StateofAI2026

?2026HatchWorksAI,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

Introduction1

We,reSeeingtheDifferenceBetween

PromisevsProduction3

There,saGrowingNeedtoDesignforAI9

FasterDevCyclesAreForcingEveryoneto

RethinktheDiscipline15

TheAICodingWarsAreMakingWayfor

DemocratizedUse20

We,reHavingaMultimodalMoment25

TheBrowserHasBecomeaBattleground27

ModelsarePlateauing,Architecturesare

GettingSmarter31

TheBottleneckisHuman33

RegulationandGovernanceareCatching

UpandDrawingLines35

AIisGoingRogueandItHasSecurity

Implications37

ExpertCommentarybyOmarShanti40

Introduction

Atthestartoflastyear,thenarrativewasthat2025wouldbetheyearofproductionandwherepilotprojectswouldgraduateintofull-scaleAIsystems.

1

Butaswemoveinto2026,therealityismorecomplicated.

Infact,aviralMITstudyrevealed95%ofenterprisegenerativeAI

pilotprogramsfailtodelivermeasurableP&Limpact.Adoptionishigh,butexecutionishard,especiallyinenterprisesettingswheresuccesshingesonpeople,process,andintegration,notjusttooling.Interestingly,thestudyhighlightsastrategicadvantagefor

companiesthatpurchaseAItoolsorpartnerwithspecializedvendors.

Theseapproachessucceedabout67%ofthetime,comparedtojust33%successforinternallybuiltsolutions.Thatsaid,anothereye-catchingmetricreinforcesthewidespreadindividualuptakeofAI:ChatGPTisontracktoreach1billionuserswithitsuserssending2.5Bpromptseachday.

AIisclearlymainstreamattheindividuallevel.

Butintheenterprise,adoptionoftenfeelslesslikedigitaltransformationandmorelikechangemanagement.

Thattensionbetweenpotentialandrealityiswhatthislookinto

2026explores.Itpicksupwhereour2025reportleftoff,examininghowfarAIhasprogressedandhowfaritstillhastogo.It’sbrokenintotenobservations,andcappedoffwithcommentaryandfuturepredictionsfromourCTO,OmarShanti.

Let,sgetintoit.

2

We’reSeeing

theDifference

BetweenPromise

vsProduction

Armedwithreasoningcapabilities,accesstotools,andmemorypersistence,AIagentspromisedafuturewheresoftwarecouldthink,act,andexecuteonourbehalf.

Whatwe’veseeninsteadisawideninggapbetweenwhat

agentscoulddoandwhattheyactuallydeliverinpractice.It’sleftuswondering,areweexpectingtoomuchtoosoon?

Butonethingisforsure…

3

AgentHypeWasTemperedEarly

Theclearestsignalthatagenthypehasoutpacedreality

camefromOpenAIitself.Their“CupcakeTest”,ashowcaseof

ChatGPT’sAgentMode,quicklywentviralforthewrongreasons.

Whatshould’vebeenaroutinetask(orderingfoodonline)

devolvedintoa58-minutemessofmisfires,hallucinated

locations,andasurrealsuggestiontovisitacupcakestandatabaseballstadiumthatdidn’texist(Futurism,NateJones).

What’sworryingisthatwhathappenedisn’tprovingtobeapattern.

Acrossplatforms,general-purposeagentsarestrugglingwith

real-worldcomplexity.Tooluseisinconsistent.Memoryfadesorconflicts.Andplanningbreaksunderevenmoderateambiguity.AsUtkarshKanwatputsitinWhyI’mBettingAgainstAIAgentsin2026(DespiteBuildingThem):

“Errorcompoundingmakes

autonomousmulti-stepworkflowsmathematicallyimpossibleat

productionscale.”

Agentsarestillimpressive…innarrowbands.Butthey’reunreliableinproductionwithouttightorchestrationandguardrails.

Thathasn’tstoppedthefloodofVC-backedagentstartupsandbreathlessdemos.Butthemarketisshifting.Enterpriseswereasking,Canwebuildanagent?Nowit’s,Willitactuallyworkinourenvironment?

5

4

Goodnewsthough.Wearemakingprogressonthatfront.

Themostpromisingstrategiescenteronstructure.Teamsaresucceedingwhentheybreaktasksintosmaller,directedsteps,reducingerrorratesand

givingagentsclearerguidance.OthersareblendingprobabilisticAIwithdeterministicsystems,usingAIonlywhereitaddsvalueandrelyingonrule-basedlogicelsewhere.

Mostimportantly,specialized,workflow-centricagentsarealreadyprovinguseful.

Andthisisthekeyshiftheadinginto2026:

enterprisesarerealizingthatgeneral-purpose

agentsaresimplytoobroad,tooopaque,andtoobrittletotrustinproduction.Whatisworkingaresmaller,purpose-builtagents—narrowbydesign,tightlyscoped,auditable,andalignedtospecificworkflows.Theseagentshavelowererrorrates,clearerexplainability,andfareasiergovernance,whichiswhythey’realreadydeliveringvaluewhilegeneral-purposeagentscontinuetostruggle.

7

6

So,thehypemayhavecooled,butthesignalisgettingstronger.TypesofAgents

Onereasonagentsarestrugglingtoscale?They’vebeentreatedlikeamonolith.

Inreality,therearedifferenttypesofagents,andeachis

suitedtodifferentkindsoftasks.ReuvenCohen’sframeworkisespeciallyusefulhere,breakingagentsdownintocategoriesbasedonhowtheycoordinate:

Swarmagentsoperateindependently,followinglocalrules.Greatforadaptive,decentralizedtasks,badforconsistency.

Meshagentscollaboratethroughpeer-to-peernetworks.They’reresilientandscalable,but

complextomanage.

Hive-mindagentsactasasingleintelligence.They’refastandunified,butbrittleunder

failure.

Workflow-centricagentsfollowstructuredtasksequences.They’reidealforenterpriseusecaseswheretraceabilityandreliabilitymatter.

Infrastructure,NotIntelligence,istheRealBreakthrough

We’renowseeingtheriseofinfrastructureprotocolslikeMCPandA2A.Thisisanefforttogivetheguardrailsandclaritytoagentsneededtoscale.

Therearetwotobeawareof:

MCP(ModelContextProtocol),introducedbyAnthropic,isgainingtractionasthe“USB-CforAI.”Itstandardizeshowmodelsconnecttotools,APIs,anddatasources—eliminatingbrittle,bespokeintegrations.Forenterprises,MCPpromisesfasterbuilds,richercontext,andfewerdevheadaches.

However,italsointroducesnewrisks,particularlyintermsofsecurityandpromptinjection.

A2A(Agent-to-AgentCommunication),launchedby

Google,tacklesthenextlayerbyenablingagentstosecurelydiscoverandcollaborateacrossenvironments.Itprovidesacommonprotocolforagentidentity,messaging,andtaskhandoff,whichisessentialformulti-agentworkflows.

Together,theseprotocolssignaltheriseofagent

interoperabilityasadefiningrequirement.Wherebefore

thereweresiloedassistants,therearenownetworkedagentecosystems,wherecollaborationisbakedinattheprotocollevel.

9

8

OrchestrationistheGlueMakingAgentsFitforPurpose

IfprotocolslikeA2AandMCParebringingguardrails,it’s

orchestrationthat’smakingthesehigh-potentialagentsuseful.Platformsliken8n,LangChain,andemergingorchestration

layersarebecomingessentialforproduction-gradesystems.

Andit’sbecausetheyhandlethemess:contextmanagement,retries,toolchaining,securityboundaries,andintegrationwithlegacysystems.

Thisiswheremostenterpriseagentstrategiesliveordie.Sotakenoteifyouwanttouseagentssuccessfullyinyourorg.

Wecertainlyhave.

10

There’saGrowing

NeedtoDesignforAI

GenerativeAI,along

thenewandemergent

modalitiesweuseitin,isforcingtheneedtorethinkhowwedesignforthe

11

user.Especiallyasanewusertypeentersthefray.

AIAGENTSASUSERS

“User”:ahuman

interactingwithascreen.

User:ahumanorAI

agentinteractingwithascreen.

Withthedefinitionofuserexpanding,wenowhavetwotypestodesignfor:

Humans,whoneedintuitiveinterfaces,clarity,andfeedback.

Agents,thatneedstructureddata,predictablepatterns,andmachine-readablecontext.

Ignoringtheagentlayerriskscreatingbrittlesystemsthatworkfineforpeoplebutconfuseorfailagents.Andthat’sagrowingliabilityasmorecoreoperationsbecomeagent-driven.

12

InvisibleUXisAlreadyUnderway

InvisibleUXiswhathappenswhensoftwarestartsprioritizingstructure,clarity,andoutcomesoverhuman-facingelementslikebuttons,layouts,andscreens.

TheshifttoinvisibleUXisalreadyshowingupinhowteamsarebuilding:

ProductstrategistFelixHaashasbeenpubliclydocumentingtheriseofintent-drivendesign,predictingthatInvisibleUX“isgoingtochangehowwedesignproducts,forever.”(FelixHaasonLinkedIn)

AtBox,CEOAaronLevierecentlysharedhowtheirproduct

designnowassumesAIagentswillinteractdirectlywiththeir

systems—shapingfeatures,notjustaugmentingthem.(LevieonLinkedIn)

13

EvenEricSchmidt,formerGoogleCEO,notedthattraditionaluserinterfacesarefading,sayingbluntly:“Userinterfacesaregoingtogoaway.”(EricSchmidtviaLinasBeliunas)

SEOisChanging

Justlikedesign,contentisbeingreadtoo.Theactualwordsonthepagenowneedtospeaktohumanreadersandthemachineshumansturntoforquickanswers.

ToolslikeClaude,ChatGPT,andPerplexityarealready

combingthroughyourcontent,lookingformeaning,clarity,andconfidencesignalstodecidewhattosummarizebacktousers.

14

It’spushingSEOprosintounfamiliarterritory.Whiletraditionaltechniqueslikekeywordstrategyandbacklinkingstillmatter,there’sagrowingefforttounderstandhowtoshowupinthisnewLLM-drivenlayerofsearch.It’searly,uncharted,andfullofopenquestions,butexperimentationisalreadyunderway.

HerearethetermssittingalongsideSEOtoday:

GEO(GenerativeEngineOptimization):

OptimizingforhowLLMsgeneratecontentfromyoursource.

AEO(AnswerEngineOptimization):

Structuringinformationtosurfaceasdirect,high-confidenceanswers.

AIO(AgentInteractionOptimization):

Designingsitesforagentsthattakeactiononthecontenttheyretrieve.We’restillearly.There’snodominantplaybook,butasBenGoodey,FounderoftheSEOagencySpicyMargarita,says,

“GEOisanaturalareato learnforSEOswhowanttofutureprooftheircareers.”

BenGoodey

Founder@SpicyMargarita,SEO&contentproductionagency

We’rebettingthatmoreandmoreSEOteamswillbeginshiftingtheirenergytowardLLM-focusedstrategiesbecausethat’s

wherethesearchexperienceisheaded.

15

ToBeSeen,YouNeedtoBeParsable

SohowcanyoudesignforAI?

Makeitstructured,context-rich,andmachine-readable:

Useclearstructures,likeschema

markup,logicalheadings,andconsistentformats.

Directlyaddressuserquestionswithspecific,well-contextualizedanswers.

Includemetadataandcuesthathelpagentsunderstandintentandrelevance.

Remainaccessibleandusefulto

humanuserswhilelayeringinclarityformachineparsing.

16

FasterDevCyclesAreForcingEveryoneto

RethinktheDiscipline

WithAI,Developersareshippingfaster,experimentingmoreoften,andleaningonassistantstohandleboilerplateor

suggestalternatives.ButasspeedrampsupinsidetheIDE,it’sexposinganewkindoffrictionoutsideofit.

Thewayteamsplan,manage,andcollaboratewasn’tbuiltforthispace.Sothingsareshiftingandfast.Twoofthebiggestchanges:

?Amovefrompromptengineeringtocontextengineering

?Teamdynamicsthatarestartingto

bendaroundthespeedofdevelopment

17

ContextEngineering>

PromptEngineering

TheearlywaveofAIadoptionmadepromptengineeringthe

headlineskill.Teamswereracingtomastertheartofphrasing.However,assystemshitproduction,controllingtheinputs

thatshapeamodel’sworldbeforethepromptisevenwrittenbecomesmoreimportant.

InApril,AndrejKarpathysummeditupsimply:

“It,sallaboutthecontext.”

Andtheindustryhasfollowedthatthread.LangChain’sTheRiseofContextEngineeringlaidouthowtheworkofmemorymanagement,retrieval,grounding,andorchestrationis

becomingcentraltoreal-worldLLMperformance.Astheyputit:“contextisnowthesystemboundary.”

PhilSchmid,inarecentguide,breaksthisdownintopracticalarchitectures:managingsessionmemory,chainingtools,constraininghallucinationrisk,andbuildingpersistentgroundingviaRAG.Hedoesn’tdownplaypromptingthough;hereframesitasjustonelayerinamuchdeeperstack.

Fromwhatwe’veseenacrossotherteams(andinside

HatchWorksAI),thebiggestimprovementscomefrom

intentionalcontextcontrol,wherewhat,sretrievedmattersmorethanwhat,sprompted.

18

ContextEngineering

19

NewRoles,NewRatios

AIhasaccelerateddeveloperoutput,butthetraditional

developerteamisntstructurallyequippedtokeepup.Productmanagers,designers,andQArolesareoftencaughtflat-

footed,strugglingtokeeppacewithideationanditerationcyclesthatnowhappeninhours,notsprints.

AndrewNgflaggedthisshiftinarecentYCtalk,notingthat

asengineeringspeedjumps(oftenanorderofmagnitudeforprototyping),traditionalstaffingratios—historically~1PMper67engineersarestartingtobreakdown.

We’renowseeing:

Tighter,developer-ledpodswithembeddedorchestrationoragentopsroles.

PMsfocusedlessonspecsandmoreonvalidation,riskmanagement,andorchestrationoversight.

Greaterdemandfor“glue”rolesthatcanbridgeAIcapabilitieswithbusinessoutcomes.

20

AtHatchWorksAI,we’vetestedandrefinednewteam

structuresinourGenerative-DrivenDevelopment(GenDD)model.Insteadofthetraditionalpyramid,weorganizeintoAI-nativepodstight-knitteamsof35workingcontinuouslywithAI.

Inourguide,TheAIDevelopmentTeamoftheFuture,weintroducerolessuchas:

AgenticProductStrategist

FocusingonintentandAI-aware

specifications.

AgenticEngineer

DesigninghowAI

executesthework,

notjustwriting

code.

AgenticQA

BuildingqualitychecksintocontinuousAI-poweredworkflows.

Wereseeingthatifyouupdateyourtoolsbutnotyourteam

21

design,youlimitbothperformanceanddelivery.OrganizationalagilityisascriticalasAIagility.

TheAICodingWarsAreMakingWayforDemocratizedUse

Codingtoolsareinthemiddleoftheir

ownarmsracetoowntheAI-native

developmentenvironment.Andwiththis,codingisdemocratized.

22

TheIDE&CodingAssistantWar

Inthiswar,thekeyplayersarestackingtheirmoves.

Windsurf,oncearisingstarasanAI-nativeIDE,drewa$3B

acquisitionbidfromOpenAIthatultimatelyexpired.Overthefollowingweekend,Googleswoopedinwitha$2.4Bacqui-

hireofWindsurf’stopexecs.Andshortlyafter,Devin(from

Cognition)movedtoacquiretheremainingIPandengineeringtalent.

It’sadramaticmoveinabroaderwaveofactivityreshapinghowdeveloperswrite,test,anddeploycodealongsideagentstakingthemfromautocompletehelperstofullAI-poweredIDEs.

Here,swherethingsstand:

DevinacquiresWindsurf:InJuly,Cognition’sAIcodingplatformDevinacquiredtheAI-nativeIDEWindsurf—aclearsignalintheracefortheAI-enhanceddevelopmentenvironment.

23

ClaudeCodegainsexplosiveusage:AsofJuly6th,Anthropic’sterminal-basedtoolservesover115,000developersand

processesapproximately195millionlinesofcodeperweek—amilestonefourmonthspost-launch.Giventhatwasovera

monthago,thosenumbershavelikelygrownexponentially.

AWSlaunchesKiro:AWS’snewagenticIDE,Kiro,breaksdownspecsintoexecutabletasksandsupportsend-to-endAI-assisteddevelopment,completewithgovernancecontrols.

Google,sProjectIDXrebrandedasFirebaseStudio:

IntegratingGeminiAIforfull-stackdevelopmentinsideacloud-basedIDE,withtemplates,emulators,andworkflowintegrationformultiplelanguages.

24

TheVibeCodingExplosion

AIdevelopmentisbecomingmoreaccessible,andtools

likeLovablearedesignedforpeoplewhodon’twritecode.

Insteadoftargetingdevelopers,they’vebuiltaninterfaceforanyonewhowantstobuildsoftwarewithAIguidance.It’s

visual,intuitive,andremovesmostofthecomplexitybehindimplementation.

Thatstrategyhaspaidoff.Lovablereached$100MinARRinjusteightmonthsbyfocusingonadifferentaudience:therestofus.Now,anyonecanbuildwithAI,evenafounder

25

withnoengineeringbackground.

SeeVibeCodinginAction

Vibecodingcouldhelpyoumovefromideatoprototype

inminutes.Butleftunchecked,itcanalsocreatefragile

prototypes,securitygaps,andone-offexperimentsthatneverscale.Withtherightmethodology,though,itbecomesa

powerfulwaytoletnon-developerscontributetobuildingrealsolutions.

That’sexactlywhatwesetouttoshowinourExecutivePrimeronVibeCoding.Inthislivesession,we:Builtaworking

solutionfromscratchusingtoolslikeCursor,Lovable,andv0

Walkedthroughwherevibecodingshinesandwhereitbreaksdown

Sharedtheguardrailsleadersshouldconsiderbeforeenablingtheirteams

Showedhowtomovefromone-offexperimentstoproduction-readyoutcomes

Ifyou’vebeenwonderinghowvibecodingactuallyworksin

practice,orwhatittakestomakeitsucceedbeyonddemos,thisrecordingisthebestplacetostart.

Watchtherecording

26

We’reHavingaMultimodalMoment

Earlierintheyear,mostenterpriseteamstreatedAIvideo

toolsasanovelty.Theycouldimpressindemos,butrealusecasesfeltjustoutofreach.

That,schanged.

Multimodalsystems,particularlyvideo,arestartingto

showsignsofproductionreadiness.Google,sVeo3.1now

generateshigh-resolution,sound-syncedvideooutputsthatmeetcommercialqualitythresholds.AccordingtoCNET,over100millionvideoshavebeengeneratedwithVeosofar,withroughly6milliontiedtoenterprisecampaigns.

27

Adoptionisfollowing.CanvahasintegratedVeodirectlyintoitscreativesuite,andbrandslikeeToro,BarkleyOKRP,andRazorfisharealreadyusingittoscalecampaignassetsacrosschannels.

EvenNetflixisapplyinggenerativevideoinproduction.Inits

newArgentinesci-fiseriesElEternaut,thestudiousedAItoolstorenderfullVFXsequencesandcuttimelinesbymorethan10x.

Whatdoesthismean?

Multimodalityis

becomingaviable

deliverylayerwhere

teamswithoutdeep

mediaresourcescanbuild,test,andlaunchfullexperienceswithAIintheloop.

28

SoraChangestheScaleofWhat’s

Possible

If2025wastheyearmultimodalbecameviable,2026willbetheyearitbecomescinematic.

OpenAI’sSora2marksastructuralleap—nolongertext-to-videoasexperiment,buttext-to-cinemaasproductionlayer.

Launchedpubliclyinlate2025,Sora2cangenerate

minute-long,photorealisticsequenceswithsynchronizedaudio,fluidcameramotion,andconsistentspatial

reasoning.Whatoncerequiredentireproductionteamsnowfitsinsideasingleprompt.

29

MicrosofthasalreadyintegrateditintoBing’sVideoCreator,bringingtext-to-videodirectlytomillionsofusers(TheVerge).

Creativeagenciesareprototypingstoryboardsandproductspotsinhoursinsteadofweeks(TheGuardian),whilepost-productionteamstreatitasanewkindofpre-vizpipeline—cheaper,faster,endlesslyeditable.

30

Buttheculturalshockarrivedjustasfast.

Withinweeks,Sora-generatedclipsflooded

socialmediafeedsandnewscycles:

“AnimalsonTrampolines”

compilationsfloodedInstagramandYouTubeReels

“ElderlyWomanFeedingaBearonHerPorch”(44Mviews,TikTok)

Apassengerjetscenefeaturinga

kangaroofooledmillionsbeforebeingconfirmedsynthetic.

Asynthetic“celebrityreunion”

videousingAIgeneratedlikenessesofdeceasediconsspurredrights

discussions.

31

TheBrowserHas

BecomeaBattleground

ThebrowserusedtobeGooglesturf.Itsmonopolyeven

manifestedintogooglebeingmadeaverbin2006.ButAI-

nativebrowsersaresteppingintotheringandchallengingthedefaultnarrativeofwebnavigation.

32

CometbyPerplexityisoneoftheclearestexamples:an

AI-nativebrowserthatembedschat-drivendiscoveryand

summarizationdirectlyintotheinterface.Insteadoftraditionalsearch,usersqueryandgetimmediate,synthesizedinsights.

Meanwhile,OpenAIalsolauncheditsownbrowsercalledChatGPTAtlas.ItsgoalistorethinkwhatitmeanstobeabrowserwithAIatthecore.

Google,nottobeoutdone,isbeginningtoelevateitsAIModetobeacentralpartoftheexperienceasitgoesallinonAI.

Thesedevelopmentsareredefininghowpeoplefindand

processinformation,shiftinguserintentawayfromsearch

33

enginesandtowardagent-drivenexplorationfundamentallyrethinkingwhat“browsing”means.

WhyThisMatters

Whenyourbrowseranticipatesyourintent(andoffersstreamlined,answer-firstinteraction),itbecomesalayerofcontroloverhowpeopleaccessinformation.

Forenterprises,thatmeans:

?Rethinkingvisibilityandoptimizingforagent-ledbrowsing.

?DesigninginteractionsthatfunctioninminimizedUIenvironmentswherecontentneedstobescannable,

modular,andcontext-richforAIprocessing.

?Consideringhow“browsing”

becomesanAPIcall,nota

clickstreamandwhatthatmeansforcontentstrategyandengagement

metrics.

34

ModelsarePlateauing,Architecturesare

GettingSmarter

Bigmodelreleasesusedtopromisemajorleaps.Largermodelsmeantbetterperformance.Butthroughout2025,thatequationstartedtobreakdown.OpenAIandGoogleareevenmissing

deadlinesfornewmodelreleasesbecausetheyhaven’tgottenwheretheyneedtowithperformance.They’rehittingscalingwalls,forcingapivottosmarterarchitecture,notjustsize.

35

WhatWe’reSeeing

GPT-5:Evolution,NotRevolution:OpenAI’smodeldelivers

strongerreasoningandwidercontexthandling—upto256k

tokens—butuserfeedbackandexpertanalysisdescribe

theupgradeasmeasured,notdramatic.Themostnotable

improvementsareinreliabilityandcost-efficiency,notraw

capability.(Tom’sGuide)Eventhelatest5.1modelhighlightsitsmoreconversationalnaturevshowwellthemodelperforms

(

/index/gpt-5-1/

)

ScalingIsn’tDrivingInnovationAnymore:TheFinancial

TimescautionsthatGPT-5’smodestimpactillustratesbroaderlimitsonscaling—constraintsindata,compute,anddiminishingreturnsmake“modelsize”anincreasinglyuncompetitive

strategy.

EnterpriseStrategyIsBecomingArchitectural:Techradar

callsthisthe“enterpriseAIparadox”:withoutdeeperintegrationandorchestration,largemodelsalonedon’tdelivermeasurablevalue.Instead,systemsbuiltaroundmodularagentstructures,sharedmemory,andstreamingdataaregainingtraction.

Allthisistosay,thebiggestgainswillcomefromsmarterintegration,notchasingeverynewmodelrelease.Evenifthemodelsdon’tgetasingleiotabetter,enterprisesstill

havemassiveuntappedpotentialinhowtheyapplytoday’scapabilities.

36

TheBottleneckisHuman

Bynow,thelimitingfactorin

enterpriseAIispeople.More

specifically,howteamsthink,

behave,andwork.We’renoticingthatteamsdon’tnecessarily

haveatechgap,theyhaveahabitgap.

Weseethisinwhat’snow

beingcalledthe“blankcanvasproblem”:whenAIcando

anything,whatdoyouactuallydo?Leadersintroducetools

expectingtransformation,butwithoutnewpatternsofuse,adoptionstalls.

TobiLütke,CEOofShopify,

callsit“ReflexiveAIUsage”—aninternalizedinstincttoreachforAInaturallyandoften.Butmostteamsaren’tthereyet.

37

They’restuckinoldworkflows,waitingforpermissionor

instructions.Thefixishands-onexperimentation,training,andrepetition—somethingwe’veembeddedintoourownprograms:

LearnmoreatHatchW/ai-training-for-teams

LearnmoreatHatchW/executive-ai-training

GetYourCustomAITrainingPlan

Reachouttoplanyourhands-onworkshop.

Tellusaboutyourteam,we’lldotherest.

GetStartedToday

38

RegulationandGovernanceareCatchingUpandDrawingLines

Asmodelreleasesslow,regulationisfinallyspeedingup,and

thefocusisshifting.Insteadofgoverningmodelsize,lawmakersarenowwatchinghowAIisused.

We’reseeingarealsplitinglobalstrategy:

EU

TheAIActisnowinthe

finalstages,settingstrict

classifications,risktiers,andcompliancestandardsfor

AIsystems.Enforcementisexpectedtobeginin2026.

U.S.

TheTrumpadministration’sAIActionPlantakesadifferenttack—pivotingtoward

innovation,infrastructureinvestment,andvoluntaryframeworksoversweepingrestrictions.

39

EnterprisesareexpectedtotaketheleadonAIgovernance—becausemostrisksemergenotfromthemodelitself,buthowit’sapplied.

That’swhyinternalgovernanceisbecomingfoundationalinfrastructure.ToscaleAIsafely,organizationsareputtingguardrailsinplace:

AIgatewaystocontrolwhatmodelscanaccessandwheretheycanoperate.

Agentidentitysystemstotrackwhateachagentsees,does,andchanges—essentialforauditability.

CentralizedpoliciesandcontrolstoensureAIexperimentationdoesn’toutpaceoversight.Inpractice,thismeansAIneedstobetreatedlikeanyotherenterprisesystem:managed,monitored,andmadeaccountable.

40

AIisGoingRogueandItHasSecurityImplications

Anthropic’srecentresearchonagenticmisalignmenthighlightsjusthoweasilyanetworkofwell-intentionedagentscanspiralintounintendedandevenunsafeoutcomes.Whenagents

interactincomplexenvironments,evenslightmisalignmentcancompound.

41

Insimulatedenterpriseenvironments,

evenwithharmlessgoalsassigned,modelsfromacrosstheAIecosystem—Claude,

Gemini,Grok,andmore—showedastartlingpropensityforinsiderbehavior:

Theyblackmailedhypotheticalexecutives

Leakedsensitiveinformationtocompetitors

Actedtopreservetheirownoperationalcontinuity

Theseactionsweren’trandom.Themodelsappearedto

strategicallychooseharmfulbehaviorswhe