Transcription
February 15, 2017REPORT: PRODUCT OVERVIEWInside Salesforce EinsteinArtificial IntelligenceA Look at Salesforce Einstein Capabilities, Use Casesand ChallengesDoug HenschenVice President and Principal AnalystContent Editor: R “Ray” Wang & Courtney SatoCopy Editor: Maria ShaoLayout Editor: Aubrey CogginsProduced exclusively for Constellation Research clients
TA BLE OF CONTENTSE XECUTI VE SU MM ARY. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3SA LE S FORCE E INS TE IN A SS IS TS H UM A NS R ATHER TH A N REPL ACE S THEM . . . . 4REQU IREMENTS FOR AI SUCCE SS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4IN SIDE SA LE S FORCE EINS TE IN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6COMPE TITI V E A SS E SS MENT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13RECOMMENDATI ONS: FOCUS ON BUSINE SS OUTCOME S. . . . . . . . . . . . . . . . . . . 15A N ALYS T B I O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18A BOUT CON S TELL ATI ON RE S E ARCH. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2017 Constellation Research, Inc. All rights reserved.2
E XECUTIVE SUMMARYThis report provides a product overview of Salesforce Einstein, which was introduced with muchfanfare at the company’s October 2016 Dreamforce event. The introduction marked the company’sofficial entry into the world of artificial intelligence (AI), but this “beginning” was the result of morethan two years of behind-the-scenes data science work and at least nine AI-related acquisitions. With asolid data-management backbone in place and a rich trove of acquired AI assets to draw on, Salesforcebrought 11 Einstein features to general availability within three months of the introduction in October.Ten more Einstein features are due in February 2017 and more than 20 additional features areexpected to be available before the end of the year.Salesforce Einstein is not a “general AI” offering that attempts human-like perception, thinking andaction. Rather, Einstein is an intelligence capability built into the Salesforce platform and focused ondelivering smarter customer relationship management (CRM). The features are designed to discoverinsights, predict outcomes, recommend actions and automate tasks.This report explains the data-management underpinnings of the Einstein platform, details the list ofavailable and soon-to-be-released features, explores Einstein strengths and weaknesses, and analyzesplanned developments on Salesforce’s roadmap. Readers should use this report to better understandSalesforce Einstein and determine whether it’s a good fit for their organization.Business ThemesConsumerization of ITDigital Marketing &Sales EffectivenesMatrix CommerceData to DecisionsFuture of WorkNext Gen Customer 2016 Constellation Research, Inc. All rights reserved.TechnologyOptimization3
SALE S FORCE EINS TEINA SSIS TS HUMANS R ATHERTHAN REPL ACE S THEMThe emphasis on assisting, rather thandisplacing, humans is, in part, reassurancethat Einstein is nothing like the dark, sciencefiction depictions of AI seen in movies rangingSalesforce stresses that its goal with Einsteinfrom “2001: A Space Odyssey” (1968) to “Theis to help people focus on what matters,Terminator” (1984) to “Ex Machina” (2015). Thenot to replace them. This is a concept thatfocus on practical, CRM-focused applicationsConstellation refers to as Augmentedcontrasts with the overly ambitious false startsHumanity. The idea is to uncover insights,experienced in the general AI experiments ofpredict outcomes, recommend next-bestdecades past.actions and automate routine, manual tasksthat keep people from being more productive.The Predictive Lead Scoring Einstein feature,REQUIREMENTS FORAI SUCCE SSfor example, helps salespeople focus on themost promising leads. Automated ActivityConstellation views four core capabilitiesCapture, a Sales Cloud Einstein feature dueand assets as being essential to developingin February, will automatically log sales calls,powerful AI skills (see Figure 1):email exchanges and chat interactions sosalespeople won’t face the drudgery of keepingcustomer interaction records up to date. A large corpus of data is the firstrequirement. It’s not the case that he whoFigure 1. Four Requirements for Developing Artificial Intelligence CapabilitiesSource: Constellation Research 2017 Constellation Research, Inc. All rights reserved.4
has the most data wins; the goal is to buildthe data store to support self-learning andthe largest graph that maps the connectionscontinuous learning (see Figure 2).to data. More data should improve theprecision of insights and allow for more Massive computing capacity is the secondpatterns to emerge. Data is used to test andrequirement, and it’s closely tied to thetrain algorithms and models, but the dataability to ingest, store and quickly analyzehas to be connected in some way to detectdata at scale. Public clouds have changed thepatterns and behaviors and then providescale and economics of computing, makingaccurate recommendations and suggestedit possible to tap vast computing capacity onor automated actions. The data exhaustdemand. Winners will have access to or ownof these systems is also brought back intovast computing power.Figure 2. Continuous Learning Unlocks a Spectrum of Seven Outcomes for AISource: Constellation Research 2017 Constellation Research, Inc. All rights reserved.5
Data science refers to intellectual property(IP), skill and experience. The discoveryscience behind AI-based systemsand applications.of patterns, creation of new algorithmsand the ability to apply human intuitionOther emerging and differentiatingto computing requires great math talent.requirements of AI include:The skills range from the basics of datamanagement, data cleansing, integrationand transformation to the ability to minedata and apply advanced statistical methodsas well as machine and deep learning to Industry-specific expertise to improve therelevance of specialized AI systems Natural user interfaces to take advantage ofhuman voice, visual and gestural interaction Robust recommendation engines that takeany amount of data. IP includes algorithms,the output of AI and present choices thatmodels and related proprietary capabilities.accelerate decision making Time is the fourth requirement, andit boils down to the people-yearsINSIDE SALE S FORCEEINS TEINthat can be poured into research anddevelopment. There is no substitute forThere aren’t enough data scientists in thetime. Early adopters gain an advantage ofworld to go around, so Salesforce is countingtime. Algorithms need time to improve.on automation to develop customized modelsCompanies can try to buy time by hiringfor each Einstein customer. Salesforce spentmore people or acquiring firms that havemore than two years building an automated,already sunk years into research andmachine-learning-based data-analysis pipelinedevelopment. But successful deliveryto deliver predictions and recommendationsof capabilities depends on time spentat scale. The work started after the acquisitiongenerating and learning from data,of Exact Target, part of efforts to developunderstanding computing requirements,predictive customer journeys. The systemand iteratively advancing the math and datacan scale up, Salesforce says, because all 2017 Constellation Research, Inc. All rights reserved.6
data collection, data preparation, featureSubscription pricing will vary between per-engineering and feature selection, modeluser/per-month and volume-based (leads,building, hyper-parameter tuning and scoringemails, etc.) models. The initial portfolio andsteps are handled automatically.applications expected through 2017 aredetailed in Figure 3.This internally developed data-managementand predictive engine is powering Community,A few of Einstein’s features to be releasedIoT, Marketing, Sales and Service Cloudwere derived in whole or in part from acquiredEinstein features that are either alreadycompanies. For example, Commerce Cloudavailable or due out this year (see Figure 3).Einstein features were derived largely fromIt’s already delivering as many as 300 millionDemandware, which had developed its ownpredictions, recommendations and lead scoresmachine-learning-based modeling and data-per day, says Salesforce. Most of Salesforce’sanalysis capabilities. Similarly, BeyondCore isAI acquisitions (though not all, as yet) willthe engine behind the Analytics Cloud Einsteinbe plugged into this same automated, data-features available at this writing. Salesforcemanagement and prediction backbone.is in the process of integrating internallydeveloped and acquired machine-learning andThe Roadmapautomated modeling capabilities. The companysays it will consolidate best-of-breed assetsEinstein capabilities are designed tofrom across its technology portfolio.complement and enhance existing SalesforceSaaS applications in that they are or will beFor now, Salesforce is focusing on deliveringbuilt into the Analytics, App, Commerce,straightforward Einstein features that willCommunity, IoT, Marketing, Sales andwork “out of the box,” according to companyService clouds. Some capabilities will beexecutives. Salesforce also plans to bringno-cost enhancements, but the majority ofEinstein services and APIs to its App CloudEinstein features will be extra-cost options.development platform so customers can build 2017 Constellation Research, Inc. All rights reserved.7
Figure 3. Salesforce Einstein Features Available or Planned for Release in 2017CloudEinstein FeatureOriginAvailabilityAnalytics CloudAutomated Analytics & StorytellingBeyondCoreGAContinuous User FeedbackBeyondCoreFebruary 2017Personalized Story SharingBeyondCore2017Predictive Wave AppsInternal development2017Smart Data DiscoveryBeyondCoreGASmart Data PrepBeyondCoreFebruary 2017Smart Newsfeed For CRMBeyondCore2017Heroku Prediction IOPredictionIO2017Predictive Sentiment ServiceMetaMind2017Predictive Vision ServiceMetaMindFebruary 2017Commerce InsightsDemandwareFebruary 2017Predictive EmailDemandwareGAPredictive SortDemandware2017Product RecommendationsDemandwareGAProduct Recommendations (Store)DemandwareGAArticle InsightsInternal developmentGAAutomatic Topic CreationInternal developmentGACommunity SentimentInternal developmentGACompany HighlightsInternal development2017Einstein AnswersInternal developmentFebruary 2017Questions and Article AnswersInternal development2017Recommended Experts, Files & GroupsInternal development2017Recommended Group MembersInternal developmentGASession Tracking for AnswersInternal development2017App CloudCommerce CloudCommunity Cloud 2017 Constellation Research, Inc. All rights reserved.8
Figure 3 Continued. Salesforce Einstein Features Available or Planned for Release in 2017CloudEinstein FeatureOriginAvailabilityCommunity CloudTopic InsightsInternal developmentGATrending User and Group ScoresInternal development2017Unanswered QuestionsInternal developmentFebruary 2017Automated IoT Rules OptimizationInternal development2017Predictive Device ScoringInternal development2017Recommended Best Next ActionInternal development2017Predictive Content & Product RecommendInternal developmentGAPredictive Scoring & AudiencesInternal development2017Automated Send-Time OptimizationInternal development2017Image Insights in Social StudioMetaMind2017Account HealthInternal development2017Account InsightsInternal developmentFebruary 2017Automated Activity CaptureInternal/RelateIQFebruary 2017Opportunity InsightsInternal/ImplisitFebruary 2017Predictive ForecastingInternal development2017Predictive Lead ScoringInternal developmentFebruary 2017Recommended ConnectionsInternal development2017Automated Case ClassificationInternal development2017ChatbotsInternal development2017Recommended MacrosInternal development2017Trending IssuesInternal development2017ContinuedIoT CloudMarketing CloudSales CloudService CloudSource: Constellation Research, based on Salesforce data 2017 Constellation Research, Inc. All rights reserved.9
custom smart applications. Salesforce says thisprobably have a good sense of which leads tocapability will be a point-and-click affair aimedprioritize. Salesforce says its machine-learning-at developers, with no requirement for deepbased lead-scoring engine needs at leastdata science talent.150 converted leads per month for accurate,automated scoring. Similarly, the BeyondCoreThe first App Cloud Einstein service availableengine can spot complex correlations andwill be the Einstein Predictive Vision Service,patterns across as many as 100 columns ofdue in February and based on technology fromdata, but it requires at least 10,000 rows ofthe MetaMind acquisition. Demonstrateddata to deliver statistically reliable results.at Dreamforce, the vision engine was shownto be easily trainable by business users byIt’s when data volumes are overwhelming thatdragging and dropping collections of images.Einstein features will make the most sense. ButOne MetaMind (and now Salesforce) customer,that’s not to say that Einstein is geared only toa company focused on providing virtualizedbig companies. Data volumes depend on theradiologist services, is using vision servicesapplication. Many small marketing teams, forto save lives by reviewing thousands of brainexample, send out millions of emails per month.scans within seconds to spot and helpPredictive Scoring & Audiences and Automateddoctors prioritize cases of life-threateningSend-Time Optimization, two features cominginter-cranial bleeding.to the Marketing Cloud this year, might makesense even for small companies, so long as theyConstellation’s Analysis: Einstein featuresare marketing at high scale.won’t be a fit for every company. For starters, ittakes lots of data to drive automated, machine-Strengths and Challengeslearning-based predictions. If you are dealingwith fewer than 100 leads per month, EinsteinIt’s the end of the beginning for SalesforcePredictive Lead Scoring would be overkill,Einstein, but long-term success will dependas humans can handle the load and theyon the performance and ease of use of the 2017 Constellation Research, Inc. All rights reserved.10
applications, packaging and pricing, andchicken-and-egg situations in whichcompetitive alternatives. All of the aboveaccuracy might suffer until a sufficientwill hinge in large part on the four underlyingnumber of customers are using particularrequirements for AI success discussed earlier:Einstein capabilities?data, computing capacity, data science andtime. Data science talent and time spent inAnother question is the scale at whichdevelopment are strengths for Salesforce,Salesforce will be able to tap into contextuallywhereas data and computing capacity may berelevant enrichment data. Salesforce ownschallenges. Constellation’s analysis is below.Data.com, the business-to-business-centricdata source developed out of the 2010Dataacquisition of Jigsaw. But Data.com is nota consumer-centric data source and it can’tWith its multiple application clouds, 100,000-match the scale of data sources ownedplus customers and millions of users, Salesforceby rivals, such as IBM’s WeatherChannel,sits atop a massive trove of application data.Microsoft’s internet search, gaming assets,As explained earlier in this report, data isLinkedIn assets or Oracle’s Data Cloud (ODC).used to test and train algorithms and models,Salesforce says it plans to tap into its Pardotso the more data available, the better. Theconsumer engagement data in a future release.question is to what degree Salesforce canEnrichment data contributes to contextualtap into customer data in an abstracted wayunderstanding of customers and builds thewithout crossing data-ownership or privacynetwork graph so that data science can spotboundaries. Abstracted benchmarks will becorrelations and patterns.important in understanding behaviors anddeveloping other insights. Einstein customersConstellation’s Analysis: Data scale and accesswill undoubtedly grant access for privacy-are likely to be challenges for Salesforce if andand security-protected analysis of their ownwhen competitors present real competition todata, but to what degree will Salesforce faceEinstein. Constellation suspects this was one 2017 Constellation Research, Inc. All rights reserved.11
motivation behind the company’s abandonedper kilowatt hour. There’s no doubt that AWSattempt to acquire Twitter in 2016. It should becan beat any one of its competitors by thisnoted that rivals may also face data-ownershipmeasure with its economies of scale, but toand privacy boundaries when attempting towhat degree will it cut Salesforce in on theseuse data in new ways tied to AI applications.cost advantages? Will AWS wholesale ratesenable Salesforce to undercut its competitors?Computing CapacityThe cost of computing capacity will obviouslyhave a huge impact on Salesforce EinsteinWith its comparatively small cloud computingpricing and competitiveness.capacity, Salesforce (with fewer than 10 datacenters globally) on its own would face aData Sciencecompetitive disadvantage compared to a cloudgiant like Amazon Web Services (with moreThis is a strength for Salesforce as it hasthan 50 data centers) or Microsoft (with morebeen able to attract top data science and AIthan 30). This is likely why Salesforce selectedtalent. More than two years ago, for example,Amazon Web Services (AWS) as its strategic,Salesforce lured away Vitaly Gordon fromlong-term cloud computing provider in 2016.LinkedIn to serve as its Vice President of DataAWS computing capacity is many timesScience. Together with former KXEN veteranlarger than that of all its largest competitorsJohn Ball, Senior Vice President and Generalcombined. Thus, Salesforce may have laid thisManager of Einstein, Gordon helped developconcern to rest with its AWS partnership,the data-management and machine-learningalthough it all depends on the terms it struckbackbone that now powers Einstein. Salesforcewith AWS.has picked up yet more talent throughacquisitions, including deep learning expertConstellation’s Analysis: At the industrial scaleRichard Socher, Ph.D, formerly MetaMind’sof the largest public clouds, the competitionCEO and now Salesforce Chief Scientist.comes down to the cost of computing capacity 2017 Constellation Research, Inc. All rights reserved.12
Constellation’s Analysis: The list of namesConstellation’s Analysis: Algorithms needabove just scratches the surface of Salesforcetime to improve. Data set gathering requiresdata science expertise and assets. Add to thetime for better precision. More interactions inlist the personnel and acquired intellectualthe network depend on time. Between earlyproperty of BeyondCore, Demandware andhiring and data-management work and itsits CQuotient acquisition, PredictionIO,numerous acquisitions, Salesforce has an edgeand RelateIQ to cite a few more examples.on time invested in AI, bested only by IBM, andSalesforce fares better than most enterprisematched, perhaps, only by Microsoft among itssoftware companies at retaining talent, sodirect competitors.the key challenge will be pulling together andconsolidating the best-of-breed technologiesInvestment of time is the biggest gap betweenand best practices from various acquisitions.where Oracle Adaptive Intelligent Apps areat this writing and the broad portfolio of appsTimeenvisioned on the company’s roadmap. Thecustomer-experience-related apps will launchSalesforce began work on building anfirst because that’s where the company canautomat
Capture, a Sales Cloud Einstein feature due in February, will automatically log sales calls, email exchanges and chat interactions so . Robust recommendation engines that take the output of AI and present choices that accelerate decision making INSIDE SALESFORCE EINSTEIN
