AIinaction

HowgenAIandagenticAI

redefinebusinessoperations

RESEARCHINSTITUTE

#GetTheFutuΓeYouWant

2

AIinaction

Tableof

contents

06

10

ExecutivesummaΓy

AsoΓganizations

realizeAIbenefits,ROIconceΓnswane

28

36

GenAIandagenticAIadoptionissoaΓing

AI-drivenprocess

transformationsare

deliveringvalueacrossbusinessoperations

CapgeminiResearchInstitute2025

3

AIinaction

CapgeminiResearchInstitute2025

Tableof

contents

54

78

Preparingyour

organizationfor

AI-poweredbusinessoperations

Conclusion

methodology

79

ReseaΓch

84OuΓseΓvices

4

CapgeminiResearchInstitute2025

AIinaction

Whoshouldreadthis

reportandwhy?

Thisreportisprimarilyintendedforseniorexecutivesin

businessoperations—particularlythoseleadingfunctionssuchassupplychain,procurement,finance,andcustomerorpeopleoperations.ItwillalsobevaluabletoCTOs,

CDOs,andothertechnologyandinnovationleadersexploringthestrategicroleofGenAIandagenticAIinenterprisetransformation.

Thisreportoffersunique,data-driveninsightsinto

howGenAIandagenticAIarebeingadoptedacross

industries,withafocusonROItrends,emergingusecases,andtheirtangibleimpactoncorebusinessfunctions.

Italsoexploresworkforceimplications,governancebestpractices,andscalingstrategies—makingita

criticalresourceforleaderslookingtomovebeyondexperimentationanddrivemeasurablevaluefromAIinvestmentsinbusinessoperations.

CapgeminiResearchInstitute2025

5

AIinaction

WhatisAI,GenAI,andagenticAIinbusiness

operations?

Inthisreport,“AI,GenAIandagenticAIinbusiness

operations”referstotheintegrationoftraditionalAI,AI-enhancedprocessautomation(IPA),generativeAI(GenAI),andagenticAIintoorganizational

workflowstoimproveefficiency,decision-making,andinnovation.

Keyterminologyusedinthereport:

?Artificialintelligence(AI)isacollectiveterm

fortheintelligentcapabilitiesinlearningsystems,typicallycategorizedintomachinevisionand

sensing,naturallanguageprocessing(NLP),

predictinganddecision-making,andactingandautomating.

?GenerativeAI(GenAI)isasubsetofAIthat

harnessesthepoweroftransformermodelsand

massivescalingofdataandcomputetoplan,reason,andcreategenerativefeaturesincludingtext,image,andvideo.

?AIagentisasoftwareprogramthatcaninteractwithitsenvironment,collectdata,andusethistoautonomouslyperform

taskstomeetpredeterminedgoals.

Asanevolutionfromtechnologieslikeroboticprocess

automation(RPA)andmachinelearning(ML),AIagents

can,perceive,reason,andactinchangingenvironmentstoachievetheirgoals.AIagentsemployarangeofadvancedtechnologiestointeractwithusersandperformtasks

autonomouslyandeffectively.Largelanguagemodels

(LLMs)areoftentheprimaryinterfacebetweenAIagentsandusers.Anagentcanunderstandandgeneratehuman-liketextorverbalresponsesusingnaturallanguage

processing(NLP),makinghuman-AIinteractionsmorenaturalandefficient.1

?AgenticAIisthedeploymentofAIagentsinareal-worldenvironmentwhereagentscandetectsignals,planand

reason,makeautonomousdecisions,andachievesetgoalswithouthumanintervention.(Notethatthedefinitionofthistermisnotyetstableacrosstheindustryandmayvarybetweendifferentprojectsandcontexts.)

Note:Inthisreport,theterm“AI”encompassestraditionalAI,AI-enhancedprocessautomation(IPA),GenAI,andagenticAIcollectively,unlessexplicitlystatedotherwise.

CapgeminiResearchInstitute2025

6

AIinaction

Havingshiftedfromexperimentalproofsofconcept(PoCs)toin-productionoperationalAIsystems,

Executive

summary

businessesarebeginningtorealizethebenefits.Oursurveyof1607organizationsshowsthat,returnsoninvestment(ROI)averaginganimpressive1.7xon

AIinvestmentsinbusinessoperations.aConfidence

inAI'scommercialviabilityisgrowing,with40%of

organizationsexpectingpositiveROIwithinonetothreeyearsandanother35%withinthreetofiveyearsbasedonoursurvey.AIagentsandmulti-agentsystemsdeliversignificantimprovementsinoperationalefficiency,costreduction,customersatisfaction,anderrorreduction.

InvestmentinintegratingAIintobusinessoperationsisrising,with62%oforganizationsincreasingtheirGenAIspendingthisyear,and36%allocatingcapitalspecificallytoGenAI.Threeoutoffourexecutivespreferusing

proprietarymodelsforAIimplementationinoperations,valuinghighperformanceandeasyintegrationwith

enterprisesystems.

In2025,GenAIdeploymentinbusinessoperations

surged,with36%oforganizationsdeployingthe

technologyatalimited/fullscale,upfrom20%in2024.Amongthese,30%haveintegratedAIagentsintotheiroperations.

aBusinessopeΓationsencompassthecoordinatedactivitiesandprocessesundertakenbyvariousdepartments

withinanorganizationtoproduce,market,anddelivergoodsorservices.Theseoperationsintegratefunctions

suchascustomerservice,productmanagement,marketing,andsupplychainmanagementtoensureefficiency,

profitability,andalignmentwiththeorganization’sstrategicobjectives.Inourresearch,wefocusonfouΓprimarybusinessfunctions:supplychainandprocurement,financeandaccounting,peopleoperations,andcustomeroperations,astheseareascollectivelyrepresentthecoreoperationalpillarsofmostmodernenterprises.

7

AIinaction

Insupplychainandprocurement,AIenhancesroute

optimizationandwarehousedesign,streamlining

fulfillmentandreducingoperationaloverhead.Inpeopleoperations,GenAIautomatestaskslikerésuméscreeningandcandidatematching,acceleratinghiringcyclesand

loweringrecruitmentcosts.Thesetransformationsare

drivingleaner,faster,andmorecost-effectiveoperationsacrosstheenterprise.Byembeddingatargetedsetof

AIcapabilitiesintocorebusinessprocesses–suchas

procurement,customerservice,supplychainoptimization,andfinancialoperations—organizationsareachieving

measurableefficiencies,leadingtocostreductionsrangingfrom26%to31%.

TheuseofAIagents,includingmulti-agentsystems,has

morethandoubled,with21%oforganizationsutilizing

themin2025(comparedwith10%in2024).Reported

Executive

adoptionratesmaybeoverstatedduetovarying

summary

definitionsofAIagentsversusGenAIassistants.While

surveysindicatestrongmomentum,clientandpartner

feedbacksuggestsactualAIagentadoptioncouldbemorelimited.Thesurvey’sbroadphrasingof“AIagentuse”mayincludeeverythingfrompilotstofull-scaledeployments.Comparedtocurrentlevels,agenticAIprojects(in

production)areexpectedtoriseby48%thisyear.

AIisreshapingbusinessprocessesandfunctionssuchassupplychainmanagement,finance,peopleoperations,andcustomeroperations,deliveringsignificant

efficienciesbyembeddingintelligenceintocoreworkflows.

8

AIinaction

?EmbraceagenticAIfortransformational

benefits:AdoptingagenticAIatscalethrough

phasedimplementationenablesoperational

transformation,betterdecision-making,andenhancedcustomerexperiences.

?Maintainastrictfocusoncostcontainment:

FinancialdisciplineinAIadoption—guidedbymetricslikecostperinferenceandROI—ensuresinnovationremainseconomicallysustainable.

?DeviseastrategyforscalingupAI-powered

processes:ScalingAIsuccessfullyrequiresaclear

build-versus-buystrategythatbalancesinnovationwithoperationalstabilityandlong-termadaptability.

TodevelopAI-drivenbusinessoperations,organizationsmustfollowsixessentialsteps:

Executive

?BuildafoundationofAIreadiness:EstablishingAIreadinessrequiresalignedleadership,strong

summary

governance,widespreadAIliteracy,digitalbusinessoperationsandrobustdatainfrastructuretoensurescalableandeffectiveAIinitiatives.

?MaketheworkforceAI-ready:SuccessfulAI

integrationdependsonchangemanagement,culturaltransformation,andempoweringemployeesto

collaborateeffectivelywithAI.

?Developastrongapproachtoprocessredesign:

AstrategicandstructuredprocessredesignembedsAIwhereitdeliversthemostvalue,drivingefficiencyandinnovation.

CapgeminiResearchInstitute2025

9

AIinaction

We'dalsolike

tothankthe

manyindustry

executiveswhosharedtheir

valuableinsightswithus.

AnnaKopp

DigitalLeadGermany,Microsoft–Germany

KishorePandrangi

GlobalDirectorofCustomerSuccess,Google–USA

DanielVassilev

Co-FounderandCo-CEO,RelevanceAI–USA

NicoleOnuta

LeadAIRiskManagement,ING–Netherlands

DeepakAnand

EnterpriseArchitectureLeader,UiPath–USA

Dr.WalterSun

SVP,GlobalHeadofAI,SAP–USA

EricPace

HeadofAI,Cox

Communications–USA

10

AIinaction

01

AsorganizationsrealizeAIbenefits,ROIconcernswane

CapgeminiResearchInstitute2025

11

Organizationsreportastrong1.7xROIfromAIinbusinessoperations

Inrecentyears,businessleadershaveraisedquestionsaboutwhetherthesubstantialexpenditureonAIandGenAIwillyieldcompensatoryAI-drivenbenefitsandreturns.2,3Butorganizationsthathaveconductedpilotprojects,achievedlimiteddeployment,orscaledtheseusecasesinvariousbusinessfunctionshavereportedaverageROIof1.7x.

AIinaction

Figure1.

OrganizationsachieveaverageROIof1.7xfromAIinvestmentsinbusinessoperations

AverageROIfromAIinvestment

575

1.7x

343

1.5x

2.1x

1.5x

1.7x

156

→140

116

91

→163108

69

75

Financeand

accounting

Totalacrossall

functions

People

Customer

operations

Supplychainand

procurement

operations

Totalamount

invested($million)

Operationscost

reduced($million)

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=1,007executiveswhoarefrombusinessfunctionssuchassupplychainandprocurement,financeandaccounting,peopleoperationsandcustomeroperations.

CapgeminiResearchInstitute2025

CapgeminiResearchInstitute2025

12

AIinaction

O∩aveΓa9e,aΓou∩d40%ofexecutivesa∩ticipateachievi∩910-20%impΓoveme∩tsi∩keymetΓicssuchasi∩si9ht

accuΓacy,pΓoductivity,timetomaΓket(TTM),a∩dcustomeΓa∩demployeesatisfactio∩oveΓthe∩extthΓeeyeaΓs,

compaΓedwith32%whoexpeΓie∩cedthesamelevelof

be∩e?tsi∩thepastyeaΓ.A∩otheΓ20%ofexecutivesexpectmoΓetha∩20%impΓoveme∩tacΓosstheaboveme∩tio∩edpaΓameteΓsi∩∩extthΓeeyeaΓ.

Twoinfiveorganizations(40%)trackingROI

expecttoachieve

positiveROIinonetothreeyears

AΓou∩d40%ofoΓganizationstΓacki∩9ROl,expectto

achievepositiveROlfΓomAlwithi∩o∩etothΓeeyeaΓs,

ΓeHecti∩99Γowi∩9co∩?de∩cei∩thetech∩olo9y’s

commeΓcialapplicability.AnotheΓ35%a∩ticipateΓealizi∩9ROlwithi∩thΓeetofiveyeaΓs,hi9hli9hti∩9abΓoadeΓtΓe∩dofstΓate9ici∩vestme∩t.

Whiletimeli∩esvaΓybasedo∩factoΓssuchasi∩dustΓya∩d

usecasecomplexity,mostoΓ9a∩izatio∩saΓeco∩?de∩ti∩Al'spote∩tialtodΓivesi9∩i?ca∩tbusi∩essimpact.

These∩ioΓdiΓectoΓfoΓ9lobalpΓocuΓeme∩ta∩alytics,data

scie∩cea∩ddi9italataphaΓmaceuticaloΓ9a∩izatio∩says:"AIsignificantlyenhancescostsavingsandcostavoidance,whicharecrucialforsupplychainefficiency.TheROIforAI-drivencontractanalysisandvalueleakagepreventionsurpasses300%."

%

ofoΓ9a∩izatio∩stΓacki∩9ROl,expecttoachievepositiveROlfΓomAlwithi∩o∩etothΓeeyeaΓs

CapgeminiResearchInstitute2025

13

AIinaction

FiguΓe2.

Around40%oforganizationswhoaretrackingROIexpecttorealizeapositiveROIinonetothreeyears

AveragetimetoachievepositiveROIonAIinbusinessoperations

40%

24%24%

16%

13%

11%

2%

<1years1to2years2to3years3to4years4to5years>5years

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=300executiveswhoarefrombusinessfunctionstrackingROIasafinancialKPItoevaluatethesuccessofAIandGenAIinitiatives.

14

Thisisconsistentwithrecentestimates.Around

half(49%)ofUSGenAIdecisionmakersexpecttheirorganizationstoachieveROIonAIinvestmentswithinonetothreeyears,while44%saythreetofiveyears.4

AgenticAIinbusinessprocesseswillboostthesebenefits

AIagentsarebeingadoptedacrossenterprises,mid-

marketfirms,andSMBs,buteachsegmentisfocusing

ondifferentpriorities.Enterprisesareleadingadoptioninoperationsandcompliance-heavyareas,with46%ofusecasescenteredonfunctionslikeprocurement,HR,

andfinance—wherescale,control,andriskmanagementarekey.Customerserviceandsalesarealsoemergingasimportantareas,reflectinggrowinginterestinAI-drivenengagement.5Improvedcustomersatisfactioncanbe

tracedtoAIagents’abilitytoprovidepersonalized,

round-the-clockservice,instantresponses,andseamlessmultichannelintegration.

Themagnitudeoferrorreduction(+40%)isstriking,

especiallygiventhecomplexityoftasksAIagentstypicallyhandle.ThisindicatesagrowingrelianceonAIsystemsforoperationswhereprecisionisessential.

CapgeminiResearchInstitute2025

AIinaction

“AIagentsareexpected

todriveefficiencies

andreduceoperational

costs,withconservative

estimatesindicatinga

minimumof10%efficiencygains,andoptimistic

projectionsreaching25%”

JojiPhilip

DirectorofAI/MLproducts,Ericsson

15

“Multi-agentsystemsallowtaskstobebrokeninto

specializedroles,improvingefficiencyandreducing

errors.Agentscanrevieweachother’sworktominimizehallucinations–atrulyfascinatingapproach.”

DanielVassilev

Co-FounderandCo-CEO,RelevanceAI

CapgeminiResearchInstitute2025

AIinaction

Figure3.

AIagent/multi-agentsystemsresultinimprovementsrangingfrom40–45%acrosskeyparameters

ImpactofAIagents/multi-agentsystemsonkeyparameters

Increaseinoperationale?ciency

Improvementincustomersatisfaction

Reductioninerrors

Decreaseinoperationalcosts

45%

44%

43%

40%

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=125executiveswhoareinvolvedintechnology

implementation,andGenAIproductowners/AIdeliverymanagerswhoareutilizingAIagents/multi-agentsystems.

CapgeminiResearchInstitute2025

16

AIinaction

Initialresultsindicateuptoa40–45%improvementinkeyparametersfollowingthedeploymentofAIagentsand

multi-agentsystems.Whilethesefiguresareencouraging,theymustbecontextualizedwithinthecurrentscope

andmaturityofimplementation.Asignificantportionoftheobservedgainscanbeattributedtotheautomationofstraightforward,repetitivetasks—representingearly-stageefficienciesratherthanlong-termtransformationalimpact.Furthermore,theunderlyingdatamaybesubjecttobias,asitisderivedfromalimitednumberofearly

adopters,manyofwhomoperateinenvironmentsthatarealreadyconducivetoAIintegration.Thesamplesizeremainssmall,andlarge-scaledeploymentsarestill

relativelyrare,whichlimitsthegeneralizabilityofthesefindings.Assuch,whiletheinitialoutcomesarepositive,furthervalidationthroughbroaderandmorediverse

implementationsisnecessarytoestablishconsistentandscalableimpact.

AFinTechorganizationimplementedanerrorpatterndetectionagentthatidentifieda23%spikeinpayment-processingerrors.Theagentnotonlyflaggedtheissuebutalsohighlightedspecificproblematiccodeblocks,reducingdebuggingtimefrom12hourstoundertwohoursperincident,cuttingoverallerrorratesby47%inthreemonths.6

YUMBrands,theparentcompanyofTacoBelland

operatorof60,000restaurantsworldwide,hasintroducedanAI-poweredrestaurantmanagerthatcantrackcrew

attendanceandplanshiftpatterns,aswellassuggest

adjustedopeninghourstoalignwithmarketconditions,andevenattendthedrive-throughwindow.Whilenotyetmarket-ready,YUMBrands,theworld'slargestfranchiseoperator,evidentlyisanillustrationofagenticAIpotentialintheindustry.7

17

AIimpactboostsinvestment

Asignificantmajorityoforganizationssurveyed,around62%,haveincreasedtheirinvestmentinGenAI,year

onyear.Amongthese,36%haveallocatedadditional

investmentcapitaltoGenAI.Thisshiftalsoreflectsa

strategicreallocationoffunds,with33%oforganizationsdivertingbudgetfromotherareas.

EvenamongorganizationswhoseleadershiparenotstrongadvocatesofGenAI,60%haveincreasedtheirinvestments.

GenAIisincreasinglyseenasastrategicinvestmenttofuture-prooforganizationsagainsttechnologicalandmarketdisruptions.

%

oforganizationswithlimitedleadershipsupporthavestillincreasedtheirGenAIinvestments

CapgeminiResearchInstitute2025

AIinaction

Figure4.

Around62%oforganizationssurveyedhaveincreasedinvestmentinGenAI

Year-on-yearchangeinbusinessinvestmentinGenAI,2025

62%

36%Additionalbudget

36%

32%

Reallocationofotherbudgets

32%

1%

Mixofboth

0%

IncreasedRemained

thesame

DecreasedUnsure/don'tknow

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=1,607executives.

CapgeminiResearchInstitute2025

18

AIinaction

Acrossvariousindustriessurveyed,thefivesectors

showingthehighestyear-on-yearriseininvestment

inGenAIareconsumerproducts(73%ofexecutives),insurance(70%),banking(67%),aerospaceanddefense(65%),andtelecom(64%).

%

ofconsumerproductsorganizationshave

increasedtheirGenAIinvestmentscomparedwithlastyear

FiguΓe5.

Nearlythree-quartersofconsumerproductsorganizationshaveincreasedtheirGenAIinvestmentscomparedtolastyear

PercentageoforganizationsinindustriessurveyedwhoincreasedtheirGenAIinvestmentscomparedtolastyear

73%

70%67%66%

ConsumerproductsInsuranceBanking

Industrialmanufacturing

65%64%

63%62%

Aerospaceanddefense TelecomAutomotiveAverage

59%

59%

EnergyandutilitiesHightech

57%

56%

55%

PharmaandhealthcareRetail

42%

Government/publicsectorMediaandentertainment

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=1,607executives.

CapgeminiResearchInstitute2025

19

%

organizationsthatincreasedtheirGenAI

investmentsinsupplychainandprocurementafterachievingsignificantcostsavings

AIinaction

FiguΓe6.

Amongorganizationsthathaveachievedsignificantcostsavingsintheirbusinessoperations,63%haveincreasedGenAIinvestments

Question1:Whatpercentageoftotaloperationscostwasreducedduetothefollowingusecases:

pilot,partiallyimplemented,fullyimplemented?

Question2:HowhasyourinvestmentlevelinGenAIchangedthisyearcomparedtolastyear?

Average

Supplychainandprocurement

Peopleoperations

Customeroperations

Financeandaccounting

63%

75%

67%

60%

50%

%organizationsexperiencingorexpecting>20%operatingcostreductionandwhoareincreasinginvestments.

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=545executiveswhoarefrombusinessfunctionsandhaveexperienced20%ormorecostreduction.

CapgeminiResearchInstitute2025

20

Amongorganizationsthathaverealized/expectmorethan20%operatingcostreductionintheirbusinessfunctions,mostincreasedtheirGenAIinvestmentsfrom2024.

Mostinvestments

willbeinproprietarymodels

Despitetheincreasingperformanceandcostadvantagesofopen-sourceAImodels,asignificantmajorityof

executivescontinuetofavorproprietarysolutionsforAIimplementation.Accordingtooursurveydata,threeoutoffourexecutivespreferproprietarymodels,with43%optingforthosedevelopedbyhyperscalersandanotherthirdchoosingmodelsfromspecializednicheproviders.ThispreferenceisparticularlystrongamongorganizationsthathavescaleduptheirinvestmentsinAIandgenerativeAI,indicatingacleartrendtowardtrusted,enterprise-

gradesolutionsthatofferrobustsupport,security,andintegrationcapabilities.

AIinaction

Figure7.

ThreeinfourexecutivessurveyedpreferproprietarymodelsforAIimplementation

PercentageofexecutiveswhoprefervariousAImodels

Proprietarymodelsfromnichemodeldevelopers

Proprietarymodelsfromhyperscalers

Open-sourcemodels

Combinationofproprietaryandopen-sourcemodels

6%

17%

34%

77%

43%

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=1,607executives.

CapgeminiResearchInstitute2025

21

Lookingahead,theadoptionofindustry-specificAIis

expectedtoaccelerate.By2027,morethan50%ofGenAImodelsdeployedbyorganizationswillbetailoredtospecificindustriesorbusinessfunctions—upfromjust1%in2023.Thisshiftunderscoresthegrowingdemandfordomain-specificintelligenceandperformance,areaswhereproprietarymodelsareoftenbetterpositionedtodelivervalue.8

Notably,organizationsthathaveincreasedinvestmentinAIandGenAIshowastrongerpreferencefor

proprietarymodels.

Overthepastyear,AIsystemshavecontinuedto

improve,exceedinghumanperformanceonseveral

benchmarks.9AccordingtoStanford’sAIIndexReport2024,theskillsgapbetweenthetopand10th-rankedAImodelsontheChatbotArenaLeaderboardwas11.9%.Byearly2025,thisgaphadnarrowedto5.4%.Similarly,thedifferencebetweenthetoptwomodelsshrank

from4.9%in2023tojust0.7%in2024.10

AIinaction

FiguΓe8.

Topfactorsdrivingpreferenceforproprietarymodels

Question:Whatfactorsmakeyouselectproprietarymodels(e.g.,MicrosoftCopilot,OpenAIGPT-4,GoogleGemini,AnthropicClaude,etc.)?

75%72%

66%

65%

Advancedsecurityfeaturesandcompliancewith

Accesstodedicated

supportandregular

updates

HighperformanceEasyintegrationwith

enterprisesystems

industrystandards

Source:CapgeminiResearchInstitute,AI-poweredbusinessoperationssurvey,February–March2025,N=1339executiveswhopreferproprietarymodelsoracombinationofproprietaryandopen-sourcemodelsforAIimplementation.

CapgeminiResearchInstitute2025

22

Twoadditionalfactorsaffectingdecisionsonmodelselection–beyondcapabilitiesandoutputreliability–aredecreasinginferencecostsandtheavailabilityofmodeloptimizationtechniques.

Inferencecosts,ortheexpenseofqueryinga

trainedmodel,arefallingdramatically(seeFigure9).Thischart,onalogarithmicscale,illustratesthetrendinAIperformanceperdollar.GPT3.5experiencedadecreasefrom$20permilliontokensto$0.07per

milliontokens,whileGPT-4hadareductionfrom$15to$0.12inayear.,,

Modeloptimizationtechniquessuchasmodel

pruning,quantization,anddistillationhelpreducethesizeandcomplexityofAImodelswithoutsignificantlycompromisingperformance.Theseoptimizedmodelsrequirefewercomputationalresources,thereby

loweringinferencecosts.Inaddition,efficient

hardwareutilization,batchprocessingofinferencerequests,dynamicscalingtoadjustthenumberof

computingresourcesbasedoncurrentdemand,andenergy-efficientalgorithmscansignificantlyreducethepowerconsumptionofAImodels.

AIinaction

Figure9.

AIinferencecostshaverapidlydeclined

Inferencepriceacrossselectbenchmark,2022-24

Inferenceprice(inUSDpermilliontokens-logscale)

10

1

0.1

oGPT-3.5level+inmultitask

languageunderstanding(MMLU)

GPT-4level+incodegeneration(HumanEval)

oGPT-4olevel+inPhD-levelsciencequestions(GPQADiamond)

GPT-4olevel+inLMSYSChatbotArenaElo

Sep-2022Jan-2023May-2023Sep-2023Jan-2024May-2024Sep-2024Publicationdate

Source:EpochAI,“ArtificialAnalysis,2025”.

CapgeminiResearchInstitute2025

23

AIinaction

Byachievingan11xreductionincomputecosts

withoutcompromisingperformance,open-source

modelssuchasDeepSeekaddressasignificant

bottleneckinAIdevelopment:accesstoadvanced

hardwareresources.Morebusinesses,research

institutions,andsmallerstartupscannowdeploy

high-qualityAImodelstailoredtotheirneeds.12

However,enterpriseadoptionofopen-sourcemodelsstillinvolvescertaintrade-offsduetothevarying

levelsofrisksandimplicationsforbusinessand

technology(seeFigure10).Theseincludetheneed

“TheadoptionofAIusecases,acceleratedbyadvancementsin

open-sourcetechnologies,availabilityofcloudAIservicesand

infrastructure,andincreasedaccessibilityinenterprisesystems,

empowersorganizationstoinnovaterapidlyandachievemeasurable

businessoutcomes.Adoptionisnolongeroptionalorsize-dependent–

it’sfundamentaltocompetitiveadvantageandoperationalproductivity."

forgreatertechnicalexpertise,potentialexposuretosecurityvulnerabilities,andrelianceoncommunity-drivensupport,whichmayaffectupdatecyclesanddocumentationquality.Whilethesechallengesare

notuniversalandarebeingactivelyaddressedby

theopen-sourcecommunity,theyremainimportantconsiderationsfororganizationsevaluatingAI

deploymentstrategies.

Despitetheseadvancements,fewerthanoneinfiveexecutivescurrentlypreferopen-sourceplatforms.

MarekSowa

HeadofGenerativeTechnologiesCenterofExcellence,Capgemini'sBusinessServices

Concernsaroundsecurity,technicalcomplexity,andtheneedforongoingcustomizationandmaintenancecontinuetodriveorganizationstowardproprietary

models.Asperformanceconvergesandcostsdecline,proprietarysolutionsremainthestrategicchoiceforenterprisesseekingscalable,secure,andspecializedAIcapabilities.

CapgeminiResearchInstitute2025

24

AIinaction

11x

claimedreductioninAIcomputecostsforopen-sourcemodelssuchasDeepSeek,without

compromisingperformance

AsFigure10shows,thechoicebetweenopen-sourceandproprietaryAImodelsisincreasinglyshapedby

aspectrumofmodelopenness,eachwithvarying

degreesoftransparency,control,andrisk.While

fullyopenmodelsofferunmatchedflexibilityand

auditability,theyalsointroduceconcernsaround

dataleakageandcompetitiveexposure.