【谷歌Google】2024生成式AI高管指南5724mb

Theexecutive’sguidetogenerative

AI.

Introduction

Aboutthisguide

......................................................................................................................................................04Chapter

1

AquickprimerongenerativeAI

........................................................................................................................05Corecapabilitiesandapplications

.......................................................................................................................09Chapter

2

Thestep-by-stepguidetogettingstartedwithgenerativeAI...............................................................

10How

to

launchyour

first

use

case

in

30

days

.....................................................................................................

12KPIstomeasuretheimpactofgenerativeAI

.....................................................................................................26Chapter

3

ThevalueofgenerativeAIineveryindustry.................................................................................................

28Seewhatyourpeersaredoing

..............................................................................................................................31RetailandCPG

...........................................................................................................................................................32Financialservices

......................................................................................................................................................35Healthcareandlife

sciences

...................................................................................................................................39Mediaandentertainment

.......................................................................................................................................

42Manufacturing

...........................................................................................................................................................46Telecommunicationsserviceproviders

...............................................................................................................

51Conclusion

Innovate

fasterwith

generative

AI

for

business

..........................................................................................

54TableofcontentsIntroductionGenerative

AI

marks

one

of

the

most

significant

technological

shifts

in

history.

Itsimpactonindividualandbusinessproductivitycanbesignificant,withthepotentialtorivalthe

adventoftheinternetorthemobiledevice.Indeed,amongorganizationsconsideringorusingAI,82%believeitwilleithersignificantlychangeortransformtheir

industry.1WhatmakesgenerativeAIdifferentfromotherformsofAIthathavecomebeforeisitseaseofuseinhelpingsolveeverydayproblemsinpeople’spersonalor

professional

lives.

Anyone

who

knows

how

toaskaquestionofasearchenginecanuseeverydaylanguagetointeractwithagenerativeAIchatbotorvirtualagent—gettingittoanswerquestions,createcontent,produceimages,summarizedocuments,andmuch

more.Evenbetter,asinglegenerativeAIplatformcandeliversolutionsformultipleusecases,creatinganetworkeffect.Asthenumberofusersandapplicationsincreases,themodelisexposedtomoredataandbecomesincreasinglyaccurateanduseful—whichinturnencouragesmore

users.OrganizationsthatusegenerativeAItospeedup,automate,scale,andimprovebusinessprocessesstandtoreapbigbenefits.AccordingtoMcKinsey&Company,generativeAI’simpactonproductivitycouldaddbetween$2.6trillionand$4.4trillionUSDannuallytotheglobal

economy.2

Notechnologyevertakesawaythefundamentalsofyourvaluepropositionandthecorevaluechaininyourindustry.Attheendofthedayinhealthcare,forexample,you’restilltryingtoimprovepatient

care.Whatcanchange,though,ishowyouusethistechnologytoenableyourteamstoimprovecoreofferings,andhowyousolvefundamentalproblemsthatgetinthewayofdeliveringthem.Indeed,withtherighttools,youcouldevenidentifyanddelivernewpointsof

difference.Google

Cloud

Gen

AI

Benchmarking

Study,

July

2023McKinsey

&

Company,

The

economic

potential

of

genera?tive

AI,

20233Thisguideisforbusinessleaderswhowanttokickstarttheirorganization’sgenerativeAI

journey.Inthefirstchapter,you’lllearnwhatgenerativeAIis,whatitiscapableof,andtheimpactyoucanexpectwhenapplyingitinabusinesscontext.Thesecondchapterprovidesastep-by-stepguideforgettingstartedwithgenerativeAI,withrecommendedbestpracticesfromGoogleCloud’sAIexperts.Next,wediveintoreal-worldexamplesofpioneersacrossindustrieswhohaveadoptedgenerativeAItoworksmarter,delivervaluetocustomersfaster,andunlocknewrevenue

channels.WithgenerativeAItechnologymovingsofast,itcanbeoverwhelming.We’reheretohelpyoucharttheright

path.Aboutthis

guide4Aquickprimerongenerative

AIChapter

015Thisgenerationoftechnologyisthemostapproachableandflexibletechnologythatcomputersciencehaseverbuilt.Itcansolveproblemsthatthealgorithmwasn’tpurpose-builtfor.And,anybusinessusercanquicklyexperimentandgetfeedbacktoshapeitsusefortheirparticularbusiness

problem.”

Philip

Moyer

Global

VP,

AI

&

Business

Solutions,

Google

CloudChapter

01

|

A

quick

primer

on

generative

AI6Everyday,peopleinyourbusinessspendtimeandenergydiggingforinformationtomakedecisions,servecustomers,andmoveyourbusinessforward.Informeddecisionsrequireinformation,andcollectingtherightinputscantake

time.

Supposeyouneedtoknowhowadspendimpactscustomerperceptions.Oryouwanttoseethetrendsinacompetitor’spatentfilings,R&Dinvestments,andtechnologyacquisitions.Informationlikethisexistssomewhereinyourorganization,often

acrossa

bunch

of

places.

You

needitto

decide

the

next

step—and,toobtainit,expertsneedtobeconvened,researchcompleted,andtheinformationcompiledandsynthesized.Ifyouhaveafollow-upquestion,thewholeprocessmighthavetostart

again.Fromexecutivesdiggingintostrategictrendstosalespeoplecreatingproductdemosornewemployeeswithbenefitsquestions—everyoneinyourorganizationcanrelatetothisfrustration.Butthat’s

changing.Imaginegivingeachpersoninyourcompanynotonlyapersonalassistant,butanexpertineverypieceofdatarelevanttotheirjoband,indeed,potentiallyeverypieceofdataacrossyourwholeorganization.Withanassistantlikethat,impatientmomentsofindecisionwoulddwindle.Everyonewouldbeempoweredtospendlesstimewaitingandmoretime

doing.WithgenerativeAI,thisisachievable.Andit’sjustoneexampleofthemanydisruptionsthistechnologyhasunlocked.Always-oncodingcollaborators.Brainstormingassistantstodraftanditeratecontent.Personalizedself-educationonanysubject.Human-likeinteractionswithcustomerswherevertheyneedyou,forwhateverreason.GenerativeAIisallthesethingsandmore.Intime,itwillaffectalmosteveryaspectofevery

business.Chapter

01

|

A

quick

primer

on

generative

AI7Foundationmodelsaretheenginepoweringgenerative

AI.GenerativeAIapplicationsarepoweredbyfoundationmodels,whicharetrainedonvastamountsofcontent.Forexample,LargeLanguageModels(LLMs)

areonetypeoffoundationmodel,trainedontextorlanguage.Othermulti-modaltypescanbetrainedonimages/photos,video,music,softwarecode,medicalinformation,orcybersecuritydata.Butaccesstomodelsalonewon’tpositionyourbusinessfor

success.Foundationmodelsarebestthoughtofasprobabilityenginesthatcanbenudgedandshapedbyhumaninput.Becausetheyareprobabilistic,theyarefundamentallydifferentfromtraditionalsoftwareparadigms.Whentoday’sappsneedtolookupproductpricesorvalidatecustomerinformation,theyusedeterministicfunctionsthatcalladatabase.Incontrast,foundationmodelsusethepatternstheylearnduringtrainingandtuningtocalculatethemostprobableoutput,suchasthemostlikelyanswertoaquestionoranaccuratecaptionforan

image.Sincethey’renotconfinedbyrowsandcolumnsinadatabase,foundationmodelsareextremelypowerful.They’reoftencapableofperformingmanydownstreamtasks—suchasQ&A,summarization,

oropen-ended8contentgeneration—withlittleornoadditionaldataortuning.Yettheycanalsobeexpensivetotrain

andrun,pronetoinaccurateoutputs,anddifficulttowork

with.Forthesereasons,generativeapplicationsaren’treducibletogenerativemodels.Yourintelligentapplicationswillneedtomixprobabilisticfoundationmodelswithtraditional,deterministic(inotherwords,constrained)programming.Deterministicmodelsareconstrainedtowhatcanflowoutofthem,andarelimitedbyendlessoptionsthathavetobe

pre-established.TraditionalAIispurpose-builtforthetaskathand.Itisallaboutoptimizingandtweakingexistingprocessessuchasforecastingspecificpatternsthathave

beenpre-determinedbyhumans.It’swhytraditionalAIcanbeusedtoautomatediscrete,standardizedprocessesinspecificareasofthebusiness,suchascustomer

service.Incomparison,generativeAImodelshaveemergentcapabilitieswheretheycanperformmultipletasks,eveniftheywerenotexplicitlytrainedforthemduringtheinstructiontuningphase.Itisthisabilitytomultitask,plusthefreedomaffordedbythepromptinginterface,thatenablesthesemodelstoperforminawidevarietyofuse

cases.Chapter

01

|

A

quick

primer

on

generative

AICorecapabilitiesand

applicationsAtthecore,generativeAIhasfour

capabilities:

Creation

Summarization

Discovery

AutomationAnd

it

tends

to

excel

in

four

applications:ChatIt’snocoincidencethatgenerativeAIhasgainedrapidpopularityandadoptionthroughsimplechatinterfaces.ChatisanaturalwaytointeractwithpowerfulgenerativeAImodels.Youcanuseittoimprovecustomerinteractions,enhanceproductcapabilities,trainemployees,and

more.SearchBycombininggenerativeAIcapabilitieswithsearch,youcananchoronaknowledgebase—eitherinternalorexternal—formoretailoredandtargetedinteractions.Usinggenerative

AIforsearchcanhelpeliminatehallucinationsbysourcinginformationfromafactualknowledge

base.Generate

contentTheabilitytogeneratehigh-qualitytext,images,speech,andcodehasenormouspotential.Whetherit’sspeedingupprocessesorhelpingemployeesturnideasintooutputfaster,generativecapabilitiescanbedeployedintoproducts,tools,and

workflows.Associative

reasoningThisistheabilitytosuggestassociationsininformationbasedoncontext,frequency,orproximity.Forexample,generativeAIcouldidentifythethreemostcommonreasonsthatacallcenterinteractionendsnegativelybyparsingthelargeamountsoftranscribed

conversations.Readourglossary

ongenerativeAItermsand

concepts.Chapter

01

|

A

quick

primer

on

generative

AI9Chapter

02Thestep-by-stepguidetogettingstarted10本報告來源于三個皮匠報告站（）,由用戶Id:262617下載,文檔Id:186866,下載日期:2024-12-28Byclusteringusecasesintoadomain,youcanstartwithone,andthenwhenthat’sworking,youcannaturallyexpandtoyoursecond,third,andfourthusecasewithinthesamedomain.Themoreyouthrowatamodel,thesmarterit

gets.Takecustomerservice,forexample.First,let’ssayyougiveyourcallcenteragentsagenerativeAItoolwithaconversationalinterfacetousewhentakingphonecalls.Customersmightcallinandsay,“Mycreditcardnolongerworks.”O(jiān)r“Ilostmypasscode.”O(jiān)r“WhenIgotraveling,howdoImakemycellphoneworkinternationally?”Youragentscananswer

thesequestionsnaturally,queryingthegenerativeAIinterfacetoprovideananswerusingsimilarwordingtohowthecustomeraskedthe

question.Then,youcanlookattheaggregateofthosequeriesandaskgenerativeAI,“Whatquestionsarebeing

askedmostfrequently?What’sourresponsetime?Whatresponsesdoweprovide?”Now,you’removingfromjustansweringcustomers’questionstosummarizingtheir

data.Third,usingthissummarization,youcanpromptgenerativeAItotakethemostfrequentlyaskedquestionsandcomparethemtotheFAQsonyourwebsite.Youcanthenaskittogenerateanswerstoanyquestionsnotansweredonyourwebsite,soyoucanpublish

them.Inthisexample,generativeAIsolvesthreeusecasesthatenhancethecustomerservicedomain—answerthequestion,summarizewhat’sbeingasked,andproduceanswerstothesequestions.Witheachusecase,themodelgets

smarter.Toenableyourfoundationmodelstogetverysmart,

very

fast,

pick

one

functional

domain

inyourbusinessandexperiment

inthatarea.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI11IdentifyaspecificdomainCreateatiger

teamSelectapersonaDeterminedata

sourcesStep

1Step

2Step

3Step

4BuildUXand

UIExpandtomore

usersDefineyourgoalsBuildanLMoperations

planExpandtomoreuse

casesDesignpromptswithtiger

teamStep

7Step

8Step

5Step

9Step

6Step

1010stepstolaunchyourfirstusecasein30

daysFollowthesestepsforafast,easy,andlow-riskapproachtostartingyourorganization’sgenerativeAI

journey.We’veincludedKPIstohelpyoudemonstrateimpacttoyourleaders,foundationalprocessestoscaleacrossdomains,andrecommendationstoallowyourteamstoexperimentinasafe,internalsettingsotheycangainexperiencewiththenew

technology.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI12Step1

of

10Identify

a

specific

domainChooseadomaininyourcompanythatcouldbenefitfromgenerativeAI,suchascustomerservice,patientintake,corporateactions,ormarketing

content.

Key

questions:Inwhichareasofthebusinessdoemployeesspendsignificanttimeonrepetitive

tasks?Isthereaprocessorpartofarolethatis,ineffect,alreadystandardized(e.g.,havingtodoaparticularsteporansweraquestionevery

time)?Wheredoemployeesgetstuckinthecreativeprocess(e.g.,writer’sblockorcreative

block)?Wouldanincorrectresponseorhallucinationcause

harm?Whichareasofthebusinessofferthelowest-riskenvironmentforinitialuse

cases?Doyouhavealargecorpusofdatathatyouwanttoactivate,ormakemore

useful?Isthereanareaofthebusinesswhereemployeesconsistentlyhavetosearchforexistinginformationusinginternalknowledgebases

and/orexternalsearch

engines?Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI13Step2

of

10SelectapersonaDeterminewhichjobcategoryorfunctionwithinthechosendomainyouwanttomakemoreproductive.

Considerthesethree

factors:Lookatjobrolesthatarehardtoretainandhardto

hire.Suchrolesareoftenrepetitiveandofferlittlecareeradvancement.Automatingthesetaskscanfreeupemployeestofocusonmorestrategic

work.Findopportunitiestoautomaterepetitive,tedioustasksthatarenecessarytogenerate

revenue.Forexample,themulti-trilliondollarpre-authorizationindustryinhealthcareisincrediblyfrustratingforpatients.ItcantakehoursorevendaystogetauthorizationforasimpleMRIorspecialist

visit,andtheprocessisoftenpaper-based.Investmentmemorandumsareanotherexampleinwhichthesameinformationmustbegatheredoverandoveragain.GenerativeAIcanhelpautomatethesetasksso

employeescanfocusonmorestrategicwork.Makesafetyandcompliance

ambient.Manyindustriesmustmeetstrictrequirementstobecompliant.Inlifesciences,forexample,everyclaimabouttheefficacyofamedicinemustbereviewedbyanattorneytoensurethelanguageiscompliant.Anattorneymustthenreviewtheclaimtoensurecertainclausesareincludedineveryinstanceofthefineprint.Thiscanbetimeconsumingandrepetitive,butisanessentialpartoftheprocessformakingmedicinepubliclyavailable.GenerativeAIcanhelpautomatecriticaltasksthathelporganizationsimproveaccuracyandreduce

risk.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI14Your

generative

AI

model

will

be

trained

on

the

data

you

gather.Thisdatashouldbehonedforthespecificbusinessordomain-levelproblemitistryingtosolve,andaccessiblethroughenterprisedata

sources.Forexample,ifyouselectamarketingmanagerasyourfirstpersona,it’simportanttounderstandthespecificsoftheirjobrole.Saytheyareresponsibleforcreatingdigitalcampaigncontentlikeebooksandreports.Theseassetsarepromotedandusedtocaptureleadsviaonlineforms.Iftheindividualfillsouttheformandoptsintoreceiveoutreachaboutrelatedtopics,theleadisenteredintoamarketingautomationprogramandscoredbasedonpre-determinedcriteriasetbymarketingandsalesoperationsteams.Iftheleadmeetsthecriteriaforqualifiedsalesopportunities,itmayalsobehandedofftothesales

team.Thisindividualjobrolerequiresseveraldatasources,

including:AwordprocessingtoollikeGoogleDocstowrite,edit,andcollaborativelyreviewthe

copyA

design

tool

to

lay

out

and

format

the

final

copyAwebplatformtopublishthecontent

onlineAmarketingautomationtooltotrackandmeasuremarketingengagement,tasks,and

workflowsACRMlikeSalesforcetoensuresales,support,andmarketingarecoordinatedacrossallusertouchpointsandcustomer

interactionsStep3

of

10DeterminethedatasourcesthepersonaneedstobeproductiveChapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI15Mitigate

hallucinations

AImodelsaretrainedtogiveuserswhattheywant,whichmeanstheyoccasionallyspitoutmade-upanswersthatsoundconvincingandcanbehardtospot.Toavoidthisproblem,youcangroundresponsesinspecificdata

ratherthanrelyingontheLLM

alone.Enhancetheexplainabilityof

AI

GenerativeAImodelscanbecomplex,andthe‘thinking’thatanalgorithmusestoproduceanoutputisn’talwaysclear.ExplainableAIislikeaslidingscale—therearedegreestowhichyoucan

explainorreliablysteeranLLM’sbehavior.Bystartingoffwiththerightdatatofeedintoandfine-tunemodels,yourorganizationwillbeable

to:

Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI16Includepeoplefrombothbusinessand

technology:Step4

of

10Createa

three-person

tiger

teamIndividualfromthe

businessPromptengineerMLoperationsleadResponsiblefordetailingthejobrequirements,workflows,challenges,andneedsoftheday-to-daytasksexecutedbythechosen

persona.Responsiblefortranslatingthebusinesspersona’sneeds,actions,andoutputintopromptsforthegenerativeAI

model(s).Responsibleforbuildingandoperating

theapplicationin

production.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI17Accelerated

innovationImprovedemployeeproductivityImprovedcustomerexperienceDevelopednewproductsor

servicesIncreasedcompetitiveness/market

shareImprovedstrategicdecision

makingIncreasedshareholder/investor

demand66%57%49%48%39%35%35%12%IncreasedoperationalefficiencyStep5

of

10Defineyourintentions,objectives,andtheoutputyouaretryingtoachieveEnsureyouhaveahumaninthelooptooverseethefirstusecasesandprovideoversight.

Keepinmind,thevalueofagenerativeAIprojectcancomefromanumberofsources.There’sdirectbusinessvalue,incrementalvalueofgenerativeAIoverlegacysystemsortraditionalAI/ML,andtheforecastedvalueofcapabilitiesoncescaledtootheruse

cases.Considerthefollowingoutcomes,whichotherorganizations

haveexperiencedafteradopting

AI:3Outcomesorganizationsreportafteradopting

AI3.

Google

Cloud

Gen

AI

Benchmarking

Study,

July

202318Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AIStep6

of

10Designpromptstogetherwiththetiger

teamWorkcollaborativelywiththetigerteamtodesignprompts

thatwillguidethegenerativeAImodel’s

responses.

Yourthree-persontigerteamhasexpertiseinthebusinessneed,theAImodel(s),tuning,

andapplicationintegration.Usetheirskillsandrefertopromptsamples

tomovequicklyinthis

step.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI19Herearesometipstokeepin

mind:Keeptheinterfaceanddesign

simple.Ensure

the

UI/UX

is

responsive

and

accessible

across

different

devices

andscreensizes,includingmobilephones,tablets,and

desktops.Considerhowthenewinterfacefitswithinthelargerecosystemofexisting

apps,

like

email

and

chat,

that

might

have

their

own

built-ingenerativeAI

capabilities.Begin

with

a

selection

screen

which

allows

users

to

choose

the

personalitytoapplytothegeneratedtext,images,oroutput.Forexample,optionscouldinclude‘formal’,‘casual’,‘technical’,or

‘creative’.Create

a

logical

and

intuitive

user

flow

that

guides

users

through

the

AImodel’sfunctionality.Ensurethattheinterfacedesignalignswiththeexpecteduser

journey.Step7

of

10Buildauserexperience(UX)anduserinterface

(UI)Createauser-friendlyexperienceandinterfacethatwillrunthegenerativeAImodelinproductionforthechosenpersona’suse

case.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI20Step8

of

10ExpandusagetoadditionalindividualsOnceyouaregettingacceptableresultsfromtuning,invitetwoorthreeotherindividualswithinthechosenpersonatostartusingthe

model.Continuetesting,measuring,andtuningwiththisgroupuntilyouaregettingconsistent,qualityoutputs,thenexpandusagetofiveto10morepeoplewithinthechosenpersonaandcontinuefine-tuningthe

process.Witheachnewindividual,makesureyouunderstandthedifferentwayseachuserinteractswiththegenerativeAImodel.Todoso,conductuserinterviews,surveys,orworkshopstogatherinsightsintouserpreferences,painpoints,and

desiredfunctionalitiesfortheAImodel’s

interaction.Chapter

02

|

The

step-by-step

guide

to

getting

started

with

generative

AI21Step9

of

10BuildaLanguageModel(LM)operationsplanDevelopaplanforproductionizingandmonitoringtheAImodel’soutputtoensureitfunctionseffectivelyand

safely.Keyquestionstoaskyourmanagement

team:Can

we

quickly

evaluate

and

experiment

with

generative

AI?Dowehaveamechanismforcontinuousimprovement?Areweabletoassess,

evaluate,

and

re-engage

to

go

deeper

within

exis