《麥肯錫：2024從LLM(大語言模型)到ROI(投資回報率)：如何在零售業中擴大生成式AI的采用規模研究報告（英文版）（10頁）.pdf》由會員分享，可在線閱讀，更多相關《麥肯錫：2024從LLM(大語言模型)到ROI(投資回報率)：如何在零售業中擴大生成式AI的采用規模研究報告（英文版）（10頁）.pdf（10頁珍藏版）》請在三個皮匠報告上搜索。

1、August 2024Retail PracticeLLM to ROI:How to scale gen AI in retailRetailers continue to experiment with generative AI and are seeing that the technology holds great promise for reigniting growthif they move quickly to seize the opportunity.This article is a collaborative effort by Alexander Sukharev

2、sky,Andreas Ess,Denis Emelyantsev,Emily Reasor,and Holger Hrtgen,with Oleg Sokolov and Sergey Kondratyuk,representing views from QuantumBlack,AI by McKinsey,and McKinseys Retail Practice.Once generative AI(gen AI)hit the mainstream,in late 2022,it took little time for retail executives to realize th

3、e potential in front of them.Mentions of artificial intelligence(AI)in retailers earnings calls soared last yearwhich was no surprise,given that gen AI is poised to unlock between$240 billion to$390 billion in economic value for retailers,equivalent to a margin increase across the industry of 1.2 to

4、 1.9 percentage points.This,combined with the value of nongenerative AI and analytics,could turn billions of dollars in value into trillions.Over the past year,most retailers have started testing different gen AI use cases across the retail value chain.Even with all this experimentation,however,few

5、companies have managed to realize the technologys full potential at scale.We surveyed more than 50 retail executives,and although most say they are piloting and scaling large language models(LLMs)and gen AI broadly,only two executives say they have successfully implemented gen AI across their organi

6、zations(see sidebar,“Our survey findings”).Some retailers have found it difficult to implement gen AI widely because it requires rewiring parts of the retail organization,such as technical capabilities and talent.Data quality and privacy concerns,insufficient resources and expertise,and implementati

7、on expenses have also challenged the speed at which retailers can scale their gen AI experiments.Retail companies that have succeeded in harnessing gen AIs power typically excel in two key areas.First,they consider how gen AI use cases can help transform specific domains rather than spreading their

8、resources too thin across a range of scenarios.Second,they effectively transition from pilot and proof-of-concept to deployment at scale.This requires not just data prioritization and technological integration but also significant organizational changes to support widespread AI adoption.In this arti

9、cle,we explore which use cases can offer the most value and what organizational transformations are necessary to scale these technologies successfully.In April 2024,we conducted a survey of 52global Fortune 500 retail executives.Our survey focused on the progress retailers made in exploring and expe

10、rimenting with generative AI(genAI).We found that most retail executives(90percent)say they began experimenting with gen AI solutions and scaling priority use cases and that these experiments had knock-on effects across their other AI initiatives;two-thirds of retail leaders say they want to invest

11、in and focus more on data and analytics.Sixty-four percent of retail leaders say they have conducted gen AI pilots that have augmented their organizations internal value chains,while 26 percent say they are already scaling gen AI solutions in this area.Whats more,82 percent of retailers say they hav

12、e conducted pilots for gen AI use cases related to the reinvention of customer service.Thirty-six percent say they are scaling gen AI solutions in this area.Off-the-shelf gen AI tools have become more readily available in the past year.More than half of retail leaders surveyed(60 percent)opted for r

13、eady-made platforms,although the adoption rate of these third-party platforms is lower in areas such as procurement(18 percent)and commercial(25 percent).The adoption of third-party gen AI solutions will likely grow as the gen AI platform market Our survey findingsmatures.Two-thirds of retailers say

14、 they intend to increase their gen AI budgets over the next year.Meanwhile,10 percent of retailers say they are adopting a wait-and-see approach to gen AI.They plan to integrate gen AI into their operations at a later date,particularly in the areas where there are no full-service platforms yet.This

15、decision may stem from factors such as having insufficient expertise or organizational resources,issues related to data quality and privacy,as well as the expenses associated with gen AI implementation.2LLM to ROI:How to scale gen AI in retailFrom the inside out:Two ways gen AI transforms retailReta

16、ilers we spoke with have already piloted gen AI use cases within their internal value chains,and some are even beginning to scale gen AI solutions.Gen AI can help streamline operations,allowing leaders to make faster,better-informed decisions across retailers internal value chains.The technology als

17、o offers immediate,no-regret efficiency gains,as well as applications that could redefine decision making in retail (more on this later).Retailers have also experimented with gen AI to reinvent the customer experience.Gen AI can deepen relationships with customers(in part,by extending the interactio

18、ns between retailers and customers across the customer journey)and help make the customer experience more personalized and fulfilling.The advanced conversational abilities of gen AI chatbots,powered by natural-language models,can make the smart-shopping assistant a primary shopping channel.Augmentin

19、g retails internal value chainGen AI has the potential to boost productivity and efficiency along each step of the retail value chain,including in marketing,commercialization,distribution,and back-office work(Exhibit 1).Retailers can start to realize gen AIs impact across the value chain through qui

20、ck-win use cases.These use cases generally require fewer resources to implement relative to their impact and compared with other gen AI use cases.In fact,retailers may more easily deploy current off-the-shelf tools without the need for much customization.Real examples of these use cases include the

21、following:Marketing.Amazon launched an AI-powered image generation tool in late 2023 to help advertisers deliver a better ad experience.The tool uses gen AI text prompts to transform basic product photos into more realistic lifestyle images.For example,rather than showing a picture of a sofa against

22、 a white backdrop,AI can place the sofa in an AI-generated living room to help shoppers envision the product in a more relevant context.The tool so far has improved advertising click-through rates by up to 40 percent.Software development.“Copilots,”or genAI tools that help employees do their jobs by

23、 providing a starting point for a task,can boost tech talent productivity by reducing the time spent on software engineering tasks by up to 60percent.Mercado Libre deployed some of these copilots to improve satisfaction and productivity among the companys development teams,empowering them to focus o

24、n higher-value work by automating more repetitive tasks.Gen AI has the potential to boost productivity and efficiency along each step of the retail value chain.3LLM to ROI:How to scale gen AI in retailExhibit 1Web Exhibit of Retail value chainThere are several touchpoints where generative AI can tra

25、nsform the retail value chain.McKinsey&CompanyBefore generative AIAfter generative AIAll supplier negotiations(including end-to-end contract creation)handled manually by associates,often leading to overlooked details Tedious supplier assessments based on limited data,leading to suboptimal choicesThe

26、 initial round of supplier negotiations handled by generative AI(gen AI)chatbotsProcurement associates utilize gen AI to assist in closing deals(eg,gen-AI-powered briefs and automated summaries of supplier terms and key insights)ProcurementEnd-to-end communicationwith third-party logistics handled b

27、y associatesDelayed response to distribution disruptionsdue to complexity of supply chain operationsInitial communication and email messages to third-party logistics handled by gen AI chatbotsReturns management process,along with a response to distribution disruption,supported by gen AIDistributionH

28、undreds of hours spent on the generation of e-commerce contentManual rule-based website personalization,consuming employees resourcesAutomated generation of e-commerce content(eg,product profles,descriptions)within a few minutesE-commerce customer experience personalized spontaneously by automated f

29、ront-end development techniquesE-commerceIndependent decision making by individual functions,leading to a recurring cycle of searching for the underlying causes of commercial events,often failing to identify the true underlying factors End-to-end optimization of decision making across the value chai

30、n through faster,more precise,and personalized insights from structured and unstructured data sources,fueling decision on pricing,promotions,stock allocation,digital marketing,and other levers of performanceEnd-to-endvalue chainOne-size-fts-all marketing approach due to limited customer insights der

31、ived from structured dataCreation of marketing materials through a lengthy,iterative processUnlimited insights extracted from diferent unstructured sources(eg,product reviews)Fully personalized marketing materials generated with increased efciency for every customerMarketingTime-consuming administra

32、tive processes,such as HR and payroll,prone to errors and inefcienciesThe next-generation“white collar”lean transferring administrative processes of support functions to gen-AI-powered chatbots and interfaces(eg,software development copilots,HR/fnancial copilots)Back ofceAnalytical tools of diferent

33、 maturity level,sometimes hard to adoptEnhancement of analytical tools with gen AI interface,automation of creative tasks(merchandiser copilot)CommercialInformation searches(eg,price,in-store location,stock level)handled manually by associates,leading to delayed customer serviceAssociates use gen-AI

34、-powered assistants for instant voice access to information(eg,prices and promos,product location,stock level)In-storeoperations4LLM to ROI:How to scale gen AI in retail In-store operations.In June 2023,Lindex,a Swedish retailer,announced the release of the“Lindex Copilot”to support its store employ

35、ees.The tool,which is trained on the companys sales and store data,provides employees with personalized advice and guidance about store operations and information about daily tasks.While the above examples can help simplify daily tasks,gen AI can also help retailers accelerate their decision making

36、by automatically generating insights,root causes,and domain-level and company-wide responses(Exhibit 2).Retail operations are affected by countless forces that are difficult to quantify and track,making performance analytics and forecasting an arduous task.Traditionally,teams might spend weeks study

37、ing competitors tactics,changes in pricing and promotion,supply chain issues,and unexpected disruptions to understand sales declines and devise strategies to avoid future sales drops.The combination of gen AI and advanced analytics can revolutionize this process:rather than manually assessing that d

38、ata,workers from across the companyfrom CEO to category managercan access a personalized report featuring key performance insights and suggested actions.Lets use a hypothetical electronics retailer as an example.The retailers television sales are 6percent lower than it had forecasted.The retailers t

39、eam Exhibit 2Web Exhibit of Internal and external inputs needed to enable generative AI decision makingGenerative AI can help bring clarity to retail decision making.McKinsey&CompanyInternal dataExternal sourcesAssortment and productConsumer behavior and loyaltyE-commerce and shopsInventory and supp

40、ly chainsPricing and promotionsTransactionsData twinA granular data model of the organization,capturing key internal and external eventsAI coreAnalytical anomaly detectionStatistical engine for noiseelimination and detectionof events requiring attentionMachine learning factortree and reasoningRoot-c

41、ause analysis basedon granular sector-andbusiness-specifc triggersRecommendation engineRequired actions generationbased on industry practice andgenerative AI technologyCompetitors actionsMacroeconomicsMicroevents and disruptionsTrends and innovationsWeather,seasons,holidaysGenerative AIpowered inter

42、facesPersonalized insights1:1 personalized andactionable reportsfor every stakeholderProactive analysisKey data points and cutsdelivered to stakeholdersproactively and automaticallyDiving deeperPrompting additionalanalysis in humanlanguage5LLM to ROI:How to scale gen AI in retailspent a week looking

43、 for the root cause of the decline and came up with a dozen potential reasons:Could the missed sales forecast have been caused by the unusually rainy weather?A delayed product release?Or were temporary out-of-stock items and a weak promotional campaign to blame?In this example,a gen AI system,traine

44、d on the retailers proprietary data,could automatically analyze the impact of not only these potential root causes but also additional scenarios,such as what actions its competitors may have taken at the same time.A cross-functional team,led by the retailers technology leaders and considering input

45、from sales and commercial teams,could work with technology providers to customize the retailers AI-and gen-AI-powered system.The gen AI platform could then create a list of causes by impact,as well as a set of actions the retailer could consider to help reduce sales drops in the future.Based on our

46、early work with retailers,we expect gen-AI-powered decision-making systems to propel up to 5 percent of incremental sales and improve EBIT margins by 0.2 to 0.4percentage points.When it comes to using gen AI copilots,companies will need to decide if they are a“taker”(a user of preexisting tools),a“s

47、haper”(an integrator of available models with proprietary data for more customized results),or a“maker”(a builder of foundation models).Across the internal value chain,most retailers will likely adopt the taker archetype,using publicly available interfaces or APIs with little to no customization to

48、meet their needs.However,many of todays off-the-shelf solutions dont offer the functionality that some retailers need to fully realize the technologys value,since the technology powering these solutions typically doesnt account for sector-and company-specific data.At the same time,most retailers won

49、t be able to adopt the maker archetype,given that the costs associated with building foundation models are outside the typical retailers budget.In these cases,retailers may opt for the shaper archetype,customizing existing LLM tools with their own code and data.The shaper archetype will also be rele

50、vant for gen AI decision-making use cases.How many resources a retailer invests in shaping its gen AI tools will depend on the market it intends to serve,which use cases it wants to prioritize,and how these use cases complement the retailers core value proposition.Reinventing the customer experience

51、Today,retailers typically engage in only three of the seven steps of the customer journey.Gen AI has the potential to increase retailer engagement and reinvent the customer experience across the entire customer journey(Exhibit 3).Gen-AI-powered chatbot assistants are one primary tool retailers can u

52、se to better engage with customers.Customers can use chatbots to receive product recommendations,learn more about a product or retailer,or add or remove items from their virtual shopping carts.Importantly,since many consumers will use these chatbots before deciding to purchase a product rather than

53、after,using chatbots allows retailers to engage with customers earlier in their shopping journey,which can help increase customers overall satisfaction.Gen AI chatbots work by recognizing the intent of a customers message.An LLM agentthe system that the chatbot relies on for its reasoning engineproc

54、esses the customers message and is then connected to various data sets(such as a retailers SKU base)and to other models,such as an analytical personalization engine.To create the best outputs,a retailer must dedicate resources toward product design and conduct frequent user testing to calibrate how

55、it wants the chatbot to process the customers message.(How customers most frequently use the chatbot will largely determine this calibration.)For example,a shopper might be interested in planning a dinner party but may not know what to buy.After the customer provides the gen AI assistant with a few

56、details about the dinner partysuch as how many people are attending,whether any guests have dietary restrictions,and overall budgetthe gen AI assistant could provide specific product recommendations based on the customers preferences or purchase history.6LLM to ROI:How to scale gen AI in retailWhile

57、 chatbots can be a convenient tool to help reduce customers mental load and shopping time,to truly transform the shopping experience and win over customers,chatbots will need to be deeply personalizedfor example,being able to remember customers order histories,product preferences,and shopping habits

58、.Many leading retailers,particularly in the grocery and fashion spaces,have already begun experimenting with chatbots,though most of these early experiments have not yet harnessed the power of personalization(see the sidebar“Retailers embark on the chatbot journey”).As is the case with internal valu

59、e chain gen AI use cases,retailers often adopt the“shaper”archetype for gen AI use cases that transform the customer experience.Determining the costs of chatbots in retail.The first concern many retailers have about integrating gen-AI-powered chatbots into their business is how much it will cost.Tha

60、t depends on a few factors.Product performance metrics(or the length of a conversation between a customer and chatbot)is one of the first considerations.The length of the conversation is inversely related to the quality of personalizationmeaning,the more personalized a chatbot is for a given custome

61、r,the shorter their conversation.Purchase conversions are another factor.The higher the conversion rate,which is linked to the effectiveness of the chatbot,the lower Exhibit 3Web Exhibit of Retail customer journeyGenerative AI can make the customer journey more efcientfor both retailer and shopper.M

62、cKinsey&CompanyBefore generative AIAfter generative AIManually searching across multiple sources,or ideating independentlyCustomer provides brief contextGenerative AI agent understands the intent and ofers suggestions based on preferencesPurchase ideaand contextChoosing based on past experiences or

63、spontaneityhard to controlGenerative AI minimizes switching,as the customer is already shopping in the retailers appSelection ofpreferredretailerAnalyzing online reviews and web forums to gain insights into product usage methodsRecommendations for videos,production usage insights,and step-by-step in

64、structions,tailored to customer contextProductusageTime-consuming browsing through multiple categories and flters online to fnd the right SKUGenerative AI intelligently navigates through catalogs to fnd items that perfectly match the customers purchase history and preferencesSearch forrequired itemM

65、anually comparing each product across multiple dimensions(eg,price,quality,reviews and ratings,size)Adaptation of basket for customer needs,including smart comparisons of products,considering factors from unstructured data(eg,style,customer and market reviews)Selectionof item andquantityNavigating t

66、hrough multiple checkout pages,entering delivery and payment information,and choosing from limited delivery slots that may not align with the customers scheduleConfrmation of order and specifying delivery preferences in a conversational style(eg,“Please deliver this order to my mothers address on an

67、y day except Wednesday”)Orderconfrmation,delivery dateselectionManually creating the purchase list,relying on customer memory or experienceGenerative AI agent provides an accurate shopping list that includes the items/ingredients needed for the purchase ideaCreation ofpurchase list7LLM to ROI:How to

68、 scale gen AI in retailthe net operational costs of that chatbot.A third factor is the price of LLM APIs.The cost of using these LLM APIs has dropped dramatically in the past year(for example,when comparing the cost of input tokens,GPT-4o,released in May 2024,is half as expensive to operate as GPT-4

69、 Turbo,released a year earlier).AI experts believe that the price of LLM APIs will continue to drop substantially,with some estimates showing a drop of as much as 80 percent within the next two to three years.Based on our experience building gen AI chatbots with retail companies across a range of re

70、alistic scenarios,a 2 to 4 percent basket uplift can justify LLM costs.Retailers can also combine the power of their generative and analytical AI products to While some retailers have adopted a wait-and-see approach to generative AI(gen AI),others have already started experimenting with the technolo

71、gy(exhibit).Leading retailers,particularly those in grocery and fashion,started experimenting with gen-AI-powered chatbots in late 2023.These chatbots have taken different forms.Walmart launched its“Text to Shop”tool,where customers can text the retailer to search for items,add or remove items from

72、their carts,reorder products,and schedule Retailers embark on the chatbot journeydeliveries.Instacart created a ChatGPT plug-in that allows users to plan a meal in ChatGPT and then convert the output into a basket on Instacarts website.Exhibit 1Eg,iMessage,Instagram direct message,WhatsApp.2SmartSea

73、rch functionality in the app;chatbot still available only in iMessage.Web Sidebar of Retailer implementation of AI-powered chatbots,nonexhaustiveRetailers,particularly those in grocery and fashion,have started experimenting with generative AIpowered conversational chatbots.McKinsey&CompanyImplementa

74、tionDec 2022Mar 2023Apr 2023Apr 2023June 2023Nov 2023Jan 2024Feb 2024Feb 2024Feb 2024Mar 2024Reported plans123456789101112RetailerWalmartInstacartKNXTZalandoCarrefourCastoramaWalmart2AmazonIKEASsenseLykoVictorias SecretRetailer categoryGroceryGroceryFashionFashionGroceryDIYGroceryGroceryDIYFashionBe

75、autyFashionAI-powered chatbot typeMessenger1ChatGPT plug-inWebsite/appWebsite/appWebsite/appWebsite/appWebsite/appWebsite/appChatGPT plug-inChatGPT plug-inWebsite/appWebsite/appDecJanFebMarAprMayJuneJulyAugSeptOctNovDecJan20232024FebMar1235671112489108LLM to ROI:How to scale gen AI in retailfurther

76、justify LLM costs.For example,companies can first use gen AI to learn more about a customer,then use analytical models to surface personal offers relevant to that customer.Together,these two technologies can help increase sales conversions.When building a business case,retailers should also consider

77、 the investment required to develop a chatbot.Sometimes,the basket uplift may not be high enough to cover the cost of the investment.To understand the full return on their investment,retailers should factor in the cost of attracting new customers who will use these tools,as well as how much the tool

78、 can increase the purchase frequency for existing customers.Measuring chatbots impact.In controlled customer experiments,weve seen chatbots create a significant increase in convenience for customers.When comparing a traditional retailer app with the minimum viable product of a gen-AI-enabled chatbot

79、,the chatbot reduced the time spent to complete an order by 50 to 70 percent.Retailers that arent ready to invest in chatbots may instead choose to launch smart-search functionality.Smart-search tools allow a customer to receive a list of recommended products by asking a question rather than needing

80、 to engage in a conversation with a chatbot.(For example,a customer might search for“dinner party supplies,”and the smart-search tool would provide a list of products that one might need for a dinner party).While traditional search uses basic algorithms and relies on keyword matching,smart-search to

81、ols powered by gen AI can better understand the context and intent of a search term,even if it veers away from keyword use.Although these smart-search tools may be limited in functionality compared with using chatbotsand therefore limited in impactthey are easier and less expensive to develop.They a

82、lso carry fewer risks compared with using a chatbot;their outputs are generally limited to a list of products rather than longer text that a chatbot would give,which means the responses are less likely to be harmful,offensive,or inaccurate.How to scale the use of gen AI in retailGen AI is no longer

83、a novelty.As companies figure out how to implement the technology to create real value,best-in-class retailers will need to move from testing to scaling or else risk falling behind their competitorsor,worse,losing customers.To scale their gen AI tools,retail executives can consider five imperatives

84、for outcompeting in digital and AI:Identify domain-level transformation candidates.Retail executives should identify the various domains where a transformation is needed,such as in customer experience,marketing,or store employee productivity,before they identify which gen AI use cases to pursue.By i

85、dentifying transformations at the domain level first,retailers can determine which tools will bolster gen AIs impact,such as robotic process automation(RPA)or advanced analytics.Upskill talent to develop gen AI skills.Retail people leaders should offer both technical and nontechnical talent the oppo

86、rtunity to engage in learning programs,such as those focused on gen AI software development and prompt engineering.Form a centralized,cross-functional team to enable scaling.While scaling gen AI will hinge on a retailers tech capabilities,retailers can gather leaders from across the organization to

87、identify how gen AI can help improve the business.These cross-functional teams,convened to help accelerate scaling in the short term,should have shared goals that extend across the business and reinforce the retailers overall strategy.Set up technology architecture to scale.Before committing to a sp

88、ecific gen AI vendor,retailers should experiment with different ones to assess which vendor will best suit their needs.The ideal gen AI architecture for retailers will be agile enough to make switching between LLMs easier to do,thus making scaling the technology across the organization easier as wel

89、l.(This means using modular components that can be easily swapped out.)9LLM to ROI:How to scale gen AI in retail Ensure data quality to fuel models.Unstructured data will be critical to powering retailers gen AI tools and providing key customer insights.Retailers should identify the unstructured dat

90、a sources that differentiate them from other retailers(grocers,for example,could develop new recipe databases or leverage existing ones)and establish metadata tagging standards so tech teams can more efficiently power a retailers gen AI models.Decisions on data should be backed by a clear understand

91、ing of the datas business application.Some of the guidance outlined above may be sector-agnostic,but scaling gen AI in retail is unique because several of the technologys use cases involve direct interactions with consumers.In retail,even a 1 percent margin of error could result in millions of custo

92、mer-facing mistakes.This emphasizes the importance of strong gen AI risk guidelines and safety testing.The stakes may be higher,but the rewards are,too.Designed by McKinsey Global Publishing Copyright 2024 McKinsey&Company.All rights reserved.Alexander Sukharevsky is a senior partner in McKinseys Lo

93、ndon office,where Sergey Kondratyuk is an associate partner;Andreas Ess is a partner in the Zurich office;Denis Emelyantsev is a partner in the Atlanta office;Emily Reasor is a senior partner in the Denver office;Holger Hrtgen is a partner in the Dsseldorf office;and Oleg Sokolov is an associate partner in the Stockholm office.The authors wish to thank Andrei Persh and Sergei Sereda for their contributions to this article.Scan Download PersonalizeFind more content like this on the McKinsey Insights App10LLM to ROI:How to scale gen AI in retail