCG content and register to receive e-alerts on this topic or others,please visit .Follow BostonConsulting Group on Facebook and Twitter.Authors1Large language models,also known as foundation models,are deep-le

81、arning algorithms thatcan recognize,summarize,translate,predict,and generate content based on its trainingdata.Today these models are mostly trained on text,images,and audio,but they can also gobeyond language and images into signals,biological data,and more.Models trained on databeyond language are

82、 called multimodal models.2“How Generative AI Is Changing Creative Work,”Harvard Business Review,November 14,2022.Franois CandelonMANAGING DIRECTOR&SENIOR PARTNER;GLOBALDIRECTOR,BCG HENDERSONINSTITUTEParisAbhishek GuptaSENIOR SOLUTION DELIVERYMANAGER,RESPONSIBLE AIMontrealLisa KrayerPROJECT LEADERWashington,DCLeonid ZhukovVICE PRESIDENT-DATASCIENCENew York