bcg -企業(yè)如何管理AI新同事-The Emerging Agentic Enterprise How Leaders Must Navigate a New Age of AI

Incollaborationwith

〉

BIGIDEAS

RESEARCHREPORT

MITsloan

ManagementReview

November2025

TheEmerging

AgenticEnterprise:HowLeaders

MustNavigateaNewAgeofAI

bySamRansbotham,DavidKiron,ShervinKhodabandeh,SeshIyer,andAmartyaDas

AUTHORS

SamRansbotham

isaprofessorofanalyticsattheCarrollSchoolofManagementatBostonCollege,aswellasguesteditorforMITSloanManagementReviewsArtificialIntelligenceandBusinessStrategyBigIdeasinitiative.

DavidKironistheeditorialdirector,research,ofMITSloanManagementReviewandprogramleadforitsBigIdeasresearchinitiatives.

ShervinKhodabandehisamanagingdirectorandseniorpartneratBostonConsultingGroup(BCG)andthecoleaderofitsAIbusiness

inNorthAmerica.HeisaleaderinBCGX

andhasover20yearsofexperiencedriving

businessimpactfromAIanddigital.Hecanbecontactedat

shervin@

.

SeshIyerisamanagingdirectorandsenior

partneratBCGandtheNorthAmericachairforBCGX,wherehehelpsclientsdrive

large-scaleAItransformations.Heisthe

globalleaderfortheAI&TechLabattheBCGHendersonInstituteandcanbereachedat

sesh@

.

AmartyaDasisaprincipalatBCGand

currentlyservesasanambassadorattheBCGHendersonInstitute,whereheleadsresearchontheimpactoftechnologyandAIonsociety,focusingonhowemergingtechnologiesreshapebothcompaniesandpublicinstitutions.Hecanbereachedat

das.amartya@

.

CONTRIBUTORS

ToddFitz,KevinFoley,SarahJohnson,MattLangione,MicheleLeeDeFilippo,AmandaLuther,JenniferMartin,SamanthaOldroyd,MeenalPore,AllisonRyder,TaylorZhang,LeilaZhu,

LeonidZhukov,DavidZuluagaMartínez

TheresearchandanalysisforthisreportwasconductedunderthedirectionoftheauthorsaspartofanMITSloanManagementReviewresearchinitiativeincollaborationwithandsponsoredby

BostonConsultingGroup.

Tocitethisreport,pleaseuse:

S.Ransbotham,D.Kiron,S.Khodabandeh,S.Iyer,andA.Das,“TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI,”MITSloanManagementReviewandBoston

ConsultingGroup,November2025.

SUPPORTINGSPONSORS

BCG

HENDERSONINSTITUTE

BCG》《

/10.63383/jAXE2583

Copyright?MassachusettsInstituteofTechnology,2025.Allrightsreserved.REPRINT#:67270

CONTENTS

1

Introduction

4

StrategicTensionsWhenAdoptingAgenticAI

1

1

AStrategicOverhaulofWorkflows,Governance,

Roles,andInvestment

17

Conclusion

18

Appendix

Introduction

Executiveshavelongreliedonsimplecategoriestoframehowtechnologyfitsintoorganizations:Toolsautomatetasks,peoplemakedecisions,andstrategydetermineshowthetwoworktogether.Thatframingisnolongersufficient.Anewclassofsys-tems—agenticAI—complicatestheseboundaries.Thesesystemscanplan,act,andlearnontheirown.Theyarenotjusttoolstobeoperatedorassistantswaitingforinstructions.Increasingly,theybehavelikeautonomousteammates,capableofexe-cutingmultistepprocessesandadaptingastheygo.Notably,76%ofrespondentstoourglobalexecutivesurveysaytheyviewagenticAIasmorelikeacoworkerthanatool.

Forstrategists,agenticAI’sdualnatureasbothatoolandcoworkercreatesnewdilemmas.Asingleagentmighttakeoveraroutinestep,supportahumanexpertwithanalysis,andcollaborateacrossworkflowsinwaysthatshiftdecision-makingauthority.Thistool-coworkerdualitybreaksdowntraditionalmanagementlogic,whichassumesthattechnologyeithersubstitutesorcomplements,auto-matesoraugments,islabororcapital,orisatooloraworker,butnotallatonce.Organizationsnowfaceanunprecedentedchallenge:managingasinglesystemthatdemandsbothhumanresourceapproachesandassetmanagementtechniques.

Theseparationoftechnologyandstrategyinsidemostorganizationsexacerbatesthischallenge.Technologyexecutivesfocusontechnologyissues,makingpilot,vendor,orinfrastructuredecisions.Strategicexecutivesfocusonmarkets,competition,andpeople.ButagenticAImakesthatseparationuntenable.Itsimultaneouslyinfluencesthedesignofprocesses,thestructureofroles,theallocationofdecisionrights,andthecultureofaccountability.

ABOUTTHERESEARCH

ThisreportpresentsfindingsfromtheninthannualglobalresearchstudyonartificialintelligenceandbusinessstrategybyMITSloanManagementReviewandBostonConsultingGroup.Inspring2025,wefieldedaglobalsurveyandsubsequentlyanalyzedrecordsfrom2,102respondentsrepresentingmorethan21industriesand116countries.Wealsointerviewed11executivesleadingAIinitiativesinabroadrangeofcompaniesandindustries,includingfinancialservices,technology,retail,energy,andhealthcare.

OurresearchexaminesthespeedofagenticAIadoptioninsideorganizations,whichhasoutpacedtheadoptionoftraditionalandgenerativeAI.ExploringhowagenticAIrelatestocapitalandlabor,thisreportoutlinesaseriesofresultanttensionsandofferssuggestionsfororganizationsseekingtoresolvethem.

1

TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI

2

MITSLOANMANAGEMENTREVIEW?BCG

TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI3

ATidalWaveofAdoption,aTrickleofStrategy

Despitethetechnology’swide-rangingimplications,organizationsarerapidlyadoptingagenticAI,wellbeforetheyhaveastrategyinplace.TheiradoptionoftraditionalAIhasclimbedto72%overthepasteightyears,accord-ingtooursurvey.(seefigure1.)GenerativeAIachieved70%adoptioninjustthreeyears.Injusttwoyears,agenticAIhasalreadyreached35%adoption,withanother44%oforganizationsplanningtodeployitsoon.Vendorsareacceleratingthistrendbyembeddingagenticcapabilitiesasfeaturesintheirofferings,causingorganizationstoimple-mentagenticAIbeforetheyhavedevelopedastrategicmanagementframework.

Theresultisagrowingstrategicrisk:AgenticAIisspread-ingacrossenterprisesfasterthanleaderscanredesignprocesses,assigndecisionrights,orrethinkworkforce

models.Ourresearch,basedon2,000-plusrespondentstoaglobalsurveyandinterviewswithleadingexecutives,findsthatorganizationshavemultipleoptionsforobtain-ingvaluefromagenticAI.Itofferspossibilitiesnotonlytoimprovecostefficiencybutalsotoexpandrevenue,accelerateinnovation,compresslearningcurves,andrestructureorganizations.Withoutastrategicapproachthatalignstheseobjectives,organizationsrisklimitingreturnsontheirinvestments.

TheExecutiveDilemma

AgenticAI’sdualnatureasbothatoolandcoworkercreatescompetingorganizationalpressuresthatman-agementframeworkscannotresolve.ITleaderswantpredictable,scalablesystemswithcleartechnicalspecifi-cations.CFOsneedinvestmentmodelswithmeasurablereturnsanddepreciationschedules.HRexecutivesrequire

PercentageoforganizationsusingAI

%

100

80

52%50%

40

20

0

201720182019202020212022202320242025

Source:MITSloanManagementReview-BCGglobalexecutivesurveys,2017-2025.

53%44%

70%72%

57%56%

46%

60

FIGURE1

TraditionalAIAdoptionContinuestoGrow

Since2023,thepercentageoforganizationspilotingordeployingAIsolutionshasrisen22percentagepoints.

performancemanagementframeworksandsupervisionprotocols.Businessleadersdemandbothefficiencyandadaptabilityfromthesamesystem.

Thesecompetingdemandsaren’timplementationchal-lenges.They’restrategicimperativesthatexposenewsourcesoforganizationaldifferentiation.Amongorganiza-tionswithextensiveagenticAIuse,wefindthat73%believeusingAIfundamentallyincreasestheirabilitytostandout,while76%oftheiremployeesbelieveitchangeshowindi-vidualsdifferentiatethemselvesfromcoworkers.

OrganizationsandindividualscanuseagenticAItodif-ferentiatethemselvesbecausetheycannowtakeadvan-tageofthewaysthatagenticAIdoesnotfittraditionalmanagementframeworks.Thesamesystemthatofferscostreductionsthroughtool-likeautomationalsoenablesrevenueexpansionthroughworkerlikeadaptability,innovationaccelerationthroughcontinuouslearning,andorganizationallearningthroughhuman-agentinter-actions.Butaccessingthesemultiplesourcesofvaluerequiresnavigatingoperationaltensionsthatnoexistingframeworkaddresses.

Thepathforwardrequiresunderstandingfouropera-tionaltensionsthatexposetheinadequacyoftraditionalmanagementapproachesandthenredesigningfunda-mentalprocesses—workdesign,governance,work-forceplanning,learning,andinvestment—toworkwithratherthanagainstagenticAI’sinherentduality.Thisreportoffersevidence-basedrecommendationsonhowtoproceed.

StrategicTensionsWhenAdoptingAgenticAI

Thecompetingpressuresexecutivesfacearen’tabstracttheoreticalproblems.Theymanifestasspecific,irrecon-cilableconflictsinday-to-dayoperations.Ourresearchidentifiedfourdistincttensionsthatemergewhenorga-nizationstrytointegrateagenticAIintoexistingwork-flows.Eachtensionrepresentsafundamentalclashbetweenestablishedmanagementprinciples,forcingleaderstochoosetoeitherapplyincompatibleapproachesordevelopentirelynew,hybridframeworks.Successfullynavigatingthischallengerequiresthatleadersmanagethefollowingtensions:

1.Sca1abi1ityversusadaptabi1ity.Toolsscale

predictably;workersadaptdynamically.Agentic

AI’sabilitytodobothsimultaneouslyrequiresneworganizationaldesignprinciples.

2.Experienceversusexpediency.Whenistherighttimetoinvestinagenticsystemsandhowshouldthoseinvestmentsbemade?Leadersarefacedwithbalancinglong-termcapabilitybuildingwithshort-termreturns.

WHYAGENTICAISPREADSSOFAST

TherapidadoptionofagenticAIalignsperfectlywiththediffu-sionofinnovationstheory,whichexplainswhysometechnol-ogiesspreadfasterthanothers.iAccordingtotheframework,innovationsdiffusemorequicklywhentheyofferaclearrelativeadvantage,arecompatiblewithexistingsystems,aresimpletotry,andhaveobservableresults.

AgenticAIexcelsonallfourdimensions.Itbuildsonthefamil-iarityofgenerativeAIthatorganizationshavealreadyadopted,integratesseamlesslyintowidelyusedplatforms,requiresmin-imalnewinfrastructureforexperimentationwithit,anddeliv-ersimmediate,visibleproductivitybenefits.AsChevron’schiefdataandanalyticsofficer,MargeryConnor,explains,“Withthecompanystandardizedonasinglevendor’splatform,morethanhalfoftheworkforcehasaccesstoAItools,and,byextension,accesstoagenticAI.”Theinfrastructureisalreadythere.

Thisbuilt-inavailabilitydramaticallyreducesbarrierstoadop-tion.Employeescanbeginusingagenticfeatureswithoutimplementingentirelynewsystems,whichencouragesorganicexperimentationacrosstheorganization.Asmorepeopleexpe-riencetangiblebenefits,adoptionacceleratesthroughwordofmouthandviademonstrationeffects—exactlywhatdiffusiontheorypredicts.

Theimplicationforleaders:AgenticAI’srapidspreadisn’tanaccident.It’shappeningbecausethetechnologyisdesignedtominimizeadoptionfriction.Thismeansthatcompetitivebene-fitswillcomenotfromearlyaccesstothetechnologybutfromsuperiororganizationaldesignaroundit.

4MITSLOANMANAGEMENTREVIEW?BCG

TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI5

3.Supervisionversusautonomy.Howdoyou

supervisesomethingdesignedtoworkautonomously?

Traditionaloversightmodelsassumeeitherfullhuman

controlorcompleteautomation,notsystemsthatrequiresomehumancontrolanddifferingdegreesofautomation.

4.Retrofitversusreengineer.When,andbyhowmuch,shouldorganizationschangeprocesses?Thatdecisionrequiresresourcesandattentionthatmostchange-managementframeworksdon’taddress.

Theorganizationsthatwillsucceedarethosethatrec-ognizeagenticAI’sdualnatureasafeature,notabug.Strategiesthatembracetheambiguityanddevelophybridapproachesratherthanforcingthesesystemsintoexistingmanagementcategoriesbenefitfromboththeirtool-likescalabilityandworkerlikeadaptability.

1.TheFlexibilityTension:ScalabilityVersusAdaptability

Humanworkersaremaximallyflexible.Theycanswitchtasks,learnnewskills,andadapttounexpectedsitua-tionswithminimalretraining.Toolsaremuchlessflexible.Machineryandinfrastructureexcelatspecificpurposesandscalepredictably,buttheycanstruggletoadapttochange.

AgenticAIsitsinbetween:moreadaptablethantoolsbut(currently)lessflexiblethanworkers.Howshouldorga-nizationsdesignprocesseswithintermediateflexibility?AgenticAI’sabilitytoscalepredictablyandadaptdynam-icallyrequiresneworganizationaldesignprinciplesthatdon’tfitneatlyintoexistingmanagementsystems.

IntheheartofGoodwillIndustries’bustlingsortingfacil-ity,ariverofdonationsflowsceaselesslywithachaoticmixofvintageT-shirts,designerjeans,andforgottenfabrics.Ateamofdedicatedworkerspainstakinglysiftsthroughthismountainoftextiles,butAIagentsareincreasinglyabletohelp.Thissystemisnotasimplecom-putervisionmachine,rigidlyprogrammedtorecognizeonlyafewspecificitems.Instead,itisaflexibleAIthatconstantlylearnsfromandadaptstotheever-changingstreamofgoods:Itcanlearntodistinguishbetweenacashmeresweaterandawoolblend,identifyararecol-lectiblefromacommontoy,andevenspotsubtlesignsofwearandtearthatmightmakeanitemunsuitableforresale.StevePreston,presidentandCEOofGoodwill,says,“Oursupplychainfrombeginningtoendisverycomplicatedandrequiresalotofhumanintervention,alotofdecisionpoints,andalotofjudgment.Sothroughout

OurresearchidentifiedfourdistincttensionsthatemergewhenorganizationstrytointegrateagenticAIintoexistingworkflows.

thatprocess,weseealotofopportunitiestoincorporateAIintheentireflowofgoods,thedecision-makingpro-cess,andmakingsurethateverythingwereceivefindsitsbesthomesomewhereinthecycle.”

AI’sRole

Coach

Mentor

Colleague

Rival

Boss

Assistant

GoodwillispilotingAItechnologythatcouldultimatelybeusedtosortbillionsofpoundsofdonationseachyear,

ensuringthateachitemfindsitswaytotherightplace,whetherit’saboutique,anonlinestore,orarecyclingfacil-ity.TheAI’sabilitytohandlecomplexjudgmenttasksisencouragingleaderstoreplaceadecades-old,human-centricworkflowwithnewprocesses.TheyhavebeguntodesignprocessesforandwithagenticAI.

13%

42%

12%

36%

11%

35

%

8%

26%

5%

18%

26%

61%

%05101520253035404550556065

NowIn3years

FIGURE2

AI’sRoleintheOrganization

OurdatashowssurveyrespondentsbelieveAIwillbecomemorelikeanassistant,coach,mentor,orcolleagueratherthanarivalorboss.

6MITSLOANMANAGEMENTREVIEW?BCG

TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI7

Asthecompositionofworkersshifts,deployingscalable,flexibleAIworkerscreatesastrategictensionforprocessdesignacrosstheenterprise.

AIsystemsalreadyactasassistants,colleagues,mentors,coaches,andevenasrivalsandbosses.MostofoursurveyrespondentsanticipatethatAIwillbecomemoreinvolvedineachoftheseroleswithinthreeyears.(seefigure2,page6.)Asthecompositionofworkersshifts,deployingscalable,flexibleAIworkerscreatesastrategictensionforprocessdesignacrosstheenterprise.Agenticsystemsmightworkbestinstandardizedenvironments,butover-standardizationcaneliminatetheirabilitytolearnanddevelophumanlikeadaptability—acapabilitythatcanhelporganizationshan-dleedgecasesandsystemfailures.Organizationsneedboththeefficiencyofstandardizedprocessesandtheflexibilityofhumanlikeimprovisation.

Thethreat:OrganizationsthatoptimizeforAIefficiencymissoutonAI’shumanlikeadaptiveresponsestosystemfailuresorunexpectedmarketshifts.

Theopportunity:CompaniesthatstriketherightbalancecanachievebothAI-drivenefficiencyandhuman-poweredadaptability,creatingstrategicbenefits.

2.TheInvestmentTension:ExperienceVersusExpediency

Traditionaltoolsrequirelargeupfrontcostsbutdeliverpredictablereturnsthroughestablisheddepreciationschedules.Humanworkersareanongoingvariableexpense,buttheirvalueappreciateswithexperienceandtraining.AgenticAIdefiesbothmodels,requiringsub-stantialinitialdevelopmentcostsandongoingvariablecosts,suchastrainingmodelsonnewdata.Whilemanytechnologysystemsrequireongoingmaintenance,agen-ticAIsystemssimultaneouslydepreciatethroughmodeldriftwhileappreciatingthroughfine-tuningandemergent

capabilities.ShouldinvestmentsinagenticAIbeviewedmorelikeinvestmentsintoolsorinworkers,orboth?

Directlyaddressingthatquestionrevealstwocriticalinvestmenttensionsthatchallengeconventionalfinan-cialplanning:

Timing:Themoving-targetprob1em.AIsystemsareevolvingquickly,creatinguncertaintyaboutwhenanorga-nizationshouldmakesignificantinvestmentsinthem.Adopttooearlyandrisktechnologicalobsolescence;waittoolongandriskstrategicbenefits.AsJeffReihl,execu-tivevicepresidentandtechnologychairmanatLexisNexisLegal&Professional,observes,“Thistechnologyischangingsofast,wemighthavetodoaquickcatch-up.”Chevron’schiefdataandanalyticsofficer,MargeryConnor,echoesthissentiment,describingtheneedtoremainadap-tivetoemergingtoolsandupdates:“Thefast-paceddevel-opmentofagenticAIrequiresorganizationstobeagilewhileconsistentlyupholdingtheirdataandAIgovernancestandards,”shesays.

Unliketraditionaltoolswithpredictableupgradecycles,agenticAIrequirescontinuousadaptationandlearning.Standardnetpresentvaluecalculationsfailwhenthemostvaluableapplicationshaven’tbeenconceivedyet,andconventionaltimingmodelscan’taccountforthespeedoftechnologicalevolution.Plus,applyingconventionalreplacementschedulesrisksrapidvaluedecayassystemsfallbehindthetechnologicalcurve.

Size:P1atformsversuspointso1utions.Largeorga-nizationsfaceafundamentalchoice:investheavilyincomprehensiveAIplatformsorpursuesmaller,targetedpointsolutions.Thescaleofrequiredinvestmentforeach

approachcanvarydramatically,makingitdifficulttoaccu-ratelygaugeupfrontandcontinuingcosts.

Platforminvestmentsdemandsubstantialupfrontcom-mitmentswithuncertainreturns.AtCapitalOne,PremNatarajan,executivevicepresident,chiefscientist,andheadofenterpriseAI,describesbuilding“dozensofusecasesatscale”fromasingle,substantialplatforminvest-ment.ThefullimpactofCapitalOne’sstrategycanbeassessedthroughthelensesoftechnologyexploitationandexploration.Similarly,WalterSun,seniorvicepresi-dentandglobalheadofAIatSAP,explainsthatcreating“agenerativeAIhub”allowsforfulllife-cyclemanagementoflargelanguagemodels,whereasbuildingLLMsindivid-uallyintoisolatedapplicationsrequiresthecostlyintegra-tionoflegacysystemsbeforeanyreturnsmaterialize.SAPfocusedonthevaluetothedeveloperecosystemandusedthattodeterminewhatwouldbeasufficientreturnfromhavingaplatform,Sunnotes.

InvestinginpointsolutionsoffersmorepredictablecostsandmeasurablereturnsbutrisksmissingthecompoundvaluethatemergeswhenAIcapabilitiesareintegratedacrossbusinessfunctions.Organizationsthatchoosethispathmayfindthemselvesmanagingdozensofdiscon-nectedAItoolswithouttheinfrastructuretoscaleoradapt.

Organizationsinvestingsolelyinoneapproach,whetherhuman-in-the-loopsystemsorfullyautonomousagents,missthecompoundvaluecreatedbyagenticAI’scon-tributionsinmultipledecision-makingscenarios.Eachautonomylevelservesdifferentrisktolerancesandbusi-nesscontexts,butmeasuringROIinisolationobscuresthestrategicvalueofadiversifiedAIportfolio.

AgenticAI’shybridfeaturesmakeitdifficulttoaccu-ratelymeasureinvestmentreturnsovermeaningfultimeframesoreventoassessupfrontinvestmentrequirements.CompaniesmeasuringagenticAIreturnsthroughconven-tionaldepreciationschedulessystematicallyundervaluethecontinuous-learningandemergentcapabilitiesthesesystemsgenerate,failingtoaccountforsignificantportionsofactualvaluecreation.

Thethreat:Organizationsthatapplytraditionalinvest-mentframeworkstoagenticAIsystematicallyunderinvestincontinuouslearningandadaptation,leadingtorapidvaluedecaywhilethecompoundreturnsavailablefromintegratedAIecosystemsaremissed.

Theopportunity:CompaniesthatembracehybridinvestmentmodelsanddiversifiedAIportfolioscancre-atecompoundingreturnsastheirsystemslearn,adapt,andgenerateunforeseencapabilitiesacrossmultipleautonomylevelsandbusinesscontexts.Thewindowforestablishingthisstrategicadvantageisnarrowingascompetitivepres-suresintensifyandtechnologicalevolutionaccelerates.Thequestionisn’twhethertoresolvethesetensionsbut,rather,howquicklyorganizationsadapttheirinvestmentframeworkstomatchtherealityofagenticAI.

3.TheControlTension:SupervisionVersusAutonomy

Toolsarefullyownedandcontrolled,behavingpredict-ablyoncedeployed.Workersmustbemanagedthroughcontracts,incentives,andoversightbecausehumanshaveautonomyandmaypursuedivergentgoals.AgenticAIrequiressupervisionandmanagementlikeaworkerdoes

Unliketraditionaltoolswithpredictableupgradecycles,agenticAIrequires

continuousadaptationandlearning.

8MITSLOANMANAGEMENTREVIEW?BCG

TheEmergingAgenticEnterprise:HowLeadersMustNavigateaNewAgeofAI9

+55%

37%

57%

+179%

14%

39%

+250%

10%

35%

0102030405060

NowIn3yearsPercentageofrespondentswhostronglyagreeoragreewitheachstatement.

AIsystems workwithambiguousinputs

AIsystemsworkindependently fromhumans

AIsystemshavedecision-makingauthority

FIGURE3

OrganizationsAnticipateContinuedIncreasesinAIAutonomy

SurveyrespondentsaretwotothreetimesmorelikelytoexpectAIsystemstoworkindependentlyfromhumansandhavedecision-makingauthorityinthreeyearscomparedtotoday.

becauseitsoutputscanbeunpredictable,eventhoughorganizationsownitlikeanautonomoustool.Howcanorganizationsdesignprocessestoeffectivelysuperviseanagentthatalsoworksautonomously?(seefigure3.)

Aprocessthatincludessupervisioninherentlykeepshumansinvolvedindecision-makingprocesses.Chevron’sConnornotes,“Wealwayshaveahumaninthelooptoreviewandanalyzetheoutputsowecandeterminewhetheritmakessenseornot.”ChandraKapireddy,for-merheadofgenerativeAI,machinelearning,andanalyticsatTruistBank,says,“Ifyoulookatthefinancialservicesindustry,Idon’tthinkthereisanyusecasethatisactually

customer-facing,affectingthedecisionsthatwewouldmakewithoutahumanintheloop.”1

Despiteowningthetechnology,organizationsmusttreatagenticAIsystemswiththesameoversighttypicallyreservedforhumanemployees.

Butatthesametime,organizationswantthebenefitsofscalethatAIagentscanofferwhileavoidingthebottle-necksthathumanoversightcouldcreate.Oneleaderwespokewithasked,“Howdoyoumakesurethatyouhavetherightsetofcontrolsinplace?We’recallingithuman-in-the-loop,or‘human-out-of-the-loop’forsomeofourlow-riskusecases.”Thedecisiontomovetohuman-out-

of-the-loopforlower-riskusecasesinvolvescontext-dependentassessmentsofrisk.Unlikeasimpletoolwithafixedfunction,agenticAIoperateswithinacomplexenvironmentwhereitsactionscanhaveunintendedconse-quences.Thatmeansorganizationsmustcraftpoliciesthatsetboundariesfordecision-makingandactiontoserveassafeguardssothatorganizationalbehaviorsalignwithstrategicobjectives,andnegativeoutcomesareprevented.Thisisanalogoustosettingpoliciesandprovidingover-sightforhumanemployees.SunatSAPdescribesbuildingagenerativeAIhubforthispurpose:“Throughthat,wecanthenconnecttodifferentlargelanguagemodels.Butwealsohavetheability,then,toputintheguardrailsforallthebusinessapplications.Wecanputinourownanalytics.Wecanputinourownsecurity,ourprivacy,ourcompli-ance,andsoon.”TheneedforthesecontrolshighlightsthefactthatAIisnotafullypredictabletoolbutanagentthatmanagersmustguideandconstrain.

Whenatoolmalfunctions,it’sadefect.Whenacoworkermakesamistake,it’samanagementandlearningoppor-tunity.ManagerscanapplythiscoworkermindsettoAI.RebeccaFinlay,CEOatthePartnershiponAI,arguesthatcompaniesneedtobemoretransparentabouttheirAIfailures:“Iwouldloveforustobeinapositionwherewecouldsharemoreaboutthingsthatgowrong,becauseinthatsharing,welearn.”ThisapproachtreatsAInotasaninfallibletoolbutasanagentwhoseerrorsmustbeunder-stood,managed,andlearnedfrom.Thelearningprocessparallelshoweffectiveteamshandlehumanerror.

Thethreat:OrganizationsthatfailtodevelopappropriategovernanceframeworksforagenticAImayfacecompli-ancefailures,misalignedoutputs,orrunawaysystemsthatdamagebusinessoperations.

Theopportunity:Companiesthatmastertheartofman-agingartificialcolleaguescanscalespecializedcapabilitieswithoutthetraditionalc