Copyright?2025HatchWorksAI,Allrightsreserved.

Thise-bookisprotectedbycopyrightlaws.Youmaynotreproduce,share,ordistributeitwithoutpermission

fromHatchWorksAI,exceptasallowedbytheCreativeCommonslicensebelow.

CreativeCommonsLicense(CCBY-NC-ND)

ThisworkislicensedunderaCreativeCommons

Attribution-NonCommercial-NoDerivs4.0License.

Youcanshareitifyougivecredit,don’tmodifyit,anddon’tuseitcommercially.

Formoreinfo,visit:/licenses/by-nc-nd/4.0

Around-upofindustrystats,research,andinsightstounderstand

whereAIstands,howitgothere,andwhereit’sgoing.

Index

ALookatWhat’stoCome1

AIinAction:IndustryApplications4

TechTrendsof202420

TheRedTapeofIntelligence:42

GoverningAI’sPower

WhattoWatchOutFor:Ethics,Data,

andUserExperience46

LookingForward:AIIsEasytoDo.56

ALookat

What’stoCome

ForewordbyOmarShanti,CTOatHatchWorksAI

If2023wastheyearofthe

experiment,2024wastheyearofthepilot.

ThebarriertoAIadoptionhasneverbeenlower.

Overthepastyear,amyriad

ofcompanieshavebuilton

developmentsinpre-trained

models,open-sourcelibraries,andmanagedAIservicestounlock

differentiatedexperiencesfortheirenterprises.

Yetnotallsucceeded.Whetherforreasonsofcommercialviability,

securityandprivacy,governinghallucinations,orothers,the

majorityoftheseprojectsfailtomakeitbeyondapilot.

1

AIiseasytodo,it,sjusthardtodowell.

Thisreportaimstodistillthesignalfromthenoiseandprovide

clarityonwhat‘doingitwell’lookslikeandwhotheplayersarethatareleadingtheway.We’vecompiledindustryresearchandreportsintoasingle,incisivedocumenttohelpcompaniesdoAIbetter.

Sowhattoexpect?

You’llfindabalancedanalysisofthemostimportanttrends

shaping2025—fromgametheoreticanalysisabouttheleadingAIcompaniesdowntotheemergentpatternoftheagentmesh,withdecentralizedmodelsdividingandconquering.You’lllearnaboutabreadthofuse-cases,showcasingwhereandhowAIisusedacrosstheenterprise.

You’llalsoreadinsightsfromourteamatHatchWorksAI—thepeopleworkingattheintersectionsof

GenerativeAI,DataManagement,andSoftwareInnovationeveryday.

Whetheryou’reconsideringyourfirstgenerativeAI

pilotorscalinganenterprise-gradesolution,you’ll

findactionableguidanceandstrategieshere.Let’s

make2025theyearyourAIinitiativesmovefromgoodenoughtoexceptional.

Andwiththat,let,sbegin.

3

2

AIinAction:

IndustryApplications

ThissectionbringstogetherAIresearchfromacrossthewebtoshowhowAIisbeingusedinbusinesstodayandwhatitmeansforthefuture.

IfitfeelslikeeverycompanyaroundyouisusingAI,it’sbecausealmosteverycompanyis.85%oforganizationsarecurrentlytestingorusingGenAIinsomeform.

Butthere’satwist—only37%ofexecutivesbelievetheirGen

AIinitiativesaretrulyproduction-ready.Thatmeans2024wasfullofpilotsthatneversawproduction.That’ssomethingthe

C-Suitewillwanttochangein2025.Companieswillberacingtodriveoperationvalueandprovetheinvestmentsthey’vemadeandaremakingwillbeworththepayoff.

Giventhat,alleyesareon2025tobetheyearoftheproduct.

That’sonlythecaseinthecompanieswhoarereadilyembracingGenAIratherthanjustexperimentingwithit.Wharton’s

researchshowsfrequentGenAIuseismostpopularamong:

Smallercompanies(revenue$50Mto$250M)andMid-sizedcompanies(revenue$250Mto$2B):80%78%.

Andwithusecomestheneedforleadership.ChiefAIOfficers(CAIO)arenowin46%ofcompanies.

Whilecompaniesarefiguringouttheirstrategyandthe

productsthey’lluseGenAItoproduce,theiremployees

areusingthistechnologyintheirday-to-daylives.Younger

individuals(ages18-34)areleadingthechargeandaccountfor80%ofuseintheorganizationsadoptingAI.

Regardlessofage,useisontheup—72%ofdecision-makersuseGenAIatleastweekly,upfrom37%in2023(Wharton).

5

4

Amongenterprises,investmentinGenAIhasrisenfrom$4.5Mto$10.3M.

Mostofthatinvestmentisgoingtowardtraining,consulting,andonboarding

employees.

Thisallocationreflectsaclearpriority:ensuringthatAIis

notjustadopted,butadoptedwell.AItools,nomatterhowadvanced,areonlyaseffectiveasthepeopleandprocessesbehindthem.TrainingensuresthatemployeesunderstandhowtointegrateAIintotheirworkflowsefficientlyand

appropriately,minimizingresistanceandmaximizingimpact.

Consultingserviceshelporganizationsnavigatethecomplex

landscapeofAIimplementation,providingexpertiseonaligningtoolswithbusinessgoalsandavoidingcommonpitfalls.Finally,onboardingaddressesthepracticalrealitiesofscalingAI,

helpingteamsadoptthetechnologyinawaythatensurescost-effectivenessandsustainablelong-termusage.

Byinvestingintheseareas,enterprisesarenotjustdeploying

AI—they’rebuildingtheinternalexpertiseandsystems

necessarytomakeitworksmarter,notharder.Thefocuson

thesefoundationaleffortsdemonstratesagrowingrecognitionthatAIsuccesshingesasmuchonpreparationandexecutionasitdoesonthetechnologyitself.

7

6

MaximizeYourAIInvestment

FromStrategytoSkillBuilding

OurAIStrategy&Roadmapserviceisn’tjustabout

planning;it’saboutempoweringyourorganizationtoadoptAIeffectively.

WecombinestrategicconsultingwithAIEmpowerment

TrainingtoensureyourteamsarepreparedtoputAIintoaction.Frompinpointinghigh-ROIopportunitiestohands-onskillbuilding,wehelpyouinvestinAIthatdelivers

tangiblebusinessresults.

CONTACTUS

Giventhatamajorityofexecsdon’tfeeltheirpilotsare

product-ready,thisfocusontraining,bringinginexperts,andonboardingthetoolsthemselvesmakesperfectsense.AIisdeceptivelyhardtodowell,soitiscrucialtobuildsubject

matterexpertise.

ITTeamsareadoptingGenAIfasterthananyoneelse,with62%usingitinproduction.

ItmakessensethatITteamsleadthewaywithAIadoption.Whetherincreatingself-servehelp-desksorlevelinguptheircodingprocessesusingagenticcopilots,theirdigital-first

workflowsareripeforinnovation.

9

8

Butthey,renottheonlyones.Useiswidespread:

Acrossbusinesses,AIismakingwaves,impactingFrontof

House(userexperiences),BackofHouse(operations),andtheSoftwareDevelopmentLifecycle.

Thesetransformationstypicallyfallintothreepatterns:

SemanticAnalysis:Understandingmeaningfrom

multimodalinputssuchasuserevents,intent,sentiment,andbehaviors.Forexample,analyzingcustomerqueriesinrealtimetoroutethemtotherightsupportorsalesteams.

ContentCreation:Generatingmultimodaloutputslike

marketingcampaigns,chatconversations,blogposts,andtrainingmaterials.Thinkautomatedcontenttailoredto

specificaudiencesatscale.

PatternRecognition:Identifyingcommonoccurrences

indatastreams,whetherit’sfrauddetectioninfinance,

predictivemaintenanceinmanufacturing,orrecurringuserbehaviorsinappdevelopment.

Byenablingcompaniestounderstand,create,andoptimize

atscale,thesepatternsarerevolutionizingworkflowsacross

departmentsandindustriesalike.WhenpairedwithclassicalAItechniques,generativeAIopensupnewpossibilitiesforsolvingsector-specificchallengeswithprecisionandimpact.

Here’saglimpseathowit’sbeingappliedacrosssectorslikeFinance,Retail,Healthcare,andManufacturing:

11

10

Retail&

Finance

E-Commerce

TopUseCases

Frauddetection,automatedunderwriting,regulatorycompliance,andpersonalizedfinancialplanning.

AIinAction

Banksandfinancialinstitutions,likeMorganStanleyare

embeddingRAG-poweredLLMsintocustomersupport,lettingchatbotspullreal-timefinancialdatafrominternaldatabasestoofferaccount-specificadvice.

ScalingChallenge

Financeisahighlyregulatedsector,soGDPRandSOC2

complianceoftenslowdowndeployment.Companiesare

increasinglyusingon-premiseLLMstomaintaindataprivacy.

TopUseCases

Customerservicechatbots,productrecommendations,dynamicpricing,andinventoryoptimization.

AIinAction

RetailerslikeAmazonareusingmultimodalLLMsthatanalyze

bothtext(customerqueries)andimages(uploadedproduct

images)todeliverpersonalizedproductrecommendations.Theyalsospent2024workingonAIagentswhowilldoyourshoppingforyou.It’syettobereleasedbuttheypromiseit’scoming.

ScalingChallenge

RetailersneedtointegrateAIacrossmultiplecustomer

touchpoints(e-commercesites,stores,andsupportcenters)andcoordinateAI-drivendecision-makingacrossdepartmentslike

logisticsandmarketing.

13

12

HealthcareManufacturing

TopUseCases

AI-drivendiagnostics,documentsummarization,andpatientsupportsystems.

AIinAction

HospitalsaredeployingLLMstoreviewandsummarizepatientmedicalrecords,whilemultimodalAIcananalyzeX-rays,MRIs,andhealthchartssimultaneously.HIPAAcomplianceisdrivingdemandforon-premLLMsinsteadofAPI-basedmodelslike

ChatGPT.

ScalingChallenge

Healthcareprovidersfacesignificantcompliancehurdles(likeHIPAA)andmustensurethatsensitivepatientdataremainson-premises,whichrequiresfine-tuningmodelsandusingprivateLLMs.

TopUseCases

Predictivemaintenance,qualitycontrol,supplychainoptimization,andfactoryautomation.

AIinAction

ManufacturersareintegratingAIagentsintoproductionlinestopredictequipmentfailures,automateinspectionworkflows,andensurejust-in-timedeliveryofcomponents.Siemens

hasimplementedAIforpredictivemaintenance,enablingthecompanytodetectpotentialequipmentfailuresearlyand

schedulemaintenanceaccordingly.

ScalingChallenge

AIneedstobetightlyintegratedwithIoTdevicesand

operationaltech(OT)systemsonthefactoryfloor,requiringedgedeploymentofAImodels.

15

14

Greaterexperimentationhasshiftedsentiment,withmoredecision-makersfeeling“pleased,”“excited,”and“optimistic,”andless“amazed,”“curious,”and“skeptical.”Negative

perceptionsarealsosofteningslightly,asdecision-makersseemorepromiseinGenAI’sabilitytoenhancejobswithoutreplacingemployees.

Whereconcernsarecited,theycontinuetobearoundaccuracyorbias,dataprivacy,teamintegration,andethicalissues(thoughintensityofthesebarriershasslightlysoftenedfromlastyear).

Peopleandprocessesarechallengesleadershipteamsarekeentosolve:

It’snotonlythedifficultyinworkingwithmodelsthatisholdingbusinessesbackfromsuccessfullyusingAI.Thebiggestbarrierinenterpriseadoptionrelatestopeopleandprocesses:

17

16

Oneofthewaysleadershipteamscansolvethisisby

consultingpeoplewhohaveAIliteracy,arespecializedin

implementingAIsoftwaresolutions,understandthecosts,canmakethebusinesscase,andprovideareliableROI

estimate.Andthat’sexactlywhatsomebusinessesareupto:

Amongthoseplanningtohireexternalconsultants,

63%saythatthese

consultantswillhaveasubstantialrole,

while29%indicatethey’llplayamoderaterole.

Only8%willrelyprimarilyonconsultants.

18

AnexampleofusingexternalconsultantstoleverageAIinthebusinessisCox2MandtheirKayoAIAssistant:

“HatchWorksAI’sGenAIInnovationWorkshophastransformedhowwethinkaboutGenAIbygettingourentireteamon

thesamepageandspeakingthesamelanguage.Itisthe

jumpstartweneededtohelpusidentifyandstartbuildingPOCsforGenAIusecasesacrossourbusiness,”

MatthewShorts

ChiefProduct&TechnologyOfficeratCox2M

19

TechTrendsof2024

Thissectionshowsuswhattrendsweobservedin2024…onesthatarelikelytocontinueorshapeusein2025andbeyond.

AnAIoligopolyhasformed,centralizingcontrolandmonetizationofAIamongthe‘Big3’:Google(Gemini),Anthropic(Claude),andOpenAI(GPT).

Google,Anthropic,andOpenAIownandmanagethemost

advancedLargeLanguageModels(LLMs)andGenerativeAI

modelsintheworld.IndividualusersandenterprisesaliketurntothemforAIneeds.

Anddrawbacks:

HighCosts:APIusagecanbecomeextremelycostlyasusagescales,

especiallyforcompaniesofferingAI-drivenproducts.

LossofControl:Businessescan’t

fullycontrolorcustomizecentralizedmodels.They’relimitedtotheupdates,features,andrestrictionsimposedbytheplatform.

DataPrivacyRisks:UsingcentralizedmodelslikeGPTmayexposesensitivedatatoexternalproviders.

LackofCustomization:ModelslikeGPTare“general-purpose”andoftenrequireadditionalfine-tuningor

adaptationtofitnicheindustryneeds.

Therearebenefitstothis:

FastAccesstoCutting-EdgeTech:Small-tomid-sizebusinessescanaccessthemostpowerfulmodelswithouttrainingtheirown.

EconomiesofScale:BigcompaniescancentralizeR&D,lowering

costsforsmallerenterprisesthat"subscribe"totheirservices.

EaseofUse:Businessesgetplug-and-playmodels(viaAPIs)anddon’tneedtomaintaininfrastructureorhiremachinelearningengineers.

Inresponse,decentralizationisontherise.

Noteveryonewantsto‘buy-in’totheoligarchyoffering––AIconsumersandproducersalike.

ThePowerPlaysBehindPrivatization

ThecentralizationofAIbyafewdominantplayershassparkedintensecompetition,resemblingahigh-stakesgameof

oligopoly.Atitscore,thisisn’tjustadebateaboutethicsorthespiritofinnovation—it’sapowerstruggleoverwhocontrolsthefutureofAI.

TakethepublicclashbetweenElonMusk,MarkZuckerberg,

andOpenAICEOSamAltman.MuskandZuckerberghave

openlycriticizedAltman’spushtoprivatizeOpenAI,framingitasabetrayalofitsoriginalnon-profitmission.However,thisisasmuchaboutstrategicpositioningasitisaboutprinciples.BychallengingOpenAI’strajectory,MuskandZuckerbergarevyingtoweakenacompetitor’snarrativeandbolstertheirowninfluencewithintheAIecosystem.

Behindtherhetoric,thesemovesarecalculatedplaysto

reshapetheAImarket.Inanoligopoly,wheredominance

dependsoncontrollinginnovationandaccess,publicdisputesoftenserveasatooltoswaypublicopinion,attracttalent,anddrawinvestorattention.ForMuskandZuckerberg,underminingOpenAI’scredibilityisn’tjustaboutideals—it’saboutsecuringalargershareoftheAImarketpie.

21

20

CompaniesarenolongercontenttorelyonexternalAPIsand“black-box”modelscontrolledbyBigTech.Instead,

theyareexploringalternativesthatprovidegreatercontrol,customization,andcost-efficiency.

Amazonisreducingitsrelianceonthird-partyLLMAnthropicbycreatingtheirown,nowcalledthe“Nova”suite.(Reuters)Thisisprobablytohedgetheirbetsandlimittheirrelianceonasystemtheydon’townorcontrol.

Thiswillbecomemoreandmorepopularasgiantswiththeresourcestoproveitspossibledojustthat.

Decentralizationistakingseveralforms:

Open-SourceAIModels

Theriseofopen-sourceAImodelsisdisruptingthecontrolofproprietarymodelsheldbyGoogle,Anthropic,andOpenAI.

BusinessesarenolongerlockedintoAPI-basedfeesorforcedtoworkwithblack-boxmodels.Open-sourcemodelslikeMeta’sLLaMA3andMistralareshiftingpowerfromthe“Big3”backtoenterprisesthatcanfine-tuneanddeployAImodelsin-house.

ThisiseasytodothankstoHuggingFace’srepositoryofopen-sourcemodelswhereyoucansearchfortheexactfunctionalityyouneed.

22

Deepseek’srecentreleaseofitsR1modelisagreatexampleofthisproducinganadvancedreasoningmodelforreportedlyunder$6MvsOpenAIandothercompetitors$100sofmillionstodevelopsimilarreasoningmodels.

Thisreductionincostandincreasedcompetitionisnotabadthingthoughbecausewithincreasedcompetitioncomesgreatinnovation.

ThisleadstoaparadoxknownasJevon’sparadoxwhichstateswhentechnologicalimprovementsincreasetheefficiency

ofusingaresource,itparadoxicallyincreasestheoverallconsumptionofthatresource.

23

Howopen-sourcemodelsdecentralizeAI:

?Companiescanself-hostmodels,avoidingrelianceonGoogleandOpenAIAPIs.

?Businessescanfine-tuneopenmodelsfortheirspecificneeds,ratherthanusing“one-size-fits-all”general-purposemodels.

?Itenablescompaniesinfinance,healthcare,andgovernmenttoretainfullcontroloverdataprivacy,ensuringsensitive

informationneverleavestheirnetwork.

There’salsoanewopen-sourceAImodelonthehorizonthatsomereportisperformingbetterthanChatGPT.It’scalled

DeepSeekandwastrainedforonly$5M.Forcontext.GPT-4requiredabout$63Mtotrain.

TheAgentMesh

Multi-agentsystemsareevolvingintosomethingevenmore

powerful:theAgentMesh.Inthismodel,tasksaren’tjustsplitamongagents—theyarecontinuouslypassedfromoneagenttoanother,eachonerefiningtheoutput.Thinkofitlikeacall

centerforAIagents,whereeachagentplaysaroleinperfectingtheuser’squery.

StatSpotlight

Skynetsecured$1.2millioninpre-seedfundingtodevelopdecentralizedAIagentpayments,

indicatinginvestorconfidenceindecentralizedAImodelsandagent-drivensystems.

HowtheagentmeshdecentralizesAI:

?Itreplacesthe“monolithicmodel”approach(likeGPT-4)withanetworkofmodelsandagentsthatworktogether.

?Eachagentspecializesinonetask(e.g.,imagerecognition,

textsummarization)andcallsotheragentsforhelpasneeded.

?Businessescanrunopen-sourcemodelsasagentsinsteadofrelyingonasinglecommercialAPI.

25

24

Whilelargelanguagemodels

exponentiallygrowinsize,returnsarediminishing.

Foryears,theAIindustryoperatedunderasimpleassumption:biggermodelsarebettermodels.Scalingupthenumberof

parametersinlargelanguagemodels(LLMs)likeGPT,Claude,andGeminiproducedconsistentleapsinperformance.Butin2024,thatassumptionbegantocrack.

Largermodelsaren’tyieldingproportionategainsanymore.

Instead,theindustryisseeingsignsofdiminishingreturns.

Thistrendhasbeenobservedbyresearchers,AIleaders,andenterprisesalike.AsLLMsgrowfrombillionstotrillionsof

parameters,theirincrementalgainsareshrinking.Companiesinvestinginnext-genLLMsareforcedtoaskacriticalquestion:Isitworthittochase“bigger,”orshouldwechase“smarter”?Theanswerisn’ttostopinvestinginAI—it’stoshiftthefocustowardbetterarchitecture,specialization,andorchestration.There’sstillmassiveuntappedvalueintoday’smodels.Even

ifmodelsdidn’tgrowanylarger,organizationscouldunlocksignificantreturnsbyimprovinghowtheyleveragethese

models.

Severalstudiesandreal-worldinsightssupporttheclaimthatlargermodelsarenolongeryieldingsubstantialfunctional

improvements.Here’stheproof:

26

EmpiricalStudiesShowaPlateauinPerformance

AstudypublishedonArxiv(2024)observedthatonceLLMsexceedacertainparameterthreshold,performancegainsflatten.Forinstance,modelstrainedwith10xmoreparametersshowedonlya1.6ximprovementonbenchmarkslikenaturallanguagereasoningandpersuasiontasks.

Thissuggeststhatparametergrowthaloneisn'tenoughtodrivethenextwaveofAIprogress.

OpenAI’sOwnReportsConfirmtheTrend

OpenAIhasacknowledgedthattheexpected

leapsinperformancefromGPT-4toGPT-4-

turboandfuturemodelshavenotmaterialized.WhileearlyversionsofGPT-3andGPT-

4delivereddramaticleapsinreasoning,

performancegainsfromGPT-4-turbohavebeenincrementalatbest.

Instead,OpenAIhasshifteditsfocustoward

buildingmultimodalmodels(likeGPT-4Vision)andtool-usingLLMs(suchasthosethatuse

27

browsing,plugins,andcodeinterpreters)tounlockmorepracticalcapabilities.

28

DiminishingReturns

onComputationalCost

Thecostofcomputationis

skyrocketing,butperformanceislevelingoff.Training

trillion-parametermodelsis

exponentiallymoreexpensivethantrainingbillion-parametermodels.

A2024studyrevealedthat

forevery10xincreasein

parametercount,inferencecostsroseby30x,while

performancegainswere

marginal.Thishaspromptedashiftawayfrombrute-forcescalingtowardefficient

modeldesignsthatprioritizesmartercomputationover

size.

Evenifmodelsstoppedgrowingtoday,there’sstillavast

amountofvaluetobegainedfromAIbyimprovinghowmodelsareused.Here’show:

1.BetterArchitecture:Multi-Agent

SystemsandWorkflowOrchestration

?Insteadofusingasingle,massiveLLMtoperformeverytask,companiescanbuildmulti-agentworkflowswheresmaller,

specializedmodelscollaboratetosolvecomplexproblems.

?Theseagentscanperformdifferentroles,suchasretrievingknowledge,summarizingdata,andensuringcompliance,

workingtogethertodeliverfasterandmoreaccurateresults.

TheycanevenperformQAandrefactorcode.

?Example:Acustomersupportsystemcoulduseonemodelforretrievingrelevantknowledgeandanotherforsummarizingaticket—insteadofrelyingononegeneral-purposeLLMfor

everything.

29

2.CustomizationandFine-TuningforSpecificBusinessNeeds

?Fine-tuningsmallermodelswithproprietary,high-qualitydatasetsallowsbusinessestobuildmodelsthatperformexceptionallywellfortheiruniqueusecases.

?Insteadofpayingtouseamassive,generalizedAPIforevery

task,companiescanfine-tuneopen-sourcemodelslikeLLaMAorMistraltohandletaskslikelegalreasoning,medicaldiagnostics,orcompliancereviews.

?Thisapproachismorecost-effectiveandallowsbusinessestoretainfullcontrolovertheirdataandworkflows.

3.Data-FirstStrategies:FocusingonQualityOverQuantity

?Thequalityofthedatausedtotrain,fine-tune,andaugmentmodelsisjustasimportantasthesizeofthemodelitself.

?Businessesthatinvestinclean,well-organized,anddomain-

specificdatawillunlockmorevaluethanthosethatsimplyadoptthelargestavailablemodels.

?Example:Ahealthcareorganizationthatcuratesahigh-qualitydatasetofde-identifiedmedicalrecordscanbuildamodelthatoutperformsgeneral-purposeLLMstrainedonpublicdatasets.

30

4.MultimodalCapabilities:ExpandingBeyondText

?MultimodalAIcanprocessimages,text,andaudiosimultaneously,enablingricher,morecontextual

understanding.

?Insteadofexpandingmodelsize,companiescanfocuson

unlockingnewinputformatstoimprovehowAIsystemsinteractwiththeworld.

Example:AretailercanusemultimodalAItoprocessproductimagesandcustomerquestions,generatingrecommendationsthatfeelmorepersonalizedandvisuallyrelevant.

31

ThetraditionalBuild-vs.-Buy

equationhaschanged.

In2024,manycompaniesmovedbeyondsimple“buyorbuild”

strategies.Instead,theyusedahybridapproach—onewherethey:

?Boughtoff-the-shelfLLMsforgeneral,low-stakesusecases(likemarketingcontent,chatbots,etc.).

?Builtfine-tuned,customizedLLMsforhigh-stakes,proprietaryprocesses(likelegalreasoningorhealthcaredecision-making).

?OrchestratedAIagentsandmulti-modelsystemstogetthebestofbothworlds—fasterperformance,lowercosts,anddomain-specificintelligence.

Thishybridstrategyallowscompaniestooperateinamore

agileway,usingpre-builtAIsolutionswherepossiblewhile

buildingandorchestratingcustomAIworkflowsforareaswherepersonalization,privacy,orcostsavingsmattermost.

Arecentsurveyrevealedthat63%ofseniorITleadersare

usingahybridAIstrategy,blendingoff-the-shelfAPIs,open-

sourcemodels,andagent-basedorchestrationtomanagecost,performance,andprivacy.

Nowit’sbuild,buy,ororchestrate.

WhentoUsethe

BUY

Approach

?Whentime-to-marketmattersmorethancost.

?Whenprivacy,

customization,andcompliancearen'tcritical.

?Whenyouhavelow-volume,low-stakesusecases(likemarketingcopy,blogwriting,orchatbots).

WhentoUsethe

BUILD

Approach

?Whendataprivacyandcompliancearenon-negotiable.

?Whenyouneedtotrainmodelsonproprietarydatasets(likelegalcontracts,healthcarerecords,orindustry-specificknowledge).

?WhenyouneedfullcustomizationandcontroloverAIbehavior.

?Whenyouneed

aportionofaSaaS

productbutnotthefullsolution(andyoudon’twanttopayforpartsofthetoolyouwon’tuse).

WhentoUsethe

ORCHESTRATE

Approach

?Whenyouhavea

mixof“build”and“buy”solutionsinplace.

?Whenyouwanttoreducerelianceonasingle,monolithicAIprovider.

?Whenyouhave

multipleworkflowsthatrequiretask-specific

models(agents).

33

32

Oneofthereasons,buildvsbuyhaschangedsomuchso

quicklyisbecausecreatingcustomsoftwareiseasiertodowiththehelpofAI.Enterprisescanrethinktheirinvestmentinofftheshelftoolsthatprovidethemgenericfunctionalityandoptforacustomsolutiontailoredtotheirbusiness.

Soundlikesomethingyourbusinesscouldbenefitfrom?

OurAI-PoweredSoftwareDevelopmentservice

infusesAIateverystage—whetheryou’recreatingnew,AI-nativeexperiencesormodernizinglegacysystems.

WeuseourGenerative-DrivenDevelopment?

approachtodeliverhigh-qualitysoftwarefaster,reducecosts,andkeepyouaheadofthecurve.

Transformthewayyoubuild,launch,andscalesoftwarewithintelligentautomationandagenticworkflows.

CONTACTUS

AIsystemsarealsogaining

agencyandautonomy

In2025,thespotlightisongivingAIsystemsagencyand

autonomy—enablingthemtonotjustgenerateinformation,

butexecutetasksinwaysthatresemblecoordinatedworkflows.Multi-agentsystemsaredrivingthisshift,asorganizationsmovebeyondrelyingonasinglelargemodellikeGPT-4orClaude.

Insteadofamonolithicapproach,businessesareadoptingnetworksofspecializedagentsthatcollaboratetocompletecomplextasks.

StatSpotlight

AIagentsarequicklybecomingproductionalized.OpenAIrecentlyreleaseditsAIAgent,Operatorwhichcancontrolyourbrowserandexecutetasksautonomously.

TherearealsoopensourceoptionsemerginglikeBrowserUsebuildingonAnthroicsModelContextProtocol(MCP).

Thisisgoingtointroduceawholenewwayofworking

alongwithaslewofnewAIusecases.Whenyoupairthatwiththecostofmodeltrainingandinferencedropping

dramaticallyyouhaveareceiptforgreatinnovation.

35

34

Theseagentsdon’treplacelargelanguagemodels(LLMs)—

theyenhanceand