人工智能的競(jìng)爭(zhēng)優(yōu)勢(shì)：走向情境人工智能理論

AcademyofManagementReview

COMPETITIVEADVANTAGESTHROUGHARTIFICIALINTELLIGENCE:TOWARDATHEORYOFSITUATEDAI

Journal:

AcademyofManagementReview

ManuscriptID

AMR-2020-0205-Original.R3

ManuscriptType:

OriginalManuscript

TheoreticalPerspectives:

Resourcebasedview,Knowledge-basedview,Learning,Adaptation,Routines,andKnowledgeManagement

OtherTheoreticalPerspectives:

TopicAreas:

Business-levelresources/capabilities<BusinessandCompetitive

Strategy<BusinessPolicyandStrategy,Knowledgemanagement<StrategicManagementProcess<BusinessPolicyandStrategy,

Technologyevolution<TechnologyandInnovationManagement

OtherTopicAreas:

Abstract:

Howcanfirmsestablishcompetitiveadvantagesusingartificial

intelligence(AI)?AlthoughAIisbeginningtopermeatebusiness

activities,ourunderstandingofhowAIcanbeusedtocreateuniquevalueislimited.Toaddressthisvoid,weintroducetheconceptof

situatedAIandilluminateitsimportanceforestablishingAI-drivencompetitiveadvantages.Thepaperhighlightstheorganizational

activitiesinvolvedinsituatingAI—grounding,bounding,andrecasting.Italsoexplainstheconditionsinwhichthesesituatingactivitiesbetter

enablefirmstodevelopAI-drivencapabilitiesthatarefirm-specific,cost-effective,andappropriateforopportunitiesinthestrategicenvironment.Thus,thispaperprovidesanintegrativeframeworkforconnectinga

firm’sAIpursuitstocompetitiveadvantage.

60

Page1of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

COMPETITIVEADVANTAGETHROUGHARTIFICIALINTELLIGENCE:TOWARDATHEORYOFSITUATEDAI

AyendaKemp

PamplinCollegeofBusiness

VirginiaTechUniversity

ayenda@

Acknowledgments:Iamgratefultotheassociateeditor,AllanAfuah,andtothereviewersfor

theirinsightfulguidanceandsupportthroughouttherevisionprocess.ThispaperhasbenefitedimmenselyfromhelpfulcommentsfrommycolleaguesCynthiaDevers,DeviGnyawali,RichardHunt,AbrahamOshotse,KarenSchnatterly,andMaxStallkamp.

60

AcademyofManagementReviewPage2of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

COMPETITIVEADVANTAGETHROUGHARTIFICIALINTELLIGENCE:TOWARDATHEORYOFSITUATEDAI

Abstract

Howcanfirmsestablishcompetitiveadvantagesusingartificialintelligence(AI)?AlthoughAIis

beginningtopermeatebusinessactivities,ourunderstandingofhowAIcanbeusedtocreate

uniquevalueislimited.Toaddressthisvoid,weintroducetheconceptofsituatedAIand

illuminateitsimportanceforestablishingAI-drivencompetitiveadvantages.Thepaper

highlightstheorganizationalactivitiesinvolvedinsituatingAI—grounding,bounding,and

recasting.ItalsoexplainstheconditionsinwhichthesesituatingactivitiesbetterenablefirmstodevelopAI-drivencapabilitiesthatarefirm-specific,cost-effective,andappropriatefor

opportunitiesinthestrategicenvironment.Thus,thispaperprovidesanintegrativeframeworkforconnectingafirm’sAIpursuitstocompetitiveadvantage.

60

Page3of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

TheprospectofusingAItoestablishcompetitiveadvantagespresentsatheoreticalpuzzle.

Estimatespredictthatby2033,somewherebetween40and50percentofjobswillbeautomatedusingintelligentalgorithms(Frey&Osborne,2013),reflectingenhancedproductivityandlowercosts.ItisalsopredictedthatAImayleadtonewproducts(Barro&Davenport,2019;

Davenport&Kirby,2015),byallowingfirmstoembedAIintotheirproductsandbyigniting

innovationsinafirm’sproductdevelopmentprocesses(Gregory,Henfridsson,Kaganer,&

Kyriakou,2021;Cockburn,Henderson,&Stern,2019).Despitethispromise,agrowingbodyofresearchhighlightsthatAImaypresentsubstantialstrategicobstacles.AImaybemyopic

(Balasubramanian,Ye,&Xu,2022),incapableofperceivinginterdependencieswithinafirm

(Raisch&Krakowski,2021),andrecalcitranttomanagerialcontrol(Murray,Rhymer,&

Sirmon,2021).ThesefactorssuggestthatusingAItolowercostsandcraftdesirableproducts

maynotbeassimpleaspreviouslysuggested.Inaddition,AIisaformofexplicitknowledge

(Broussard,2018;Shrestha,He,Puranam,&vonKrogh,2021),andresemblesageneral-purposetechnology(Teece,2018).SoevenwhenAIleadstovaluecreationwithinafirm,theactivitiesunderpinningtheseoutcomesmaybereplicablebyafirm’srivals.Thus,whileAIholdspromiseforpromptingcompetitiveadvantages,itisunclearhowthispromisecanberealized.

ThispaperbeginstoresolvethispuzzlebydevelopingatheoryofsituatedAI—AIwhoseagencyiscircumscribedinafirm’sexperiential,structural,andrelationalsystems.Wegroundourframeworkintheorganizationalcapabilitiesliterature,whichholdsthatcompetitive

advantagesemergeprimarilywhenfirmsdeploytheirstrategicassetsusingorganizational

capabilitiesthatareidiosyncratic(Barney,1991),inexpensivetodevelop(Winter,2000),andalignedwiththefirm’sinternalandexternalenvironment(Mahoney&Pandian,1992;Sirmon,Hitt,&Ireland,2007).WearguethatachievingtheseoutcomesismadedifficultbyAI’s

60

AcademyofManagementReviewPage4of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

propensitytoactwithagency(Murrayetal.,2021),whichmaybecounterproductivewhennotproperlycontextualizedwithinthefirm.WealsoacceptthatuncontextualizedagencymaybeAI’sbaselinestate(Balasubramanianetal.,2022).

Weaddressthestrategiclimitationsofartificialintelligencebyexplaininghowfirmsmay(1)circumscribeAI’sagencyinthefirm’suniqueexperiencesandsystemsand(2)embedthis

transformedAIinthefirm’sorganizationalcapabilitiesthroughthreesituatingactivities:

grounding,bounding,andrecasting.Groundinginvolvesorchestratingwhichexperiencesone’sAIwillbeallowedtolearnfromacrosstheorganization.Boundinginvolveseffortstoshapetheexperiencesanchoringacompetitor’sAI.Recastinginvolvesorchestratingthecontinual

adaptationofalgorithmsandtheirsurroundingroutinestoenhanceAI’salignmentwith

interdependentactivitiesinafirm.Wealsoconsiderhowtechnologicalconstraintsand

environmentaldynamisminfluencethebenefitsofsituatingAI.Thus,thispaperacknowledgesAI’sstrategiclimitationswhileexplaininghowfirmscanovercometheselimitationstobetterrealizeAI’spotentialasanewfoundationforcompetitiveadvantage.

CONCEPTUALBACKGROUNDThePromiseofAIforCompetitiveAdvantage

Artificialintelligence(AI)broadlyreferstomachinesthatcancompletecognitivetasks

previouslypossibleonlyforhumans(Davenport,2018).Whilethereisalonghistoryof

machinesdisplacinghumanworkers,theriseofAIisuniqueinthatmachines,forthefirsttime,can“l(fā)earn”andperformtheirworkwithagency(Faraj,Pachidi,&Sayegh,2018).Withprevioustechnologies,machinescompletedtheirworkbyfollowingintricateif-thenstatements

programmedbyhumanactors.Themachinehadnoagencytospeakof;itsactionswereadirectreflectionoftheknowledgeofitsprogrammers(Dreyfus&Dreyfus,2005;Norman,2017).In

60

Page5of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

contrast,withAI,thecomputerisprovidedasetofinputdata,alearningobjective,anerror

function,andamathematicalalgorithmforminimizingthaterrorfunction(Chenetal.,2020;

Alpaydin,2016).Armedwiththisbasicdescriptionofaproblem,thecomputerthenlearnsits

own“rules”forlinkingtheinputdatatothedesiredoutcomes.Whatiscriticalabouttheserulesisthattheyarenotcreatedbyhumanactorsand,inmanycases,cannotevenbeexplainedby

humans(Castelvecchi,2016).Thus,AIcanbethoughtofaspossessingadistinctformofagenticrationalitythatincreasinglyallowsmachinestoperformcognitivetasksatalevelequalingor

surpassinghumanperformance(Murrayetal.,2021).

ThepowerofAIhasledmanytobelievethatAIwillrevolutionizeeconomicproductionbymakingfirmsmoreefficientthroughintelligentautomationandbyassistinghumansin

solvingnovelproblemsthatmayleadtovaluecreationthroughthedesignofnewproductsandtheimprovementofoldones(Barro&Davenport,2019;Brynjolfsson&Mitchel,2017;Frey&Osborne,2013).Indeed,wearebeginningtoseeprocessandproductimprovementswithAI

acrossmultipleindustries(Tarafdar,Beath,&Ross,2019).Asoneexample,DBSBankrecentlyimplementedAIthatpredictswith85percentaccuracywhetheranemployeewillleavewithinthreemonths.ThefirmisnowusingAItopoweradigital-onlybankinIndiathatemploys90

percentfeweremployeesthanatraditionalbank(Davenport,2018).Asasecondexample,

fragrancedesignersnowuseAIduringproductdevelopmenttoproduceperfumesthatappealtoconsumersmorethanfragrancescreatedbyhumanexpertsalone(Goodwinetal.,2017).

Despiteimprovementsinoperationsandproductdesign,however,firms’investmentsinAImayfailtomaterializeasprofits.InarecentsurveybytheBostonConsultingGroup,nineoutoftentopmanagersreportedthatAIrepresentsalargebusinessopportunityfortheirfirms

(Ransbothametal.,2019),and43%ofexecutivesreportedhavingimplementedAIintheir

60

AcademyofManagementReviewPage6of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

organizations.Thereportalsonoted,however,that“mostcompanieshaveahardtimegeneratingvaluewithAI.”Lui,Lee,andNgai(2022)offerempiricalsupportforthisconcern,findingthatmarketspenalizeAIadoptionsatthefirmlevel.Thus,whileAIisbeginningtoleadtoproductandprocessimprovements,firm-levelbenefitsofAI,suchasimprovedmarketperformanceorcompetitiveadvantage,maybemoredifficulttoachieve.

ThreeStrategicLimitationsofAI

Whatexplainsthisdisconnect?WhileexistingtheorydoesnotexplainhowfirmscansystematicallyleverageAItodevelopcompetitiveadvantages,recentresearchshedssomelightonobstaclestodoingso.Weexaminetheseobstaclesfromanorganizationalcapabilities

perspectiveandidentifythreereasonswhyfirmsmaystruggletoestablishcompetitive

advantageswiththeirAIinvestments.WefocusonAI’sgeneric,explicit,andmyopicnature.WhileAIundoubtedlyraisesmanyothernewchallengesforfirms,thesethreelimitationstakeprominenceinourtheoryduetotheiradverseeffectsoncapabilitydevelopment,whichwewillargueiscentraltoestablishingcompetitiveadvantageswithAI.

ThefirststrategicchallengeofAIisitsgenericnature.ThelogicemergingfromanAIalgorithmisgenerallynotuniquetotheuserapplyingthatalgorithm.Instead,itcanbe

“rediscoveredbyanyoneusingthesameprocedure”(Shresthaetal.,2021:4).Forexample,allelsebeingequal,aneuralnetworkalgorithmwillarriveatthesamelogicforconnectingasetofinputstooutputsregardlessofwhetherSpotifyorPandoraoperationalizestheneuralnetwork.ThissuggeststhatAImaydisplaycommonbehavioralpatternsacrosscompetingfirms.This

pointregardingthegenericnatureofAIiscriticalbecauseAIisregardedasageneral-purposetechnologyakintoelectricity,thesteamengine,ortheinternet(Brynjolfsson&Mitchell,2017;Frey&Osborne,2013;Lynch,2017).General-purposetechnologiesarelikelytobewidely

60

Page7of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

adoptedamongcompetingfirms(Bresnahan&Trajtenberg,1995).Asaresult,these

technologiestendtogenerateeconomy-widebenefitsratherthanprivaterents(Bresnahan&Trajtenberg,1995;Teece,2018).Thus,whileAImayhelpafirmtodevelopbetterandcheaperproducts,theex-facieexpectationisthatAIwillhelpafirm’scompetitorstodothesame.

ThesecondstrategiclimitationofAIisthatAImanifestsasaformofexplicitknowledge.AIalgorithms,andthedatathatdrivethem,mustbeavailabletothecomputerintheformof

explicitinstructionsormathematicalformulas(Broussard,2018).Consequently,the

organizationalknowledgethatdrivesafirm’sAIprocessesmaybeextractedfromcyber-attacksandmaybehighlyportableduringemployeeturnover(Tramér,Zhang,Juels,Reiter,&

Ristenpart,2016).Thisobservationisconsistentwiththegeneralideathatexplicitknowledgediffusesrelativelyquicklyacrossorganizationalboundaries,makingitchallengingtobuild

competitiveadvantages(Grant,1996;Nickerson&Zenger,2004).ThischallengeisespeciallysalientforAIbecauseintellectualpropertylawsdonot(currently)allowforpatenting

mathematicalformulasandprocedures(Gaudry&Hayim,2018;Liyange&Berry,2019).Thus,identifyingmechanismsforpreventingthespreadofAIassetsacrossfirmboundariesiscritical.

ThethirdstrategiclimitationofAIismyopia.AIismyopicinthesensethatAI

algorithmslackcontextualawarenessofactivitiesandeventsbeyondthescopeoftheirassignedtasks(Balasubramanianetal.,2022;Dreyfus,2012;Raisch&Krakowski,2020).AsingleAI

algorithmcantypicallyexecuteonlyasmallsubtaskwithinanentireorganizationalroutine

(Davenport,2018).Thus,AI-drivenroutineswillnormallyemploycollectivesofAIalgorithms(see,forexampleKumar,Venugopal,Qiu,&Kumar,2018).Yet,analgorithm’sabilityto

recognizeinterdependenciesbetweenitstaskandothertaskswithinthefirmislimited.Thismayresultinexpensivetechnicalandoperationalfailures(Balasubramanianetal.2022;Dreyfus,

60

AcademyofManagementReviewPage8of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

2012).Moreover,becauseAIcompletesitstaskwithhighdegreesofagency,managersfinditdifficulttocorrecttheseerrorswhentheyoccur(Murrayetal.,2021).

ArelatedconsequenceofAI’smyopiaisthatAIwilllackasophisticatedunderstandingofafirm’sstrategy.Thus,evenwhenAIbehavesinamannerthatisoptimalforcarryingoutatask,thereisnoguaranteethatthiswillresultinbehaviorthatisappropriateforthekindsof

marketopportunitiesafirmispursuing(orshouldbepursuing)(Balasubramanianetal.,2022).Forinstance,abudgetairlinemaydevelopAIthatcorrectlyidentifiesthatacustomercanpay$2,000forashort-haulflight.Still,thisAImaybeincapableofunderstandingthatmakingsuchanofferisinconsistentwiththefirm’smarketidentityandlow-costproviderstrategy.Inotherwords,theultimatevalueofAItothefirmdependsnotonlyonitstaskeffectivenessbutalsoonitsfitwiththefirm’soverallstrategy.Thus,organizationalmechanismsforovercomingAI’s

myopiaarenecessaryforestablishingcompetitiveadvantages.

Thegeneric,explicit,andmyopicnatureofAIallendangeritsfirm-levelbenefits,butareuniquelydifficulttosurmountbecauseAIisbothamachineandagentic.Forexample,AI’s

genericnatureisakintogeneralhumancapital.Generalhumancapitalcannottypicallyunderlyinterfirmadvantagesbecauseafirm’scompetitorscanusuallyacquireanddeploythat

knowledgeinwaysthatcloselymimicthefocalfirm(Barney&Wright,1998).UnlikewithAI,however,generalhumancapitalisconvertiblewithinafirmthroughsocializationandcanbe

madecontext-specificashumanemployeesengageinrichinformalsocialinteractionswithinthefirm(Coff,1997).ThisoptionisunavailableforAIbecauseitisamachine.Ontheotherhand,AI’scapacitytoactwithagencyremovesthetypicalmodesofcontrollingtechnologyinan

organization.Consider,forexample,thecaseofexpertsystems,whichwereheavilyusedbeforetherecentriseofAI(Cholletetal.,2022).LikeAI,thesesystemsweremyopicbecausethey

60

Page9of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

coulderroneouslyoverlookorganizationalinterdependencies.However,whensuchproblems

arose,theycouldbeaddressedbysupplyingthemachinewithmorerules.ThisoptionisnotasfeasiblewithAI,whichdependmoreondatathanrulesasbehavioralconstraints.Thus,AIaltersthefirm’sknowledgeproductionfunction,makingpathstocompetitiveadvantageelusive.

ConceptualBuildingBlocksforTheorizingAI-drivenCompetitiveAdvantages

WedevelopatheoryofsituatedAItoexplainhowfirmscanovercomethesestrategiclimitationstocraftcompetitiveadvantages.Weuseasourconceptualfoundationsthe

organizationalcapabilitiesliterature(Eggers&Kaplan,2013;Nickerson&Zenger,2004),andtheworkonhumanagency(Emirbayer&Mische,1998;Westphal&Zajac,2013).Webrieflydescribeeachconceptualbuildingblockbelow.

CapabilitiesandCompetitiveAdvantage.Anorganizationalcapabilityisacollectionofroutinesthat,togetherwiththeirimplementinginputflows,conferuponanorganization’s

managementasetofdecisionoptionsforproducingsignificantoutputsofaparticulartype

(Winter,2003).Theorganizationalcapabilitiesperspectiveviewsfirmperformanceasafunctionofsystematicandrandomfactors(Winter,2000).Marketsaremodeledascollectivesof

competingfirmssolvingarelatedproblemundertechnicalandbehavioraluncertainty(Afuah&Tucci,1997;Nickerson&Zenger,2004).Profitsareviewedasephemeralintheabsenceof

competitiveadvantagesgrantingsomefirmsastructuraledgeoverothers(Barney,1986).Andaafirm’sprimaryconcernliesinidentifyingandorchestratingpatternsoforganizationalactivitiesthatcanbereliablyleveragedtocreateandcapturevalue(Winter,2003).

Thecapabilitiesperspectivefocusesonthreecoresourcesofcompetitiveadvantage:

firm-specificity,capabilitydevelopmentcosts,andenvironmentalfit.Firm-specificcapabilitiesarethoseproducedusingco-specializedknowledge,leadingthemtohavegreatervalueinsidea

60

AcademyofManagementReviewPage10of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

firmthanexternally(Helfat,1994;Mahoney&Pandian,1992).Also,craftingfirm-specific

capabilitiesismorelikelywhenatleastsomeoftheco-specializedknowledgeneededtoproduceordeployacapabilityistacitorsociallycomplex(Grant,1996;King&Zeithaml,2001;

Nickerson&Zenger,2004).Second,becausecraftingcapabilitiesrequiresorganizationaleffortandresources,thecostofdevelopingacapabilitymustnotsupersedethevalueearnedfrom

deployingthecapability(Argyersetal.,2019;Winter,2000).Finally,customersarelikelyto

respondpositivelytoafirm’sofferingsonlywhenthefirm’scapabilitiesareadequatelymatchedtotheirneeds(Sirmonetal.,2007).

Situatedagency.Organizationsactwithagencytoinfluencetheircapabilitiesandthe

environmentsthatbindthem(Gavetti,Helfat,&Marengo,2017;Gavetti&Torras,2021;Nayak,Chia,&Canales,2020).Agencygenerallyinvolvesfreechoicewithconstraints(Emirbayer&Mische,1998;Giddens,1979).Ourbasicargumentisthat,whilefirmscannotalwayslimitAI’sagencydirectly(andmightnotwantto),firmscanbalanceamachine’sagencywithhuman

agency,bystrategicallystructuringthecontextinwhichAImakessenseofproblemsandappliessolutions.Agencyhasthreedimensionsthatinformhowanactor’sbehaviorisconstrainedandhowfreechoicemaymanifest(Emirbayer&Mische,1998).Theiterationaldimensionofagencyinvolvesactionsanchoredinanorganization’spriorexperiences.Thepracticalevaluative

dimensionofagencyconsidersactionanchoredinanorganization’spresentsocialcontext.Theprojectivedimensionofagencyregardsactionsbasedinanactor’sabilitytoreimaginetheir

organization’spresentarrangementstomeetfuturegoals.Agencyisconsideredtobesituatedwhenconstraintsonagencyoriginatepredominantlyinthesamecontextinwhichtheagentacts(Botti,1998;Westphal&Zajac,2013).WebuildonthisworkbyconceptualizingsituatedAIasthetechnologicalanalogofsituatedagency.

60

Page11of42AcademyofManagementReview

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

Capabilitydevelopment.WearguethatfirmscansituateAIduringcapability

development.Capabilitydevelopmentinvolvesorchestratingorganizationalactionacrossfour

steps:bundlingstrategicassets,embeddingassetsinroutines,assemblingroutinesascapabilities,andmatchingcapabilitiestoopportunitiesintheenvironment(Collis,1994;Eggers&Kaplan,2013;Sirmonetal.,2007).Withinthescopeofthismodel,wetreatinputdataasthemajor

strategicasset(Gregoryetal.,2021),andwereplacethefocusontraditionalroutinesinpriorcapabilitymodelswithafocusonconjoinedroutines,whichreferto(partially)automated

organizationalroutineswheretheroutine’sdesignandexecutioninvolvesamixofhumanandnonhumanagency(Murrayetal.,2021).

We,therefore,defineanAI-drivencapabilityasacollectionofconjoinedroutineswhich,alongwiththerequiredinputdata,allowafirmtoexecutespecificvaluechainactivitiesina

repeatableandreliablemanner(Gregoryetal.,2021;Helfat&Winter,2011;Winter,2003).OurmodelintroducesthreesituatingactivitiesthatfirmsmayleveragetoorchestratethedevelopmentofAI-drivencapabilities:grounding,bounding,andrecasting.

AcorepartofourtheorizinginvolvesaccountingforhowsituatedAIisadaptedinafirm’sAI-drivencapabilitiesovertime.Therefore,webuildonpriorresearchviewing

organizationaladaptationasincrementallychangingafirm’scorestructuresandstrategies

throughexperimentationandproblem-drivensearch(Ethiraj&Levinthal,2004).Wefocusonthelearningliteraturecenteredonthecognitiveunderpinningsofadaptivecapabilityformation(Gavetti&Levinthal,2000;Tripsas&Gavetti,2000).

Afinalconsiderationadoptedfromtheorganizationalcapabilitiesperspectiveisattentiontodynamisminthefirm’senvironment.Asafirm’sstrategicenvironmentbecomesmore

dynamic,newmarketopportunitiesmayemerge,existingopportunitiesmayevaporate,andthe

60

AcademyofManagementReviewPage12of42

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

technologiesandcapabilitiesneededtocaptureopportunitiesevolve(Sirmonetal.,2007;Teece,2007;Tripsas&Gavetti,2000).EnvironmentaldynamismhassignificanceasitrelatestoAI-

drivencapabilitiesbecauseAIshiftstheextenttowhichafirm’sroutinesareresponsivetochange(Balasubramanianetal.,2022;Murrayetal.,2021).We,therefore,considerhowenvironmentaldynamisminfluencesthebenefitsofsituatingAIforcompetitiveadvantage.

AIcharacteristics.OurtheoryisintendedtoaccountforthelargeandgrowingformsofAItechnologies.WefollowthemachinelearningliteraturebycharacterizingAIalgorithms

basedontheirtrainingparadigms(supervisedversusunsupervisedlearning)andtheirdegreeofexplainability.AnAIalgorithmismoreexplainablewhenitiseasierforhumanoperatorsto

describethelogicthroughwhichthealgorithmlinksinputstooutputs(Aryaetal.,2019;Gilpinetal.,2019).Thismayinvolvedescribingwhichfactorsanalgorithmweighsheavilywhen

arrivingatasolutionorprovidingsomeintuitionforhowthealgorithmtreatstheinteractionbetweendifferentfactors(Gilpinetal.,2019;Hendricksetal.,2019).Whilesomealgorithmsallowforahighdegreeoftransparencyregardingtheirinnerworkings,otherAIalgorithmsdonot.Figure1showshowsomestandardAIalgorithmsfitthistaxonomy.

***Figure1abouthere****

ThelearningparadigmtellsushowAIisinstructedtomakeinferencesfromthedata.

Withsupervisedlearning,AIisprovidedwithtrainingdatainwhichthe“rightanswers”fora

problemhavebeenlabeled(Dike,Zhou,Deveerasetty,&Wu,2018;Murphy,2012).Labelsmaybeprovidedbyhumanactorsormaybeinferredusinghumanknowledgeofthedatasource.Forexample,afirmmaytrainAItowritecomputercodebyscanningsiteslikeforuser-providedcodingquestionsandthenusingthehighest-ratedinputasthecorrectanswer.Incontrast,unsupervisedlearningprovidesAIwithdatabutnotwith“ans