生成式人工智能重置:2024年重新布线将潜力转化为价值(英)-2024.3_第1页
生成式人工智能重置:2024年重新布线将潜力转化为价值(英)-2024.3_第2页
生成式人工智能重置:2024年重新布线将潜力转化为价值(英)-2024.3_第3页
生成式人工智能重置:2024年重新布线将潜力转化为价值(英)-2024.3_第4页
生成式人工智能重置:2024年重新布线将潜力转化为价值(英)-2024.3_第5页
已阅读5页,还剩10页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

March2024

Mcsey

Quartery

AgenerativeAIreset:

Rewiringtoturnpotentialintovaluein2024

ThegenerativeAIpayoffmayonlycomewhencompaniesdodeeperorganizationalsurgeryontheirbusiness.

byEricLamarre,AlexSingla,AlexanderSukharevsky,andRodneyZemmel

It’stimeforagenerativeAI(genAI)reset.Theinitialenthusiasmandflurryofactivityin2023isgivingwaytosecondthoughtsandrecalibrationsascompaniesrealizethatcapturinggenAI’senormouspotentialvalueisharderthanexpected.

With2024shapinguptobetheyearforgenAItoproveitsvalue,companiesshould

keepinmindthehardlessonslearnedwithdigitalandAItransformations:competitiveadvantagecomesfrombuildingorganizationalandtechnologicalcapabilitiestobroadlyinnovate,deploy,andimprovesolutionsatscale—ineffect,rewiringthebusinessfor

distributeddigitalandAIinnovation.

CompanieslookingtoscoreearlywinswithgenAIshouldmovequickly.ButthosehopingthatgenAIoffersashortcutpastthetough—andnecessary—organizationalsurgery

arelikelytomeetwithdisappointingresults.Launchingpilotsis(relatively)easy;gettingpilotstoscaleandcreatemeaningfulvalueishardbecausetheyrequireabroadsetofchangestothewayworkactuallygetsdone.

Let’sbrieflylookatwhatthishasmeantforonePacificregiontelecommunications

company.ThecompanyhiredachiefdataandAIofficerwithamandateto“enablethe

organizationtocreatevaluewithdataandAI.”ThechiefdataandAIofficerworkedwith

thebusinesstodevelopthestrategicvisionandimplementtheroadmapfortheusecases.Afterascanofdomains(thatis,customerjourneysorfunctions)andusecaseopportunitiesacrosstheenterprise,leadershipprioritizedthehome-servicing/maintenancedomainto

pilotandthenscaleaspartofalargersequencingofinitiatives.Theytargeted,inparticular,thedevelopmentofagenAItooltohelpdispatchersandserviceoperatorsbetterpredict

thetypesofcallsandpartsneededwhenservicinghomes.

2

Leadershipputinplacecross-functionalproductteamswithsharedobjectivesand

incentivestobuildthegenAItool.Aspartofanefforttoupskilltheentireenterpriseto

betterworkwithdataandgenAItools,theyalsosetupadataandAIacademy,which

thedispatchersandserviceoperatorsenrolledinaspartoftheirtraining.Toprovide

thetechnologyanddataunderpinningsforgenAI,thechiefdataandAIofficeralso

selectedalargelanguagemodel(LLM)andcloudproviderthatcouldmeettheneedsofthedomainaswellasserveotherpartsoftheenterprise.ThechiefdataandAIofficer

alsooversawtheimplementationofadataarchitecturesothatthecleanandreliable

data(includingservicehistoriesandinventorydatabases)neededtobuildthegenAItoolcouldbedeliveredquicklyandresponsibly.

OurbookRewired:TheMcKinseyGuidetoOutcompetingintheAgeofDigitalandAI(Wiley,June2023)providesadetailedmanualonthesixcapabilitiesneededtodeliverthekindof

broadchangethatharnessesdigitalandAItechnology.Inthisarticle,wewillexplorehowtoextendeachofthosecapabilitiestoimplementasuccessfulgenAIprogramatscale.Whilerecognizingthatthesearestillearlydaysandthatthereismuchmoretolearn,ourexperiencehasshownthatbreakingopenthegenAIopportunityrequirescompaniestorewirehowtheyworkinthefollowingways.

FigureoutwheregenAIcopilotscangiveyouarealcompetitiveadvantage

ThebroadexcitementaroundgenAIanditsrelativeeaseofusehasledtoaburstof

experimentationacrossorganizations.Mostoftheseinitiatives,however,won’tgenerateacompetitiveadvantage.Onebank,forexample,boughttensofthousandsofGitHub

Copilotlicenses,butsinceitdidn’thaveaclearsenseofhowtoworkwiththetechnology,progresswasslow.Anotherunfocusedeffortweoftenseeiswhencompaniesmove

toincorporategenAIintotheircustomerservicecapabilities.Customerserviceisa

commoditycapability,notpartofthecorebusiness,formostcompanies.WhilegenAImighthelpwithproductivityinsuchcases,itwon’tcreateacompetitiveadvantage.

Tocreatecompetitiveadvantage,companiesshouldfirstunderstandthedifference

betweenbeinga“taker”(auserofavailabletools,oftenviaAPIsandsubscriptionservices),a“shaper”(anintegratorofavailablemodelswithproprietarydata),anda“maker”(abuilderofLLMs).Fornow,themakerapproachistooexpensiveformostcompanies,sothesweetspotforbusinessesisimplementingatakermodelforproductivityimprovementswhile

buildingshaperapplicationsforcompetitiveadvantage.

MuchofgenAI’snear-termvalueiscloselytiedtoitsabilitytohelppeopledotheir

currentjobsbetter.Inthisway,genAItoolsactascopilotsthatworksidebysidewithanemployee,creatinganinitialblockofcodethatadevelopercanadapt,forexample,ordraftingarequisitionorderforanewpartthatamaintenanceworkerinthefield

canreviewandsubmit(seesidebar“CopilotexamplesacrossthreegenerativeAI

archetypes”).Thismeanscompaniesshouldbefocusingonwherecopilottechnologycanhavethebiggestimpactontheirpriorityprograms.

3

Copilotexamplesacrossthree

generativeAI

archetypes

•“Taker”copilotshelp

realestatecustomers

siftthroughproperty

optionsandfindthemostpromisingone,write

codeforadeveloper,

andsummarizeinvestor

transcripts.

•“Shaper”copilotsprovide

recommendationstosales

repsforupsellingcustomersbyconnectinggenerativeAItoolstocustomerrelationshipmanagementsystems,

financialsystems,and

customerbehaviorhistories;createvirtualassistantsto

personalizetreatmentsforpatients;andrecommendsolutionsformaintenanceworkersbasedonhistoricaldata.

•“Maker”copilotsarefoundationmodels

thatlabscientistsat

pharmaceuticalcompaniescanusetofindandtest

newandbetterdrugs

morequickly.

Someindustrialcompanies,forexample,haveidentifiedmaintenanceasacriticaldomainfortheirbusiness.

Reviewingmaintenancereportsandspendingtimewithworkersonthefrontlinescanhelpdeterminewhere

agenAIcopilotcouldmakeabigdifference,suchas

inidentifyingissueswithequipmentfailuresquickly

andearlyon.AgenAIcopilotcanalsohelpidentify

rootcausesoftruckbreakdownsandrecommend

resolutionsmuchmorequicklythanusual,aswellas

actasanongoingsourceforbestpracticesorstandardoperatingprocedures.

Thechallengewithcopilotsisfiguringouthowto

generaterevenuefromincreasedproductivity.In

thecaseofcustomerservicecenters,forexample,companiescanstoprecruitingnewagentsanduseattritiontopotentiallyachieverealfinancialgains.

Definingtheplansforhowtogeneraterevenuefromtheincreasedproductivityupfront,therefore,iscrucialto

capturingthevalue.

Upskillthetalentyouhave

butbeclearaboutthegen-AI-specificskillsyouneed

Bynow,mostcompanieshaveadecentunderstandingofthetechnicalgenAIskillstheyneed,suchasmodelfine-tuning,vectordatabaseadministration,prompt

engineering,andcontextengineering.Inmany

cases,theseareskillsthatyoucantrainyourexistingworkforcetodevelop.ThosewithexistingAIand

machinelearning(ML)capabilitieshaveastronghead

start.Dataengineers,forexample,canlearnmultimodalprocessingandvectordatabasemanagement,MLOps(MLoperations)engineerscanextendtheirskillsto

LLMOps(LLMoperations),anddatascientistscan

developpromptengineering,biasdetection,andfine-tuningskills.

Thelearningprocesscantaketwotothreemonthsto

gettoadecentlevelofcompetencebecauseofthe

complexitiesinlearningwhatvariousLLMscanandcan’tdoandhowbesttousethem.Thecodersneedtogain

experiencebuildingsoftware,testing,andvalidating

4

answers,forexample.Ittookonefinancial-servicescompanythreemonthstotrainitsbestdatascientiststoahighlevelofcompetence.Whilecoursesanddocumentation

areavailable—manyLLMprovidershavebootcampsfordevelopers—wehavefound

thatthemosteffectivewaytobuildcapabilitiesatscaleisthroughapprenticeship,

trainingpeopletothentrainothers,andbuildingcommunitiesofpractitioners.Rotatingexpertsthroughteamstotrainothers,schedulingregularsessionsforpeopletoshare

learnings,andhostingbiweeklydocumentationreviewsessionsarepracticesthathaveprovensuccessfulinbuildingcommunitiesofpractitioners(seesidebar“AsampleofnewgenerativeAIskillsneeded”).

It’simportanttobearinmindthatsuccessfulgenAIskillsareaboutmorethancoding

proficiency.OurexperienceindevelopingourowngenAIplatform,Lilli,showedusthat

thebestgenAItechnicaltalenthasdesignskillstouncoverwheretofocussolutions,

contextualunderstandingtoensurethemostrelevantandhigh-qualityanswersare

generated,collaborationskillstoworkwellwithknowledgeexperts(totestandvalidate

answersanddevelopanappropriatecurationapproach),strongforensicskillstofigure

outcausesofbreakdowns(istheissuethedata,theinterpretationoftheuser’sintent,the

qualityofmetadataonembeddings,orsomethingelse?),andanticipationskillstoconceiveofandplanforpossibleoutcomesandtoputtherightkindoftrackingintotheircode.A

purecoderwhodoesn’tintrinsicallyhavetheseskillsmaynotbeasusefulateammember.

Whilecurrentupskillingislargelybasedona“learnonthejob”approach,weseearapid

marketemergingforpeoplewhohavelearnedtheseskillsoverthepastyear.Thatskill

growthismovingquickly.GitHubreportedthatdeveloperswereworkingongenAIprojects“inbignumbers,”andthat65,000publicgenAIprojectswerecreatedonitsplatformin

2023—ajumpofalmost250percentoverthepreviousyear.IfyourcompanyisjuststartingitsgenAIjourney,youcouldconsiderhiringtwoorthreeseniorengineerswhohavebuiltagenAIshaperproductfortheircompanies.Thiscouldgreatlyaccelerateyourefforts.

Formacentralizedteamtoestablishstandardsthatenableresponsiblescaling

ToensurethatallpartsofthebusinesscanscalegenAIcapabilities,centralizing

competenciesisanaturalfirstmove.Thecriticalfocusforthiscentralteamwillbeto

developandputinplaceprotocolsandstandardstosupportscale,ensuringthatteamscanaccessmodelswhilealsominimizingriskandcontainingcosts.Theteam’swork

couldinclude,forexample,procuringmodelsandprescribingwaystoaccessthem,developingstandardsfordatareadiness,settingupapprovedpromptlibraries,andallocatingresources.

WhiledevelopingLilli,ourteamhaditsmindonscalewhenitcreatedanopenplug-inarchitectureandsettingstandardsforhowAPIsshouldfunctionandbebuilt.They

developedstandardizedtoolingandinfrastructurewhereteamscouldsecurely

experimentandaccessaGPTLLM,agatewaywithpreapprovedAPIsthatteamscouldaccess,andaself-servedeveloperportal.Ourgoalisthatthisapproach,overtime,can

5

AsampleofnewgenerativeAI

skillsneeded

Thefollowingareexamplesofnewskillsneededforthe

successfuldeploymentof

generativeAItools:

•datascientist:

–promptengineering

–in-contextlearning

–biasdetection

–patternidentification

–reinforcementlearningfromhumanfeedback

–hyperparameter/largelanguagemodelfine-

tuning;transferlearning

•dataengineer:

–datawranglinganddatawarehousing

–datapipelineconstruction

–multimodalprocessing

–vectordatabasemanagement

helpshift“Lilliasaproduct”(thatahandfulofteamsusetobuildspecificsolutions)to“Lilliasaplatform”(thatteamsacrosstheenterprisecanaccesstobuildotherproducts).

ForteamsdevelopinggenAIsolutions,squad

compositionwillbesimilartoAIteamsbutwithdata

engineersanddatascientistswithgenAIexperienceandmorecontributorsfromriskmanagement,compliance,

andlegalfunctions.Thegeneralideaofstaffingsquadswithresourcesthatarefederatedfromthedifferent

expertiseareaswillnotchange,buttheskillcompositionofagen-AI-intensivesquadwill.

Setupthetechnologyarchitecturetoscale

BuildingagenAImodelisoftenrelativelystraightforward,butmakingitfullyoperationalatscaleisadifferentmatterentirely.We’veseenengineersbuildabasicchatbotin

aweek,butreleasingastable,accurate,andcompliantversionthatscalescantakefourmonths.That’swhy,ourexperienceshows,theactualmodelcostsmaybeless

than10to15percentofthetotalcostsofthesolution.

Buildingforscaledoesn’tmeanbuildinganewtechnologyarchitecture.Butitdoesmeanfocusingonafewcore

decisionsthatsimplifyandspeedupprocesseswithoutbreakingthebank.Threesuchdecisionsstandout:

•Focusonreusingyourtechnology.Reusingcode

canincreasethedevelopmentspeedofgenAIuse

casesby30to50percent.Onegoodapproachis

simplycreatingasourceforapprovedtools,code,

andcomponents.Afinancial-servicescompany,for

example,createdalibraryofproduction-gradetools,

whichhadbeenapprovedbyboththesecurityandlegalteams,andmadethemavailableinalibraryforteams

touse.Moreimportantistakingthetimetoidentifyandbuildthosecapabilitiesthatarecommonacrossthe

mostpriorityusecases.Thesamefinancial-services

company,forexample,identifiedthreecomponentsthatcouldbereusedformorethan100identifiedusecases.Bybuildingthosefirst,theywereabletogeneratea

significantportionofthecodebaseforalltheidentifiedusecases—essentiallygivingeveryapplicationabig

headstart.

6

•FocusthearchitectureonenablingefficientconnectionsbetweengenAImodels

andinternalsystems.ForgenAImodelstoworkeffectivelyintheshaperarchetype,theyneedaccesstoabusiness’sdataandapplications.Advancesinintegrationandorchestrationframeworkshavesignificantlyreducedtheeffortrequiredtomake

thoseconnections.Butlayingoutwhatthoseintegrationsareandhowtoenable

themiscriticaltoensurethesemodelsworkefficientlyandtoavoidthecomplexity

thatcreatestechnicaldebt(the“tax”acompanypaysintermsoftimeandresourcesneededtoredressexistingtechnologyissues).Chiefinformationofficersandchief

technologyofficerscandefinereferencearchitecturesandintegrationstandardsfortheirorganizations.Keyelementsshouldincludeamodelhub,whichcontainstrainedandapprovedmodelsthatcanbeprovisionedondemand;standardAPIsthatactas

bridgesconnectinggenAImodelstoapplicationsordata;andcontextmanagement

andcaching,whichspeedupprocessingbyprovidingmodelswithrelevantinformationfromenterprisedatasources.

•Buildupyourtestingandqualityassurancecapabilities.OurownexperiencebuildingLillitaughtustoprioritizetestingoverdevelopment.Ourteaminvestedinnotonly

developingtestingprotocolsforeachstageofdevelopmentbutalsoaligningtheentire

teamsothat,forexample,itwasclearwhospecificallyneededtosignoffoneachstageoftheprocess.Thissloweddowninitialdevelopmentbutspeduptheoveralldelivery

paceandqualitybycuttingbackonerrorsandthetimeneededtofixmistakes.

Ensuredataqualityandfocusonunstructureddatatofuelyourmodels

TheabilityofabusinesstogenerateandscalevaluefromgenAImodelswilldependonhowwellittakesadvantageofitsowndata.Aswithtechnology,targetedupgradesto

existingdataarchitectureareneededtomaximizethefuturestrategicbenefitsofgenAI:

•Betargetedinrampingupyourdataqualityanddataaugmentationefforts.While

dataqualityhasalwaysbeenanimportantissue,thescaleandscopeofdatathatgen

AImodelscanuse—especiallyunstructureddata—hasmadethisissuemuchmore

consequential.Forthisreason,it’scriticaltogetthedatafoundationsright,from

clarifyingdecisionrightstodefiningcleardataprocessestoestablishingtaxonomiessomodelscanaccessthedatatheyneed.Thecompaniesthatdothiswelltietheir

dataqualityandaugmentationeffortstothespecificAI/genAIapplicationanduse

case—youdon’tneedthisdatafoundationtoextendtoeverycorneroftheenterprise.

Thiscouldmean,forexample,developinganewdatarepositoryforallequipment

specificationsandreportedissuestobettersupportmaintenancecopilotapplications.

•Understandwhatvalueislockedintoyourunstructureddata.Mostorganizationshave

traditionallyfocusedtheirdataeffortsonstructureddata(valuesthatcanbeorganizedintables,suchaspricesandfeatures).ButtherealvaluefromLLMscomesfromtheirabilitytoworkwithunstructureddata(forexample,PowerPointslides,videos,and

text).Companiescanmapoutwhichunstructureddatasourcesaremostvaluableandestablishmetadatataggingstandardssomodelscanprocessthedataandteamscan

7

findwhattheyneed(taggingisparticularlyimportanttohelpcompaniesremovedatafrommodelsaswell,ifnecessary).Becreativeinthinkingaboutdataopportunities.Somecompanies,forexample,areinterviewingsenioremployeesastheyretire

andfeedingthatcapturedinstitutionalknowledgeintoanLLMtohelpimprovetheircopilotperformance.

•Optimizetolowercostsatscale.Thereisoftenasmuchasatenfolddifference

betweenwhatcompaniespayfordataandwhattheycouldbepayingiftheyoptimized

theirdatainfrastructureandunderlyingcosts.Thisissueoftenstemsfromcompanies

scalingtheirproofsofconceptwithoutoptimizingtheirdataapproach.Twocosts

generallystandout.Oneisstoragecostsarisingfromcompaniesuploadingterabytes

ofdataintothecloudandwantingthatdataavailable24/7.Inpractice,companies

rarelyneedmorethan10percentoftheirdatatohavethatlevelofavailability,and

accessingtherestovera24-or48-hourperiodisamuchcheaperoption.Theother

costsrelatetocomputationwithmodelsthatrequireon-callaccesstothousandsof

processorstorun.Thisisespeciallythecasewhencompaniesarebuildingtheirown

models(themakerarchetype)butalsowhentheyareusingpretrainedmodelsand

runningthemwiththeirowndataandusecases(theshaperarchetype).Companies

couldtakeacloselookathowtheycanoptimizecomputationcostsoncloudplatforms—

forinstance,puttingsomemodelsinaqueuetorunwhenprocessorsaren’tbeingused(suchaswhenAmericansgotobedandconsumptionofcomputingserviceslikeNetflixdecreases)isamuchcheaperoption.

Buildtrustandreusabilitytodriveadoptionandscale

BecausemanypeoplehaveconcernsaboutgenAI,thebaronexplaininghowthesetoolsworkismuchhigherthanformostsolutions.Peoplewhousethetoolswanttoknowhowtheywork,notjustwhattheydo.Soit’simportanttoinvestextratimeandmoneytobuildtrustbyensuringmodelaccuracyandmakingiteasytocheckanswers.

Oneinsurancecompany,forexample,createdagenAItooltohelpmanageclaims.As

partofthetool,itlistedalltheguardrailsthathadbeenputinplace,andforeachanswerprovidedalinktothesentenceorpageoftherelevantpolicydocuments.ThecompanyalsousedanLLMtogeneratemanyvariationsofthesamequestiontoensureanswer

consistency.Thesesteps,amongothers,werecriticaltohelpingendusersbuildtrustinthetool.

PartofthetrainingformaintenanceteamsusingagenAItoolshouldbetohelpthem

understandthelimitationsofmodelsandhowbesttogettherightanswers.Thatincludes

teachingworkersstrategiestogettothebestanswerasfastaspossiblebystartingwith

broadquestionsthennarrowingthemdown.Thisprovidesthemodelwithmorecontext,

anditalsohelpsremoveanybiasofthepeoplewhomightthinktheyknowtheanswer

already.Havingmodelinterfacesthatlookandfeelthesameasexistingtoolsalsohelps

usersfeellesspressuredtolearnsomethingneweachtimeanewapplicationisintroduced.

Gettingtoscalemeansthatbusinesseswillneedtostopbuildingone-o

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论