mckinsey -2025 年美妝企業(yè)如何擴(kuò)大生成式人工智能的規(guī)模 How beauty players can scale gen AI in 2025

Mcsey

&company

ConsumerPackagedGoodsPractice

HowbeautyplayerscanscalegenAIin2025

BeautybrandsandretailershavealreadybeguntestinggenAI.Toscaletheseexperimentsquickly,beautyplayersshouldfocusonhigh-valueusecasesandcustomizegenAItoolstomeettheirneeds.

ThisarticleisacollaborativeeffortbyAnnaCheca,KristiWeaver,MeganPacchia,SaraHudson,

andWeiWeiLiu,withAlexisWolferandAnaBujosa,representingviewsfromMcKinsey’sConsumerPackagedGoodsPractice.

January2025

Beautyisnolongerintheeyeofthebeholder;it’satthefingertipsofthegenerativeAI(genAI)prompter.GenAIcouldadd$9billionto$10billiontotheglobaleconomybasedonitsimpactonthebeautyindustryalone,1andearlymovershavealreadybeguntestingthetechnology.But

scalingtheseexperimentswillbechallenging,giventhevelocityofgenAIinnovation.

Thegapbetweenthe

laggardsandleadersinthebeautyindustry

willonlygrowonceleaderssuccessfullydeploygenAIatscale.Thefastwillbecomefaster,moreresponsive,andbetterequippedtoanticipateanddeliverwhatconsumerswant,whilethoseleftbehindmayfindit

hardertoholdontosliversofmarketshare.

BeautyplayerswhofocusonpriorityusecasesandcustomizinggenAItomeettheirneedscanhelp

realizethetechnology’sfullpotential

.ThisarticleoutlinesfourgenAIusecasesthatbeautyplayerscanprioritize,explainshowtobringgenAItotheorganization,andlaysoutasetof

imperativestosupporttheuseofgenAIinbeautyoverthelongterm.

FourusecasesofgenAIinbeauty

MorethanadozengenAIusecasesthatapplytothebroaderconsumersectoralsoapplyto

beauty.Theseusecasesspantheorganizationfromfronttoback,includingfunctionsfromuserexperiencetocustomersupport(table).

Toprioritizetheusecases,weconsiderthatthebeautysectorreliesonspeedinbringingproductstomarketandrespondingtoconsumerfeedback.Onthatbasis,fourgenAIusecasesarelikelytohavethegreatestimpact:hyperpersonalizedtargeting,experientialproductdiscovery,rapidpackaging-conceptdevelopment,andinnovativeproductdevelopment.TheseemploygenAI

toolsatvariousstagesofadoption.Some(forexample,genAIcustomerchatbots)arealreadyinfairlywideuseamongbeautyplayers,whereasothersarenascentbutpromising.

Hyperpersonalizedtargeting

Oneofthemostimportantmovesabeautybrandcanmaketosurviveinthecompetitive

beautysectoristodevelopauniquevalueproposition.Butbeautyplayersmustalsoensurethattheproductstheyhavethoughtfullypositionedreachtheconsumerswhowillbemostreceptivetothem.

Today,mostbeautycompaniescanaffordtotargetonlyahandfulofconsumersegments

becausetheyhavelimitedcapabilitiestopersonalizemessagesonabiggerscale.Thisbroadapproachtoconsumersegmentationleavesmuchofthemarketuntapped.ButwithgenAI,beautybrandscancreatehyperpersonalizedmarketingmessages,whichcouldimprove

conversionratesbyupto40percent,basedonourobservations.

AIcananalyzelargeconsumerdatasets,detectpatterns,andcreatemicrosegmentsbased

onpatternrecognitionalgorithms.Fromthere,abeautybrandcantrainitsgenAIplatform

usingavarietyofinputs,includingcustomerdata,inputsthatdescribethebrandvoice,and

productinformation.Whenenteringnewmarkets,beautybrandscantraingenAImodels

oninternalproductdataaswellasexternalmarketresearch,suchascustomersurveys.GenAIcanthencreateandtestvariationsoftextandimagestoseewhatresonatesbestwitheach

consumersegment.

1“

TheeconomicpotentialofgenerativeAI:Thenextproductivityfrontier

,”McKinsey,June14,2023.Thebeautymarketisdefinedasskincare,fragrance,makeup,andhaircare.FiguresassumearelativelyhighrangeofgenAIimpactonconsumerpackagedgoods(1.4to2.3percentoftotalindustryrevenue)becauseofthehighimpactofmarketingandsalesstrategiesinbeauty.

HowbeautyplayerscanscalegenAIin20252

Table

BeautyplayerscanusegenerativeAIacrosstheirconsumer-facingfunctionsandwithintheirinternalvaluechain.

Front

Middle

Consumerinsightsanduserexperience

Trendidentificationandsociallistening(eg,risingtrends,

sentimentanalysison

existingproducts,assortmentvs

competitors)

Microsegmentationandcomprehensiveconsumerpersonasbasedonconsumerdata

Dynamic‘clienteling’forretailstore

associates(eg,

summarizedcustomerinformation)

Marketing

Creativecontent

generationand

versioning(creation,editing,and

translationof,eg,

productimages,

productdescriptions,socialmediaads)

Experiential

productdiscovery(eg,synthesisof

productreviews,

conversational

websiteexperienceforeasybrowsing)

Hyperpersonalizedtargeting(eg,

streamlined

customerprofiles,individualized

recommendations,offers)

Mediaoptimization(towardhigher-ROIchannels),media

generation,andSEOoptimization

Innovation

Rapidpackaging-concept

development(eg,digitalproduct

packagingdesigns,3Dvisualmodels)

Innovativeproductdevelopment

throughquickly

testingcombinationsofingredientsor

chemicals

Storedesignandplanning(eg,

AI-generatedstorelayoutoptionsforinspiration,storeofthefuture)

Technology

Codeautomation

(eg,codegeneration,translation,

refactoring,review)

andqualityassurancetestingtoensure

high-qualitycode

Productmanagercopilottosupportcreationofkey

documents(eg,

PRFAQ1),planning,andbacklog

prioritizationandreporting

Back

Corporatebusinessfunctions

Performanceanalyticsandreporting(eg,campaign

performance,

e-commercemetrics,financereporting,

demandplanning)

Internalsearchandtaskautomationforemployees(eg,keyproductinsights,

applicationsforHR

benefits,requestsfortechreplacement)

Onboardingand

training(eg,conciselearningmaterialstoelevateperformance)

Recruiting

(eg,firstdraftofjobdescriptions,

screeningof

applicants,targetedinterviewquestions)

Salesandcustomersupport

Pre-callpreparationforsalesor

customeragent(eg,prepautomationforsalespitches,AI

role-playfortraining)

During-callsupport(eg,recommendations

basedonpastinteractions,

automatedcustomerservicechatbot)

Post-calltranscriptandperformance

analytics,including

adhocreportgenerationand

overarchinginsights

HowbeautyplayerscanscalegenAIin20253

Considerthehypotheticalautomatedtextsthatmightbedeliveredtoanimaginarycustomer

namedCamille.ThebeautybrandknowsthatCamillelivesinFrance,hasalowannualspend,

andrecentlypurchasedafacesunscreen.Camillehasrespondedpositivelytopromotionsinthepast.BeforegenAI,anautomatedtexttoCamillemightsay,“Excitingnews!Newproductsare

here.Takeupto20percentoffwhenyoushopsale.”AftergenAI,theautomatedtextmightsay,“Bonjour,Camille!Didyouknowthatourspecialcleansingfoamforfacesunscreenremovalis

now20percentoff?Itwillpairperfectlywithyourrecentfacesunscreenpurchase.”

MarketingspecialistsshouldreviewAI-generatedmessagesbeforetheyaresenttoensuretheyreflectthebrand’sethosandvaluepropositionwhileavoidingplagiarismorpotentiallyharmful

connotations.Somemessagesthatseeminnocuouscanbedetrimentaltoabrand’simage.In

thepreviousexample,thegen-AI-createdgreetingmighthavesaid,“Goodevening,lovelylady,”insteadof“Bonjour.”Acustomermayfindthetoneofthismessageoffensiveorinappropriate,orthemessagemightbeatoddswiththebrand’soverallethos.ThemarketingteamshoulddeliverfeedbacktothegenAImodel—perhapsratingitsoutputswithathumbs-uporthumbs-down

mechanismandenteringdetailedcommentsinfree-textfields.ThegenAIplatformcanthenprocessthefeedbackandconvertitintonewtrainingdata.

BeautybrandswillalsoneedtointegratetheirgenAImodelswithassetsfromtheirdigital-asset-management(DAM)systems,whichserveastherepositoryforallthedigitalcreativeassetsa

branduses,aswellasintegratethemodelswiththebrand’scampaignmanagementtools.GenAIcancategorizethecreativeassetsintheDAMsystem—ataskthatwouldotherwisehave

tobedonemanually.Thisautomationfreesuptimeforthemarketingteamtofocusonhigher-valuetasks.

Evenastheycontinuetoworkwithmarketingagenciestodeveloptheirbrandstrategyand

deliverspecializedcampaigns,largebeautyenterprisesmightconsiderinvestinginin-househyperpersonalizationcapabilities.Thiswouldoffertwomainadvantages:companiescan

usetheirownconsumerdatatotraingenAImodels,andtheycancreateandtestpersonalizedcommunicationswithgreaterspeedandagility.

Experientialproductdiscovery

Despitethetech-poweredinnovationsinconsumerproductdiscoveryoverthepastfewyears,

thereisampleroomforimprovement.Thefirstgenerationofconsumerchatbots,forinstance,providerelativelyrigidanswersandcanbefrustratingforconsumerstouse.Whenaconsumer

asksforarecommendationforanewblushforadarkercomplexion,forexample,achatbotmightgiveagenericlistofproducts,ratherthanpersonalizingtheconversationforaspecificshopperandengagingthemindeeperconversation.Virtualtry-onsarehelpfulbutcanbeglitchyorfailtoaccuratelyreflectwhataproductwouldlooklikeonaconsumer.Inthesecases,onlinepurchasesoftencanleadtocostlyreturns,sincereturnedbeautyproductsgenerallycannotberesold.

Gen-AI-poweredchatbotscanhelpimprovetheshoppingjourneyanddecreasethelikelihoodofreturns.Theselargelanguagemodel(LLM)chatbots,whicharetrainedonproductdataand

consumerpreferences,canrespondtoawidervarietyofquestionsandoffermorepersonalizedrecommendations,bothofwhichcanimproveconversionrates.Onegloballifestyleplayer

developedagen-AI-poweredshoppingassistantandsawitsconversionratesincreasebyasmuchas20percent.

HowbeautyplayerscanscalegenAIin20254

Thevirtualtry-onexperience—whichhasalreadyprovensuccessfulinotherconsumercategories,

suchasaccessoriesandeyeglasses—mightalsobeenhancedwithgenAI.Usingthesame

technologythatpowersimage-generatinggenAItools,consumerscanseethelookofdifferentproductsontheirskinindifferentsettingsorseethepotentialbenefitsaproductcouldhaveto

theirappearanceovertime.Anonlineshopperwhowishestolightendarkspots,forexample,couldvirtuallytryonabrand’sspot-lighteningserumbyuploadingaphotoonabeautyplayer’swebsiteandrunningasimulationoftheserum’spossibleeffectontheirskinoverseveralmonths.

GenAIcouldalsoenhanceexperientialproductdiscoveryinphysicalstores.Today,interactivetouchscreenmonitorsinstorescanshowproductsavailablebothin-storeandonline,allowingcustomerstobrowsethroughSKUs,selectitemstheywanttoseeinperson,orscanQRcodesforexclusiveoffers.Evenwiththeirlimitedfunctionality,thesescreenshavebeenshownto

improvethein-storeshoppingexperienceandconversionrates2GenAIcanboostthe

effectivenessofthesescreens.Forexample,whenashopperwhohaslocationservicesenabledonabeautyplayer’sappwalksintothecompany’sstore,genAIcouldgeneratecontent

personalizedtothatconsumerbasedoncustomerprofileandpurchasehistory.Givenwhatweknow

abouttheeffectivenessofpersonalizedcontent

,theseprinciplescouldtranslatetothestoresetting,thoughlarge-scaleimplementationhasn’thappenedyet.

Rapidpackaging-conceptdevelopment

Whenevaluatingabeautyproduct,consumersconsiderboththeproductitselfanditsbrandingandpackaging.Beautybrandstypicallyspendmonthsdevelopingnewbrandingandpackaging

concepts—aprocessthattypicallyrequiresdesigners,copyeditors,strategists,andpackagingexpertstoiterateonideas.

GenAIwouldn’tnecessarilyeliminatethisprocess,butitcoulddramaticallyaccelerateit.Here’showitcouldwork.ApackagingdesignerasksagenAIplatformthefollowingprompt:“Showme

fivepackagingoptionsforanighttimemoisturizer,emphasizingskincarebenefitsandsustainablepackagingmaterials.”ThedesignerthenmodifiesthegenAIplatform’soutputbasedon

informationaboutcustomerpreferences,whichcouldcomefromfocusgroupsandcustomer

surveys.Next,anadvertisingdesignerusesmockupsofthenewpackagingindigital

advertisementstotestwhethertheimagesappealtoconsumers,basedononlineengagementwiththenewads.Thatdataisthenusedtofurtherrefinegen-AI-poweredconceptcreationand

prototyping.Withthisbasicapproach,onebeveragecompanyreduceditsconceptdevelopmenttimeby60percent.

Innovativeproductdevelopment

Creatingnewbeautyproductformulasisamultiyearprocess.Itrequiresbeautyplayersto

partnerwithlaboratoriestoresearchingredientsandexperimentwithformulastodeterminethesafety,stability,andefficacyofanewproduct.

GenAIcanspeedupthisprocess.AgenAImodel—onceithasbeentrainedonabeautyproduct’sbillofmaterials,rawmaterialusage,processparameters,internalresearchdata,andotherdata

(suchasproductpatentsorpreviousproducttrials)—canidentifytheingredientsthatmaybebestsuitedforanewproduct,predicttheproduct’sbenefits,andrecommendformularecipes.

2“Digitalscreensandkiosksaredrivingshopperengagement,”AVMagazine,December18,2023.

HowbeautyplayerscanscalegenAIin20255

Returningtotheexampleofanighttimemoisturizer,assumethataformulationscientistcouldpromptthegenAItooltocreateanewformulathatemphasizesneuropeptides,apopularskin

careingredient,andprioritizesanti-agingbenefitswhilealsoreducingformulationcosts.Oncethetoolcreatesapotentialrecipe,thescientistwouldrunlabteststoassessthecompatibilityandstabilityofingredientsintheformulation,aswellasadditionalsafetyandconsumertestingandclinicaltrials,ifapplicable.Formulaiterationwouldcontinuebasedonconsumerfeedback.

Whilethephysicaltestingprocesswillstilltaketime,McKinseyanalysishasfoundthatgenAItoolscanreducethetimeittakestoresearchnewproductsfromweekstodays.Thiscanhelpsaveupto5percentonrawmaterialswhendevelopingthoseproducts.

Buy,borrow,orbuild?

ThemarketforgenAIenterpriseplatformsisgrowing.Butwhichapproach—ifany—isbestsuitedforbeautyplayers?

OrganizationscanbringingenAItoolsinanyofthreeways—whatwecallthetaker,shaper,and

makerapproaches.Mostbeautyplayerslikelywon’ttakethemakerapproach,wherecompaniesbuildtheirownLLMmodelsfromscratch.Thatwouldrequirecapitalexpendituresandtalent

investmentsgreaterthanmostbeautycompaniescanjustify;itcouldalsounhelpfullydilutea

beautyplayer’sfocusonitscorecompetencies.However,beautyplayerscanstillgetvalueoutofthetwootherapproaches:

—Takerapproach.Thetakerapproachentailsintegratingoff-the-shelfthird-partygenAI

solutionsintoabusiness’sworkflows,withlittletonocustomization.Thisistheleastcostlyandresource-intensiveofthethreeapproaches,soitisanattractiveoptionforbeauty

brandsthatrelyonretailersfordistribution(andthereforehavelessconsumerdatawithwhichtocustomizemodels),havelesstechtalent,orhavelesscashforinvestments.

InevaluatingagenAItoolorplatform,beautyplayersshouldaskquestionssuchasthe

following:Whatarethedataprivacyandencryptionprotocolsinplaceatthevendor?Willthevendorusethebrands’datatotrainthird-partyorfirst-partyproprietarymodels?Whoownsthecopyrightstotheoutputs?Howeasyistheintegrationwiththebeautyplayer’sinternal

systems?(Forexample,doesthevendorhaveanApplicationProgrammingInterface?AretheyintegratedwithplayerslikeGoogleAnalyticstoenablebroaderusecases?)

Pilotingthetooliscrucial,ofcourse.MostreputablegenAIvendorsofferalow-costpilotforalimitedtime—usuallyaroundonemonth.

—Shaperapproach.Beingashapermeanstrainingthird-partygenAImodelsonthecompany’sowndataandinsightsrelatedtospecificgeographic,sector,organization,andbusinesscaseneeds.Forexample,forhyperpersonalizedtargeting,thedatamayincludeinformationaboutabrand’svoice,customerdemographicsandpreferences,orsuccessfulcampaigns.For

innovativeproductdevelopment,rawdatafromclinicaltestresultscouldhelptrainmodels.

Largerbeautybrandsorretailerswithawealthofconsumerdatamaychoosetheshaper

approach.TheywillneedabenchoftechtalentthatcanaddnewcomponentstothegenAItool,integrateitintoexistingworkflows,anddeployitacrosstheorganization.

HowbeautyplayerscanscalegenAIin20256

BeautyplayerscanuseamixofthetakerandshaperapproachestogenAI,dependingon

theirspecificneedsandusecases.Speed—ingettingtomarketandrespondingtoconsumerdemand—isparticularlyimportantforbeautyplayers.Forthisreason,beautyorganizations

shouldconsidermodulargenAIcomponents,whichmakeswitchingbetweenLLMproviderseasiertodo,soscalingiseasier.GenAImayenablestreamliningandautomationinbeauty,but

theindustryisasmuchscienceasitisart;itwillbecriticaltokeepahumaninthelooptocheckforrisksandinjectuniquelyhumancreativityinto,say,marketingandpackagingdesign.

HowtoimplementgenAIatscale

TooutcompeteindigitalandAI,

consumer-packaged-goodsplayersshouldconsidercritical

questions

suchas“Whereisthevalue?”and“Areleadersfromthebusinesssideactivelypartofthetransformation?”Inaddition,beautyplayerscantakefourstepstotrulyintegrategenAIintothebusiness:

—Alignleadershiponthevision,value,androadmap.Tomovefromexperimentingtoscaling,beautyplayersshouldidentifywhichofthefourusecasesdescribedearlierinthisarticle

willyieldthegreatestrevenuelift,timeandcostsavings,andcustomerexperienceimpact.Tocalculatethispotentialandthenshapetheroadmapaccordingly,executivesmustbring

togetherleadersfromacrossvariousfunctions,suchasmarketing,customerservice,andproductdevelopment.

—Bolstercapabilities.AspromisingasgenAIis,usingiteffectivelyoverthelongrunrequires

thatbeautyleadersassesshowitfitsintoandissupportedbythe

organization’scapabilities,

includingitsoperatingmodel,dataandtechnologypractices,andtalent

.Companiesshouldsetupcross-functionalteamstoevaluatetheorganization’sexistingcapabilitiesand

requirementsforadditionalcapabilities.Theseteamsshoulddeployupskillingprogramsthathelpaddresscapabilitygapswithintheirranks.

—Test,learn,refine,repeat.BeautyplayersshouldtestgenAI’soutputincontrolledsettings

todeterminewha