BCG

AIAGENTS

HowAgentsAre

AcceleratingtheNextWaveofAIValue

Creation

By

VincianeBeauchene

,

NicolasdeBellefonds

,

DjonKleine

,

RichLesser

,

Amanda

Luther

,TomMartin,and

DanielSack

ARTICLEDECEMBER15,20258MINREAD

AI

istransformingthewayworkgetsdone,butithasn’tfundamentallyalteredthewaymostcompaniesoperate—yet.

That’sabouttochange.

?2025BostonConsultingGroup1

?2025BostonConsultingGroup2

Weareenteringtheeraof

AIagents

.WhilepreviouswavesofAIfocusedeitheronlogicand

optimizationorcreativityandsynthesis—the“l(fā)ebrain”and“rightbrain”ofpredictiveAIand

generativeAI

,respectively—thenextwavewillinvolveagenticsystemsthattakeonentire

work?ows,applyingjudgmentshapedbyacompany’sowninstitutionalknowledge.(Seethesidebar“TheFrontalCortexofAI.”)

TheFrontalCortexofAI

IfpredictiveAIisthelebrainforlogicandoptimization,andgenerativeAItherightbrainforcreativityandsynthesis,thenagenticAIservesastheexecutivefunctionthatturnscreativeprobabilityintobusinessimpact.

PredictiveAIopenedvalueindecision-makingfunctions.GenerativeAIopenedvalueinknowledgeandcontentproductionfunctions.AgenticAIisnowopeningvalueinprocess-heavyfunctionswhereexecutionde?nesperformance.

?2025BostonConsultingGroup3

Thesesystemsdomorethananalyzeandcreate.Theyobserve,plan,andacton

theirown.Theylearncontinuously.Andwhenembeddedinthebusiness,theytake

onthestepsthatsloworganizationsdown,whethercoordinatinginternal

work?owsorengagingdirectlywithcustomersandpartners.

Earlyadoptersarealreadyshowingwhatthislookslikeinpractice:

.Ashipbuildercutdesigne?ortforakeycomponentby45%andleadtimeby80%byusingorchestratorandworkeragentstorunamultistepdesignprocessacrossmorethan14

systems.

.Alenderreducedmortgageapprovaltimeby20%andfreedhalfofitsmortgage-assessorcapacitywithanagentsystemthathandles?aggedapplicationsendtoend.

.Apayrollproviderresolvedanomaliesautomaticallythroughasupervisoragentsupportedbyspecializedworkeragents,improvingprocessingspeedbymorethan50%.

Resultslikethesearespurringrapidinvestmentacrossindustries.InarecentBCGstudywith

MIT

SloanManagementReview

,35%oforganizationssaidtheyarealreadyusingagenticAI,andanother44%saidtheyplantodososoon.Thatsameresearchfoundthatthree-quartersofextensive

adoptersbelieveAIisnowenablingnewsourcesofvalueandcompetitiveadvantage.

Productivitygainsareoentheinitialbene?tofAIadoption,buttherealprizeintheagenticageisdi?erentiation.AndtherealchallengeissustainingadvantagewhenAIcanenablenew,disruptive

?2025BostonConsultingGroup4

competitorsandshrinktraditionaladvantages.

The?rststepintheagenticAIjourneyistonarrowtheagendatotheprioritiesthatcreate

advantageandvalue.DrawingonBCG’sexperiencewithearlyadopters,we’velaidoutaroadmapforwhatCEOsshoulddonow,whattheyshouldanticipatenext,andtheprinciplestheyshould

alwaysfollowtoachievethehighestreturnsfromAI.

Now:ThreeMovestoMakeToday

Leadersshouldusethenextyeartobuildtheiragenticfoundation.Thiswillnotbecompletedin12months,ofcourse.Butmakingthesemovesnowwillhelpbuildthecapabilitiesandcon?dencetoscalebroadlyintheyearsahead.

Redesignworkaroundzero-based,outcome-drivenprocesses.Acommonmistakeisautomatingwhatalreadyexists.Realvaluecomesfroma“zero-based”approach,startingwiththeoutcomeyouwantandreinventinghowtodeliverit.Rethinkhowtasks?ow,wheredecisionssit,andhowroles

changeasagentstakeonmoreexecution.Spellouttheskills,responsibilities,andtheshapeoftheteamyou’llneedoncethesystemisupandrunning.

Oneglobal

consumergoods

companydidthiswithitsproductinnovationprocess.Insteadofspeedingupindividualsteps,theystartedfromscratchandrebuiltthework?owaroundmeta-

?2025BostonConsultingGroup5

agentsthatorchestratedtheworkandworkeragentsthatgeneratedinsights,testedideas,andcreatedconcepts.Theuni?edsystemcutcycletimebyroughly60%andimprovedproduct?t.

Embedyourcompany’scontextandproprietaryintelligence.Contextdrivesindividualbehavior,whichshapesoverallperformanceinhuman-centricorganizations.Thesameistruefor

organizationsleveragingAIagents.Theyneedthesamerichcontextyourbestworkersareguidedby,fromdataandtoolstonorms,tacitknow-how,constraints,andintent.ForAIagentstomove

fromgenericoutputstodistinctiveinstitutionalintelligence,youmustbuildabusinesscontextfabric.Itconsistsofthreecomponents:

.Objectives:De?netheoutcomes(notoutputs)theagentmustdeliver,suchashigherconversionorfastercompliance.

.Resources:Connecttheagenttothedata,tools,anddecisionlogicyourtopteamsuse.

.Constraints:Settheguardrailsthatkeepautonomousactionssafeandconsistent,includingregulatoryrules,brandtone,andriskappetite.

Oneglobalconsumergoodscompanyidenti?editsbusinesscontextfabricbyholdingexpert

workshopsthathelpedsurfacetheunwrittenprinciplesitscreativeteamfollowedindeveloping

strongcampaigns.Theytranslatedthattacitknowledgeintothepromptsanddecisiontreestheiragentsfollow.Theresultwasbrand-consistentrecommendationswithfarlessrework.

AssembleasharedAIplatform.Toscale,youcannothaveeveryfunctionbuildingitsown

infrastructure.YouneedasharedAIplatformthatgivesP&Lowners“freedomwithinaframe.”

Thinkofitasanenterprisesystemthateveryonebuildson.

Theplatformshouldincludedataconnectors,orchestrationframeworks,modelmanagement,

monitoring,andtieredautonomywith

responsibleAI

controls.Thesesharedcomponentsletteamsinnovatefasterwithoutrebuildingthebasicsorcompromisingsecurity.

Inthisnewmodel,ITandthebusinessfunctionsoperateaspartners:

.ITmaintainsthechassis:Theymanagethedata,infrastructure,andsafeguardsthatensuresystemsandAIagentsarereliable,compliant,consistent,anduptodate.

.Thebusinessdrivesthedesign:Functionalleadersandteamsusetheplatformtodesignagentsthatdeliveroutcomes,embeddingtech-?uenttalentthatownsthenewagentic-

enabledprocessesasproductsandenhancesthemovertime.

Theimpactissigni?cant.Inourworkwithclients,we’veobservedthatsharedtoolsanddatacan

cutcostsbyupto30%andliproductivityby25%.Asreuseincreases,timetomarketcanimproveby50%.

?2025BostonConsultingGroup6

Next:TwoDisruptionstoPlanFor

Asagentstakeonexecution,CEOsmustpreparefortwomajordisruptions.

ARadicallyAlteredEnterpriseOperatingModel.Our

researchwithMIT

revealsthatasAIagentshandleroutineexecution,thetraditionaloperatingmodelwillberadicallyaltered.Amongthe

changes:

.Theworkforcewillberebalanced.Juniortalentwillspendlesstimeproducing“?rstdra”

workandmoretimedirectingit.Amongheavyadopters,29%ofcompaniessaythey’llneed

fewertraditionalentry-levelroles,and43%expectgreaterdemandforgeneralistswhocan

managehuman-agentteamsandhybridwork?ows.Earlyevidenceofworkforcedisruptionisalreadyvisible—StanfordUniversityresearchonpayrolldatafounda16%declinein

employmentamongearly-careerworkers(ages22to25)inAI-exposedroles.1Thebalance

acrossfunctionsandrolesalsowillshi:somework?owswillbefullyautomated,otherswill

stayhuman-led,andnewroleswillappearasolderonesfade.Thismakesskills-based

strategic

workforceplanning

essential.

.Theorganizationdesignwillshi代.Ashumansmoveawayfromexecutionandtoward

outcome-driven,cross-functionalteams,45%ofAIleadersexpecttoneedfewermiddle-

managementlayers.Workismovingtodiverseteamsthatcombinefunctionalexperts,tech

anddata-scienceknow-how,andrelevantlegalandhuman-resourcetalent.Theirworkcentersonoversight,orchestration,andcontinualimprovementoftheagent-drivenprocessesratherthanmanualdelivery.

.Governancewillevolve.Rightnow,thevastmajorityofcompaniesarebuildingmost

processeswithhumansintheloopandwithultimatedecisionauthority.ButasAItakeson

moredecisionrights,58%ofheavyadoptersexpectafundamentalshiingovernanceoverthenextthreeyears,andone-thirdbelieveAIwillhavemoredecision-makingauthorityinthesameperiod.TheseusersaredrawingclearerlinesbetweenhumanandAIroles.They’realsobuildingcross-functionalgovernancemodelstomonitorandcoordinateautonomousactivity,developandenforcestandards,andmeasureoutcomesandrisksacrosssystems.“Freedomwithinaframe”mustincludecarefulriskmanagementthatdrawsoncoreresponsibleAI

principles—clearrisktaxonomies,rigoroustesting,continuousmonitoring,andbuilt-in

resiliencefromthestarttoensureagentsactwithinboundaries.(Seethesidebar“TheTrustProtocol:AFrameworkforGraduatedAutonomy.”)

TheTrustProtocol:AFrameworkforGraduated

?2025BostonConsultingGroup7

Autonomy

ThegreatestbarriertoscalingagenticAIisn'ttechnology,it’strust.CEOshesitateforgoodreason.Noonewantsasystemthathallucinatesadiscount,o?endsa

client,ormisallocatescapital.

Tomovefrompilotstoscale,leadersshouldadoptagraduatedautonomy

framework.Thinkofitasapromotionpathinwhichagentsearntheirautonomythroughprovenperformancethresholdsandrigoroustesting.Thisframeworkhasfourtiers:

.Tier1:ShadowMode(Agent-Assisted).Inthismode,theagentsuggestsandthehumanacts.Theagentoperatesasa“digitalwhisperer,”observingdata

anddraingoptionsoridentifyingmissinginformation,muchlikeweuse

LLMssuchasChatGPTtoday.Thehumanretainsfullcontrolandmanually

executesthe?nalstep.Intermsofstrategicvalue,thisisthetrainingground.Itallowstheorganizationtovalidatetheagent’slogicagainstthebusiness

contextwithoutoperationalrisk.

.Tier2:SupervisedAutonomy(HumanintheLoop).Inthistier,theagentactsandthehumanapproves.Theagentpreparestheactionandstagesit.

Executionispauseduntilahumanprovidesadigital“thumbsup.”Ifthe

agent’scon?dencescoredropsbelowasetthreshold(suchas90%),itrevertstoTier1.Thisapproachisforhigh-stakesdecisionswhereaccuracyis

paramount,suchas?nancialauditingor?nalizingcomplexbusinessproposals.

.Tier3:GuidedAutonomy(HumanontheLoop).Inthistier,theagentactsandthehumanmonitors.Theagentexecutesautonomouslywithinstrict

guardrails.Humansshifromapproverstoexceptionhandlers,interveningonlywhenthesystem?agsananomalyorduringrandomspot-checks.Thisapproachunlocksspeedforroutinetasks(suchassupplychainrestocking)whilemaintainingasafetynetforedgecases.

.Tier4:FullAutonomy(HumanoutoftheLoop):Inthistier,theagent

operatesindependentlywithoutexplicithumanoversightforde?ned

work?ows.Thistierisreservedforhighlymature,low-riskenvironmentswherethecostoferrorisnegligiblecomparedtothegainine?ciency.

SuccessfulimplementationsrequireagentstoruninTier1(ShadowMode)untiltheyprovetheyalignwithyourorganization’sriskappetite.Autonomyisajourneyoftrust,quanti?edbyaccuracy.

?2025BostonConsultingGroup8

ARecalibratedTechStrategy.Aslaborcostsriseanddigitalcostsfall,companiesareshiingtheirspendingfrompeopletotechnology.Inthatcontext,highertechnologyspending(forexample,35%to45%ofoperatingcostinretailbanking,upfrom20%to30%today)re?ectsarebalancedcost

mix,notabudgethike.Thisshimakestechstrategyinseparablefromcompetitivestrategy.

Asthetechnologyecosystemconsolidatesaroundafewmajorplatformsandalongtailof

specialists,leadersneedtokeepasteadyhandontheincreasinglycomplexpartnership

environment.Cloudandmodelprovidersareexistentialdependenciesthatmustbetreatedas

strategicalliances.Leadersalsowillneedtokeepsystemsmodulartoavoidlock-inandprotecttheproprietaryvalueoftheirdata.

Always:FiveEnduringPrinciples

Technologychanges,butthefundamentalsofAItransformationremainconstant.HerearesomeprinciplesCEOsshouldfollownowandintheyearstocome.

Focusonafew“reshapeandinvent”bigbets.Don'tspreadAIacrosstheorganizationtoachieveamyriadofsmallproductivitygains.Tocreaterealvalueanddistinctiveadvantage,concentrate

resourcesonreshapingend-to-endwork?owsandinventingnew,revenue-generatingbusiness

models.Gettheearlywinsfromthesebigbetsincriticalareastobuildthecon?denceandmuscletoprovidethefoundationtoexpand.

CombinemultipleformsofAIandtechnologyforimpact.Themostpowerfulsystemscombine

predictiveAItooptimizedecisions,generativeAItocreatecontent,andnowagenticAIto

orchestrateexecution.Tobuildthem,leveragewhatalreadyexists,integratingcurrentsystemsanddataratherthantreatingthemassilos.

Followthe10/20/70rule.Algorithmsaccountfor10%oftheworkinAItransformation,thetech

backbone20%.Theremaining70%comesfrompeopleandprocesses.AIagentsfundamentally

changehowworkgetsdoneandbywhom;assuch,thelion’sshareoftheleadershipteam’se?ort

mustgointoredesigningroles,managingchange,andtrainingyourworkforcetoprovidetherightoversightandguidancetoAIagents—andtoreinforceexpertisewherehumansmakeadi?erence.

?2025BostonConsultingGroup9

Makethetransformationbusiness-led.TheITfunctioniscriticaltobuildingtheinfrastructure,buttheycannotdesigntheworkonyourbehalf.Businessleadersunderstandthenuanceofthe

work?ows,wherejudgmentmatters,andwhereAIagentscantrulymovetheneedle.Theymust

leadthetransformation,inclosepartnershipwithIT,toensureagentsaresolvingtherightbusinessproblemsandbeinggiventherightcontext,data,andconstraintstoensurehighqualityand

responsibleoutcomes.

Investinarobustdatafoundation.AIagentsrelyonhigh-quality,connecteddatatoreason

e?ectively.Don’twaitforaperfectdatalake.Investincrementallyanduseeachnewagent

implementationtoexposeandclosespeci?cdatagaps,sothatthecorefoundationgetsstrongerwitheverystep.

TheCEO’sNextStep

AIagentsmarkaturningpointinhoworganizationscreatevalueandcompete.WithAI-native

?rmsalreadyachieving25to35timesmorerevenueperemployeecomparedtotraditionalpeers,thepotentialisclear.

Themodelsareready.Thechallengeiseverythingaroundthem—thesystems,data,andwork?owstheydependonandtheorganizationalcapabilitiesneededtousethematscale.Asbasic

productivitybecomesuniversal,advantagewillshitothosethatmasterintegrationanduse

agentsnotonlytodrivee?ciency,buttounlockproprietaryintelligenceandcreatenewrevenuestreams.

Inthislandscape,strategytakesongreaterweight.CEOsneedaclearsenseofwheretostart,howtosecureearlywinsthatstrengthencapabilityandcon?dence,andhowtoextendthat

momentumacrossthebusiness.The“Now,Next,andAlways”frameworkhelpsleadersmakethosechoices—whilelayingthefoundationforenduringadvantage.

?2025BostonConsultingGroup10

Nicolasde

Bellefonds

ManagingDirector&SeniorPartner

Paris

RichLesser

GlobalChair

NewYork

TomMartin

PlatinionDirector

London

Authors