麥肯錫-什么是人工智能？What is AI artificial intelligence

April2024

Mcsey

&company

AIbyMckinsey

QuantumBlack

McKinseyExplainers

WhatisAI(artificialintelligence)?

Artificialintelligenceisamachine’sabilitytoperformsomecognitivefunctionsweusuallyassociatewithhumanminds.

Humansandmachines:amatchmadein

productivity

heaven.Ourspecieswouldn’thave

gottenveryfarwithoutourmechanizedworkhorses.Fromthewheelthatrevolutionizedagricultureto

thescrewthatheldtogetherincreasinglycomplexconstructionprojectstotherobot-enabled

assemblylinesoftoday,machineshavemadelifeasweknowitpossible.Andyet,despitetheir

seeminglyendlessutility,humanshavelongfearedmachines—morespecifically,thepossibilitythat

machinesmightsomeday

acquirehumanintelligence

andstrikeoutontheirown.

Butwetendtoviewthepossibilityofsentient

machineswithfascinationaswellasfear.This

curiosityhashelpedturnsciencefictionintoactualscience.Twentieth-centurytheoreticians,like

computerscientistandmathematicianAlanTuring,envisionedafuturewheremachinescould

performfunctionsfasterthanhumans.Thework

ofTuringandotherssoonmadethisareality.

Personalcalculatorsbecamewidelyavailableinthe1970s,andby2016,theUScensusshowedthat

89percentofAmericanhouseholds

hadacomputer.Machines—smartmachinesatthat—arenowjust

anordinarypartofourlivesandculture.

Thosesmartmachinesarealsogettingfasterandmorecomplex.Somecomputershavenowcrossedthe

exascale

threshold,meaningtheycanperformasmanycalculationsinasinglesecondasan

individualcouldin

31,688,765,000years

.Andbeyondcomputation,whichmachineshavelongbeenfasteratthanwehave,computersandotherdevicesare

nowacquiringskillsandperceptionthatwereonceuniquetohumansandafewotherspecies.

AIisamachine’sabilitytoperformthecognitive

functionsweassociatewithhumanminds,suchasperceiving,reasoning,learning,interactingwiththeenvironment,problem-solving,andevenexercisingcreativity.You’veprobablyinteractedwithAIevenifyoudon’trealizeit—voiceassistantslikeSiriand

AlexaarefoundedonAItechnology,asaresomecustomerservicechatbotsthatpopuptohelp

younavigatewebsites.

AppliedAI

—simply,artificialintelligenceappliedtoreal-worldproblems—hasseriousimplicationsfor

thebusinessworld.Byusingartificialintelligence,

companieshavethepotentialtomakebusiness

moreefficientandprofitable.Butultimately,the

valueofAIisn’tinthesystemsthemselves.Rather,it’sinhowcompaniesusethesesystemstoassist

humans—andtheirabilityto

explain

toshareholdersandthepublicwhatthesesystemsdo—inaway

thatbuildstrustandconfidence.

FormoreaboutAI,itshistory,itsfuture,andhowtoapplyitinbusiness,readon.

Learnmoreabout

QuantumBlack,AIbyMcKinsey.

Beyondcomputation,computersand

otherdevicesarenowacquiringskills

andperceptionthatwereonceuniquetohumansandafewotherspecies.

WhatisAI(artificialintelligence)?2

Whatismachinelearning?

Machinelearningisaformofartificialintelligence

thatcanadapttoawiderangeofinputs,including

largesetsofhistoricaldata,synthesizeddata,or

humaninputs.(Somemachinelearningalgorithmsarespecializedintrainingthemselvestodetect

patterns;thisiscalleddeeplearning.SeeExhibit1.)Thesealgorithmscandetectpatternsandlearn

howtomakepredictionsandrecommendationsbyprocessingdata,ratherthanbyreceivingexplicit

programminginstruction.Somealgorithmscanalsoadaptinresponsetonewdataandexperiences

toimproveovertime.

Thevolumeandcomplexityofdatathatisnowbeing

generated,toovastforhumanstoprocessand

applyefficiently,hasincreasedthepotentialof

machinelearning,aswellastheneedforit.Intheyearssinceitswidespreaddeployment,which

beganinthe1970s,machinelearninghashadanimpactonanumberofindustries,including

achievementsin

medical-imaginganalysis

andhigh-resolutionweatherforecasting.

Whatisdeeplearning?

Deeplearningisamoreadvancedversionofmachinelearningthatisparticularlyadept

atprocessingawiderrangeofdataresources(textaswellasunstructureddataincludingimages),

requiresevenlesshumanintervention,andcanoftenproducemoreaccurateresultsthan

traditionalmachinelearning.Deeplearninguses

neuralnetworks—basedonthe

waysneurons

interactinthehumanbrain

—toingestdataand

processitthroughmultipleneuronlayersthat

recognizeincreasinglycomplexfeaturesofthedata.Forexample,anearlylayermightrecognize

somethingasbeinginaspecificshape;buildingonthisknowledge,alaterlayermightbeableto

identifytheshapeasastopsign.Similartomachinelearning,deeplearningusesiterationtoself-correctandimproveitspredictioncapabilities.Forexample,onceit“l(fā)earns”whatastopsignlookslike,itcan

recognizeastopsigninanewimage.

Learnmoreabout

QuantumBlack,AIbyMcKinsey.

Exhibit1

Artiicialintelligenceisamachine’sabilitytoperformsomecognitivefunctionsweusuallyassociatewithhumanminds.

Theevolutionofartiicialintelligence

Artiicialintelligence

Thescienceand

engineeringof

makingintelligent

machines

AIisthebroadieldofdevelopingmachinesthatcanreplicate

humanbehavior,

includingtasksrelatedtoperceiving,

reasoning,learning,andproblem-solving.

Machinelearning

Amajor

breakthrough

inachievingAI

Machinelearning

algorithmsdetect

patternsinlarge

datasetsandlearntomakepredictionsbyprocessingdata,ratherthanby

receivingexplicit

programming

instructions.

Deeplearning

Anadvanced

branchofmachine

learning

Deeplearningusesneuralnetworks,inspiredbythe

waysneuronsinteractinthehumanbrain,toingestdataandprocessitthrough

multipleiterationsthatlearnincreasinglycomplex

featuresofthedataand

makeincreasingly

sophisticatedpredictions.

GenerativeAI

Anadvancedbranch

ofdeeplearning

GenerativeAIisabranchofdeeplearningthatuses

exceptionallylargeneural

networkscalledlarge

languagemodels(with

hundredsofbillionsofneurons)

thatcanlearnespecially

abstractpatterns.Language

modelsappliedtointerpretandcreatetext,video,images,and

dataareknownasgenerativeAI.

McKinsey&Company

WhatisAI(artificialintelligence)?3

Thevolumeandcomplexityofdatathatisnowbeinggenerated,toovastfor

humanstoprocessandapplyefficiently,hasincreasedthepotentialofmachinelearning,aswellastheneedforit.

WhatisgenerativeAI?

GenerativeAI

(genAI)isanAImodelthatgeneratescontentinresponsetoaprompt.It’sclearthat

generativeAItoolslikeChatGPTandDALL-E(atoolforAI-generatedart)havethepotentialtochange

how

arangeofjobs

areperformed.Muchisstill

unknownaboutgenAI’spotential,butthereare

somequestionswecananswer—likehowgenAI

modelsarebuilt,whatkindsofproblemstheyare

bestsuitedtosolve,andhowtheyfitintothebroadercategoryofAIandmachinelearning.

FormoreongenerativeAIandhowitstandsto

affectbusinessandsociety,checkoutourExplainer“

WhatisgenerativeAI?

”

WhatisthehistoryofAI?

Theterm“artificialintelligence”was

coinedin1956

bycomputerscientistJohnMcCarthyforaworkshopatDartmouth.Buthewasn’tthefirsttowriteabout

theconceptswenowdescribeasAI.AlanTuring

introducedtheconceptofthe“

imitationgame

”ina1950paper.That’sthetestofamachine’sability

toexhibitintelligentbehavior,nowknownasthe

“Turingtest.”Hebelievedresearchersshouldfocusonareasthatdon’trequiretoomuchsensingandaction,thingslikegamesandlanguagetranslation.Researchcommunitiesdedicatedtoconcepts

likecomputervision,naturallanguageunderstanding,andneuralnetworksare,inmanycases,several

decadesold.

MITphysicistRodneyBrooks

shared

detailsonthefourpreviousstagesofAI:

—SymbolicAI(1956).SymbolicAIisalsoknownasclassicalAI,orevenGOFAI(goodold-fashionedAI).Thekeyconcepthereistheuseofsymbolsandlogicalreasoningtosolveproblems.For

example,weknow

aGermanshepherdisadog

,whichisamammal;allmammalsarewarm-

blooded;therefore,aGermanshepherdshouldbewarm-blooded.

ThemainproblemwithsymbolicAIisthathumansstillneedtomanuallyencodetheirknowledge

oftheworldintothesymbolicAIsystem,ratherthanallowingittoobserveandencode

relationshipsonitsown.Asaresult,symbolicAIsystemsstrugglewithsituationsinvolving

real-worldcomplexity.Theyalsolacktheabilitytolearnfromlargeamountsofdata.

SymbolicAIwasthedominantparadigmofAIresearchuntilthelate1980s.

—Neuralnetworks(1954,1969,1986,2012).

Neuralnetworksarethetechnologybehind

therecentexplosivegrowthofgenAI.Looselymodelingthe

waysneuronsinteractinthe

humanbrain

,neuralnetworksingestdataandprocessitthroughmultipleiterationsthatlearnincreasinglycomplexfeaturesofthedata.Theneuralnetworkcanthenmakedeterminations

WhatisAI(artificialintelligence)?4

aboutthedata,learnwhetheradeterminationiscorrect,andusewhatithaslearnedtomake

determinationsaboutnewdata.Forexample,onceit“l(fā)earns”whatanobjectlookslike,itcanrecognizetheobjectinanewimage.

Neuralnetworkswerefirstproposedin1943

inanacademicpaperbyneurophysiologist

WarrenMcCullochandlogicianWalterPitts.

Decadeslater,in1969,twoMITresearchers

mathematicallydemonstratedthatneural

networkscouldperformonlyverybasictasks.In1986,therewasanotherreversal,when

computerscientistandcognitivepsychologistGeoffreyHintonandcolleaguessolvedthe

neuralnetworkproblempresentedbytheMITresearchers.Inthe1990s,computerscientistYannLeCunmademajoradvancementsin

neuralnetworks’useincomputervision,whileJürgenSchmidhuberadvancedtheapplication

ofrecurrentneuralnetworksasusedinlanguageprocessing.

In2012,Hintonandtwoofhisstudents

highlightedthepowerofdeeplearning.They

appliedHinton’salgorithmtoneuralnetworks

withmanymorelayersthanwastypical,

sparkinganewfocusondeepneuralnetworks.

ThesehavebeenthemainAIapproachesof

recentyears.

—Traditionalrobotics(1968).Duringthefirstfew

decadesofAI,researchersbuiltrobotstoadvance

research.Somerobotsweremobile,moving

aroundonwheels,whileotherswerefixed,with

articulatedarms.Robotsusedtheearliest

attemptsatcomputervisiontoidentifyand

navigatethroughtheirenvironmentsorto

understandthegeometryofobjectsand

maneuverthem.Thiscouldincludemoving

aroundblocksofvariousshapesandcolors.

Mostoftheserobots,justliketheonesthathave

beenusedinfactoriesfordecades,relyon

highlycontrolledenvironmentswiththoroughly

scriptedbehaviorsthattheyperformrepeatedly.

Casestudy:VistraandtheMartinLakePowerPlant

Vistraisalargepowerproducerinthe

UnitedStates,operatingplantsin12stateswithacapacitytopowernearly20millionhomes.Vistrahascommittedtoachievingnet-zeroemissionsby2050.Insupport

ofthisgoal,aswellastoimproveoverall

efficiency,

QuantumBlack,AIbyMcKinsey

workedwithVistratobuildanddeploy

anAI-poweredheatrateoptimizer(HRO)atoneofitsplants.

“Heatrate”isameasureofthethermal

efficiencyoftheplant;inotherwords,it’s

theamountoffuelrequiredtoproduce

eachunitofelectricity.Toreachtheoptimalheatrate,plantoperatorscontinuously

monitorandtunehundredsofvariables,suchassteamtemperatures,pressures,oxygenlevels,andfanspeeds.

VistraandaMcKinseyteam,includingdatascientistsandmachinelearningengineers,builtamultilayeredneuralnetworkmodel.Themodelcombedthroughtwoyears’

worthofdataattheplantandlearned

whichcombinationoffactorswouldattain

themostefficientheatrateatanypoint

intime.Whenthemodelswereaccurateto

99percentorhigherandrunthrougha

rigoroussetofreal-worldtests,theteam

convertedthemintoanAI-poweredenginethatgeneratesrecommendationsevery

30minutesforoperatorstoimprovethe

plant’sheatrateefficiency.Oneseasonedoperationsmanageratthecompany’s

plantinOdessa,Texas,said,“Thereare

thingsthattookme20yearstolearnaboutthesepowerplants.Thismodellearnedtheminanafternoon.”

Overall,theAI-poweredHROhelpedVistraachievethefollowing:

—approximately1.6millionmetrictonsofcarbonabatedannually

—67powergeneratorsoptimized

—$60millionsavedinaboutayear

ReadmoreabouttheVistrastory

here

.

WhatisAI(artificialintelligence)?5

TheyhavenotcontributedsignificantlytotheadvancementofAIitself.

Buttraditionalroboticsdidhavesignificant

impactinonearea,throughaprocesscalled

“simultaneouslocalizationandmapping”(SLAM).SLAMalgorithmshelpedcontributetoself-

drivingcarsandareusedinconsumerproductslikevacuumcleaningrobotsandquadcopter

drones.Today,thisworkhasevolvedinto

behavior-basedrobotics,alsoreferredtoashaptictechnologybecauseitrespondsto

humantouch.

—Behavior-basedrobotics(1985).Inthereal

world,therearen’talwaysclearinstructionsfornavigation,decisionmaking,orproblem-solving.Insects,researchersobserved,navigatevery

well(andareevolutionarilyverysuccessful)withfewneurons.Behavior-basedrobotics

researcherstookinspirationfromthis,lookingforwaysrobotscouldsolveproblemswith

partialknowledgeandconflictinginstructions.Thesebehavior-basedrobotsareembedded

withneuralnetworks.

Learnmoreabout

QuantumBlack,AIbyMcKinsey.

Whatisartificialgeneralintelligence?

Theterm“artificialgeneralintelligence”(AGI)wascoinedtodescribeAIsystemsthatpossess

capabilitiescomparabletothoseofahuman

.Intheory,AGIcouldsomedayreplicatehuman-like

cognitiveabilitiesincludingreasoning,problem-solving,perception,learning,andlanguage

comprehension.Butlet’snotgetaheadofourselves:thekeywordhereis“someday.”Mostresearchers

andacademicsbelievewearedecadesawayfromrealizingAGI;someevenpredictwewon’tsee

AGIthiscentury,orever.RodneyBrooks,anMIT

roboticistandcofounderofiRobot,doesn’tbelieveAGIwillarriveuntil

theyear2300

.

ThetimingofAGI’semergencemaybeuncertain.Butwhenitdoesemerge—anditlikelywill—

it’sgoingtobeaverybigdeal,ineveryaspectof

ourlives.Executivesshouldbeginworkingto

understandthepathtomachinesachievinghuman-levelintelligencenowandmakingthetransitiontoamoreautomatedworld.

FormoreonAGI,includingthefourpreviousattemptsatAGI,readour

Explainer.

WhatisnarrowAI?

NarrowAIistheapplicationofAItechniquestoa

specificandwell-definedproblem,suchaschatbotslikeChatGPT,algorithmsthatspotfraudincredit

cardtransactions,andnatural-language-processingenginesthatquicklyprocessthousandsoflegal

documents.MostcurrentAIapplicationsfallinto

thecategoryofnarrowAI.AGIis,bycontrast,AIthat’sintelligentenoughtoperformabroadrangeoftasks.

Learnmoreabout

QuantumBlack,AIbyMcKinsey.

HowistheuseofAIexpanding?

AIisabigstoryforallkindsofbusinesses,butsomecompaniesareclearlymoving

aheadofthepack

.

OurstateofAIin2022surveyshowedthatadoptionofAImodelshasmorethandoubledsince2017—

andinvestmenthasincreasedapace.What’smore,thespecificareasinwhichcompaniesseevalue

fromAIhaveevolved,frommanufacturingandrisktothefollowing:

—marketingandsales

—productandservicedevelopment

—strategyandcorporatefinance

Onegroupofcompaniesispullingaheadofits

competitors.Leadersoftheseorganizations

consistentlymakelargerinvestmentsinAI,leveluptheirpracticestoscalefaster,andhireandupskill

thebestAItalent.Morespecifically,theylinkAI

strategytobusinessoutcomesand“

industrialize

”AIoperationsbydesigningmodulardataarchitecturethatcanquicklyaccommodatenewapplications.

WhatisAI(artificialintelligence)?6

WhatarethelimitationsofAI

models?Howcanthesepotentiallybeovercome?

WehaveyettoseethelongtaileffectofgenAI

models.Thismeanstherearesomeinherentrisksinvolvedinusingthem—bothknownandunknown.

TheoutputsgenAImodelsproducemayoften

soundextremelyconvincing.Thisisbydesign.Butsometimestheinformationtheygenerateisjust

plainwrong.Worse,sometimesit’sbiased(becauseit’sbuiltonthegender,racial,andotherbiasesof

theinternetandsocietymoregenerally).

Itcanalsobemanipulatedtoenableunethicalor

criminalactivity.SincegenAImodelsburstontothescene,organizationshavebecomeawareofuserstryingto“jailbreak”themodels—thatmeanstryingtogetthemtobreaktheirownrulesanddeliver

biased,harmful,misleading,orevenillegalcontent.

GenAIorganizationsarerespondingtothisthreatintwoways:foronething,they’recollecting

feedbackfromusersoninappropriatecontent.They’realsocombingthroughtheirdatabases,

identifyingpromptsthatledtoinappropriatecontent,

andtrainingthemodelagainstthesetypesofgenerations.

Butawarenessandevenactiondon’tguaranteethatharmfulcontentwon’tslipthedragnet.

OrganizationsthatrelyongenAImodelsshouldbeawareofthereputationalandlegalrisks

involvedinunintentionallypublishingbiased,offensive,orcopyrightedcontent.

Theseriskscanbemitigated,however,inafewways.“Wheneveryouuseamodel,”saysMcKinseypartnerMarieElHoyek,“youneedtobeableto

counter

biases

andinstructitnottouseinappropriateor

flawedsources,orthingsyoudon’ttrust.”How?Foronething,it’scrucialtocarefullyselecttheinitial

datausedtotrainthesemodelstoavoidincluding

toxicorbiasedcontent.Next,ratherthanemployinganoff-the-shelfgenAImodel,organizations

couldconsiderusingsmaller,specializedmodels.

Organizationswithmoreresourcescouldalso

customizeageneralmodelbasedontheirowndatatofittheirneedsandminimizebiases.

It’salsoimportanttokeepahumanintheloop(thatis,tomakesurearealhumancheckstheoutput

ofagenAImodelbeforeitispublishedorused)andavoidusinggenAImodelsforcriticaldecisions,

suchasthoseinvolvingsignificantresourcesorhumanwelfare.

Itcan’tbeemphasizedenoughthatthisisanewfield.Thelandscapeofrisksandopportunitiesislikely

tocontinuetochangerapidlyinthecomingyears.AsgenAIbecomesincreasinglyincorporated

intobusiness,society,andourpersonallives,wecanalsoexpectanewregulatoryclimatetotake

shape.Asorganizationsexperiment—andcreatevalue—withthesetools,leaderswilldowelltokeepafingeronthepulseofregulationandrisk.

Learnmoreabout

QuantumBlack,AIbyMcKinsey.

WhatistheAIBillofRights?

TheBlueprintforanAIBillofRights,preparedby

theUSgovernmentin2022,providesaframeworkforhowgovernment,technologycompanies,and

citizenscancollectivelyensuremoreaccountable

AI.AsAIhasbecomemoreubiquitous,

concerns

havesurfaced

aboutapotentiallackoftransparencysurroundingthefunctioningofgenAIsystems,thedatausedtotrainthem,issuesofbiasandfairness,potentialintellectualpropertyinfringements,

privacyviolations,andmore.TheBlueprintcomprisesfiveprinciplesthat

theWhiteHousesays

should

“guidethedesign,use,anddeploymentofautomatedsystemstoprotect[users]intheageofartificial

intelligence.”Theyareasfollows:

—Therighttosafeandeffectivesystems.Systemsshouldundergopredeploymenttesting,risk

identificationandmitigation,andongoing

monitoringtodemonstratethattheyareadheringtotheirintendeduse.

—Protectionsagainstdiscriminationbyalgorithms.Algorithmicdiscriminationiswhenautomated

systemscontributetounjustifieddifferent

treatmentofpeoplebasedontheirrace,color,ethnicity,sex,religion,age,andmore.

WhatisAI(artificialintelligence)?7

—Protectionsagainstabusivedatapractices,viabuilt-insafeguards.Usersshouldalsohave

agencyoverhowtheirdataisused.

—Therighttoknowthatanautomatedsystemisbeingused,andaclearexplanationofhow

andwhyitcontributestooutcomesthataffecttheuser.

—Therighttooptout,andaccesstoahumanwhocanquicklyconsiderandfixproblems.

Atpresent,morethan60countriesorblocshave

nationalstrategiesgoverningtheresponsible

useofAI(Exhibit2).TheseincludeBrazil,China,theEuropeanUnion,Singapore,SouthKorea,and

theUnitedStates.Theapproachestakenvaryfromguidelines-basedapproaches,suchasthe

BlueprintforanAIBillofRightsintheUnitedStates,

tocomprehensiveAIregulationsthatalignwith

existingdataprotectionandcybersecurity

regulations,suchastheEU’sAIAct,duein2024.

Therearealsocollaborativeeffortsbetween

countriestosetoutstandardsforAIuse.TheUS–EUTradeandTechnologyCouncilisworking

towardgreateralignmentbetweenEuropeandtheUnitedStates.TheGlobalPartnershiponArtificialIntelligence,formedin2020,has29members

includingBrazil,Canada,Japan,theUnitedStates,andseveralEuropeancountries.

EventhoughAIregulationsarestillbeingdeveloped,organizationsshouldactnowtoavoidlegal,

reputational,organizational,andfinancialrisks.Inanenvironmentofpublicconcern,amisstep

couldbecostly.Herearefourno-regrets,preemptiveactionsorganizationscanimplementtoday:

—Transparency.Createaninventoryofmodels,classifyingtheminaccordancewith

regulation,andrecordallusageacrosstheorganizationthatiscleartothoseinside

andoutsidetheorganization.

—Governance.ImplementagovernancestructureforAIandgenAIthatensuressufficient

oversight,authority,andaccountabilityboth

withintheorganizationandwiththird

partiesandregulators.

—Data,model,andtechnologymanagement.

?Datamanagement.Properdata

managementincludesawarenessofdatasources,dataclassification,data

qualityandlineage,intellectualproperty,andprivacymanagement.

?Modelmanagement.OrganizationsshouldestablishprinciplesandguardrailsforAI

developmentandusethemtoensureallAImodelsupholdfairnessandbiascontrols.

Exhibit2

RegulationsrelatedtoAIgovernancevaryaroundtheworld.

AsofNovember2023,nonexhaustive

Typeofpolicy:

Nonbindingprinciples(eg,OECD)

Japan

Singapore

UnitedArabEmirates

UnitedKingdom

.UnitedStates

Source:OECD;McKinseyanalysis

GeneralAIlegislationproposedorbeinginalized

●Brazil

Canada

China

SouthKorea

EuropeanUnion

Examplecountrieswithoutgeneral

AIlegislation

Australia

India

.NewZealand

SaudiArabia

McKinsey&Company

WhatisAI(artificialintelligence)?8

?Cybersecurityandtechnologymanagement.Establishstrongcybersecurityand

technologytoensureasecureenvironmentwhereunauthorizedaccessormisuse

isprevented.

—Individualrights.MakeusersawarewhentheyareinteractingwithanAIsystem,andprovideclearinstructionsforuse.

Howcanorganizationsscaleup

theirAIeffortsfromadhocprojectstofullintegration?

MostorganizationsaredippingatoeintotheAI

pool—notcannonballing.Slowprogresstowardwidespreadadoptionislikelyduetocultural

andorganizationalbarriers.Butleaderswho

effectivelybreakdownthesebarrierswillbebestplacedtocapturetheopportunitiesoftheAIera.

And—crucially—companiesthatcan’ttakefull

advantageofAIarealreadybeing

sidelined

bythosethatcan,inindustrieslikeautomanufacturing

andfinancialservices.

ToscaleupAI,organizationscanmake

three

majorshifts

:

1.Movefromsiloedworktointerdisciplinary

collaboration.AIprojectsshouldn’tbelimitedtodiscretepocketsoforganizations.Rather,

AIhasthebiggestimpactwhenit’semployedbycross-functionalteamswithamixofskills

andperspectives,enablingAItoaddressbroadbusinesspriorities.

2.Empowerfrontlinedata-based

decisionmaking

.AIhasthepotentialtoenablefaster,better

decisionsatalllevelsofanorganization.Butforthistowork,peopleatalllevelsneedtotrustthealgorithms’suggestionsandfeelempoweredto

makedecisions.(Equally,peopleshouldbeabletooverridethealgorithmormakesuggestionsforimprovementwhennecessary.)

3.Adoptandbolsteran

agile

mindset.Theagiletest-and-learnmindsetwillhelpreframe

mistakesassourcesofdiscovery,allayingthefearoffailureandspeedingupdevelopment.

Learnmoreabout

QuantumBlack,AIbyMcKinsey,

andcheckout

AI-relatedjobopportunities

ifyou’reinterestedinworkingatMcKinsey.

Articlesreferenced:

—

“AsgenAIadvances,regulators—andrisk

functions—rushtokeeppace

,”December21,2023,AndreasKremer,

AngelaLuget

,

Daniel

Mikkelsen

,

HenningSoller

,MalinStrandell-Jansson,andSheilaZingg

—“

WhatisgenerativeAI?

,”January19,2023

—“

Techhighlightsfrom2022—ineightcharts

,”December22,2022

—“

GenerativeAIishere:HowtoolslikeChatGPT

couldchangeyourbusiness

,”December20,2022,

MichaelChui

,

RogerRoberts

,and

LareinaYee

—“

ThestateofAIin2022—andahalfdecadein

review

,”Decembe