Einstein 101 - Salesforce AI Demystified

Fourth Industrial Revolution is here! And, thanks to Industry 4.0 (also known as 4IR) we are experiencing a smarter and connected world than ever before. Going beyond the industry jargon: we have seen how computer and internet with technology advancement had changed our life in past decades. This digital revolution is now taken to next level with AI, Robotics, Blockchain, Quantum computing, IOT, 3D printing, Autonomous vehicles, Virtual reality and so on. As consumer our day to day experience is upgraded and taken to another level.

Artificial Intelligence (AI) and Machine Learning (ML) are key cogs in this revolution. Your personal phone assistant (Siri on iOS, Google assistant on Android, Alexa), effective recommendations from your shopping apps or video streaming apps, GPS navigation to check best available route, Gmail smart composer and on. We have been using AI in our daily life knowingly or unknowingly for quite some time now.

While for consumers its the best ever experience, its not so easy for businesses. As consumers are expecting smarter, unified, connected experiences, businesses need to step up to power their processes with latest technology advancement including AI. And, implementation of AI is always complex and time consuming. Beginning with identification of right business use case to collecting & preparing appropriate data sets, examining data to create patterns & correlations, model generation & testing, computing power, appropriate skills - it takes quite an investment of money and time. Its get huge when a given business has multiple use cases and this investment is to be multiplied by n times.

Here comes Salesforce Einstein - the AI offering from Salesforce.com ecosystem. Salesforce AI focuses specifically around customer experience related scenarios so that businesses can have smarter business applications easily and faster.

Salesforce Einstein offers ready-to-use AI capabilities as extension to the Salesforce SaaS clouds. On-demand capabilities are built into the proven Salesforce Lightning platform. This leverages data from existing Salesforce ecosystem to spot correlations & patterns and combines seamlessly with other platform capabilities such as process builder, flows, workflow rules etc. Based on the given business problem, external data sets can also be used.

In the background, Salesforce deploys an automated machine learning framework which takes care of all necessary steps such as data enrichment, feature engineering, model selection, score calibration, training and develops unique predictive model for each Einstein customer. With this automation, customers don't necessarily need deep data science skill anymore to implement AI in respective businesses. Thanks to Salesforce Einstein - AI is for everyone!

Salesforce started this journey years back. But it achieved the breakthrough with acquisition of RelateIQ (a Sales automation company, now offered as SalesforceIQ) in 2014. More internal research & development and acquisition were followed. This included Tempo AI (Smart calendar assistant) in 2015, PredictionIO (machine learning startup), MetaMind (Palo Alto-based AI Company), BeyondCore, and Implisit Insights (Tel Aviv-based AI company) in 2016. It's in 2016 Dreamforce, Salesforce launched Salesforce Einstein, the AI offering from Salesforce.com. And, evolution through internal research and acquisition continued - with each release, we have better, easier, richer Salesforce Einstein capabilities.

Salesforce Einstein capabilities can be categorized into four high level groups.

• Out of box Einstein Apps (Sales cloud einstein, Commerce cloud einstein)
• myEinstein - Point and Click
• myEinstein - API services
• PredictionIO based custom model

Salesforce Einstein offering at a glance

Out-of-the-Box Apps

Built-in intelligent features with each Salesforce SaaS cloud covers most common AI use cases for respective business areas. Driven by the automated machine learning framework, these features are offered as out-of-the-box apps.

Sales Cloud Einstein - Features include: Einstein Lead & Opportunity scoring (to prioritize the most likely ones to convert), Einstein Account & Opportunity insights (to observe key developments on the accounts & opportunities), Einstein Forecasting, Einstein Activity capture & Einstein Automated contacts (to capture data and add new contacts automatically), Einstein Recommended connections (to retrieve insights about team’s network to identify who knows the customers and can help out on a deal). This app needs additional license fees (per user per month).

Service Cloud Einstein - Einstein Bots, Einstein Case Classification features are part of this app. And, more features are on the way. At this time Service Cloud Einstein is included with live agent chat feature license or unlimited edition licenses.

Marketing Cloud Einstein - Offered at additional license fees (subject to specific capabilities). Features include - Einstein Social Insights (analyze social conversations to get deeper insights about customers), Einstein Recommendations (suggest the next best product, content or offer), Einstein Engagement Scoring (predict customer's likelihood to engage with emails or convert on the web), Einstein Splits (send customers to right journey builder path based on engagement score or persona), Einstein Segmentation (From Salesforce DMP - Data Management Platform), Einstein Journey Insights (From Salesforce DMP), Einstein Vision for Social Studio (gather insights from customer's images on the social web), Einstein Lookalikes(From Salesforce DMP) etc.

Commerce Cloud Einstein - Features include: Einstein Product Recommendations (setup personalized recommendations for each shopper), Einstein Search Recommendations (personalized type-ahead search), Einstein Predictive Sort (tailored product sorts based on customer data), Einstein Commerce Insights (to transform customer data into actionable merchandising insights). Included with commerce cloud license.

Community Cloud Einstein - Einstein Recommendations (serves the most relevant and tailored content), Einstein Experts (to find certified experts on the topics that matter most to them), Einstein Feed Insights (to display the most interesting discussions intelligently), Einstein Feed Search. Licenses are included with Community cloud.

myEinstein - Point and Click

More than 80% of the records in Salesforce are on custom objects. And, every business has a different problem to address. Hence, Einstein allows to create customized AI model on any Salesforce object or field to predict business outcomes. There are several capabilities in this group and few are yet not available generally. But at the core, Salesforce Einstein offers automated machine learning framework which creates customer specific models; these models are self-learning and improve the more they are used.

Prediction Builder (beta) - Custom AI model for the given business scenario. Step-by-step wizard to configure prediction model through point-and-click. This supports Binary classification (meaning answers business problem in Yes or No) and Regression (answers business problem in numeric fields). The predicted outcomes can be stored in appropriate custom fields and can drive other platform based actions/automation.

Einstein Next Best Action (pilot) - Combines platform capabilities such as Recommendation object, Strategy builder, Flows with predictive models and provides optimal intelligent recommendations at the point of maximum impact. The given predictive models can be designed using Salesforce and non-Salesforce data. This feature can be used with or without AI capability.

Einstein Bots - Both standard and AI powered chat-bots are supported by Salesforce. Intelligent chat-bots leverage Natural Language Understanding (NLU). NLU trains chat-bots to create a learning model.

Einstein Voice (pilot) - New kid in the block. At this time, the Voice capabilities are offered in two folds: Einstein Voice Assistant (for employees) and Einstein Voice Bots (for customers). Einstein Voice leverages Automatic Speech Recognition (ASR) and NLU technologies in the background.
Audio frames are converted into sequences of text by ASR model and then, NLU model works on these unstructured data inputs to convert them into machine friendly structured form. Intents and entities are identified and mapped to appropriate Salesforce actions.

Einstein Discovery - This analytics cloud feature helps to find important correlations and make accurate predictions based on Einstein Analytics datasets. This feature can run on organization wide data (Salesforce as well as non-Salesforce) - including encrypted information to answer critical questions - such as What happened, Why, Predictions & Improvements, in context of the given business problem.

myEinstein - API Services

In addition to out-of-the-box app features and point-and-click platform features, API services are available for developers to create and train their own custom model and use the same as 'service' to predict relevant business outcomes. With APIs, these prediction services can be made available for business applications as easily as any other service.

Salesforce Einstein platform offers the following types of API services at this time:

• Einstein Vision - Image classification and Object detection
• Einstein Language - Intent and Sentiment analysis

At high level this group of Einstein features work in the following steps:

0) Prerequisite - Registration of Einstein Platform services, Download key file (.pem), Token generation

1) Define/create label/classes - This is the first step for any AI implementation with platform services. Identify the problem business needs to address and accordingly define appropriate label/classes. For example, (in context of Intent API services) we can create labels like sales, information, complaint for analyzing the incoming tweets from end customers.

2) Gather relevant data, prepare/cleanse, and upload dataset - In the next step, appropriate dataset need to be identified which can be used to spot correlations and patterns. This identified dataset to be cleansed as appropriate and uploaded to the Einstein platform using 'upload' api (…/vision/datasets/upload).

curl -X POST -H "Authorization: Bearer <token>" -H "Cache-Control: no-cache" -H "Content-Type: multipart/form-data" -F "data=@<zip file path>" -F "name=mercedesvsbmw" -F "type=image"  https://api.einstein.ai/v2/vision/datasets/upload

This API can be consumed as REST service from apex or any other programming constructs.

3) Train the model - Once dataset is successfully uploaded, this is used to train the given predictive model. API (…/vision/train) is used to train the model/engine.

curl -X POST -H "Authorization: Bearer <token>" -H "Cache-Control: no-cache" -H "Content-Type: multipart/form-data" -F "name=mercedesvsbmw" -F "datasetId=<dataset id>" https://api.einstein.ai/v2/vision/train

Training is very important in the whole scheme and better & relevant data used in the training - better the predicted outcome is.

4) Make prediction - Once the model is trained successfully, it is ready for use to make prediction based on specific inputs. Once again, developers can use API service to request for prediction (…/vision/predict).

curl -X POST -H "Authorization: Bearer <token>" -H "Cache-Control: no-cache" -H "Content-Type: multipart/form-data" -F "sampleContent=@<image path>" -F "modelId=<model id from last step>" https://api.einstein.ai/v2/vision/predict

5) Feedback, Retrain, and Improve - Once the service is deployed and in use, the predicted outcomes can be assessed, corrected/enriched (as needed) and feed into the model through retrain/feedback API (…/vision/feedback).

curl -X POST -H "Authorization: Bearer <token>" -H "Cache-Control: no-cache" -H "Content-Type: multipart/form-data" -F "modelId=<model id>" -F "trainParams={\"withFeedback\" : true}" https://api.einstein.ai/v2/vision/retrain

This allows the model to improve its performance over the time.

PredictionIO based Custom Model

This fourth variation is not directly part of the Salesforce Einstein pack. This is more of custom AI implementation with or without Salesforce platform. This is useful if business needs to implement custom prediction model, custom algorithm, combination of more than one algorithms and customization in all elements. In this, the custom AI model is deployed using PredictionIO framework.

Apache PredictionIO® is an open source Machine Learning Server. PredictionIO's Event Server continuously collects data from different application(s). It can also unify data coming from multiple platforms. The PredictionIO engine then builds predictive model(s) with one or more algorithms using the data. Once deployed as a web service, it listens to queries and respond with predicted results in real-time.

PredictionIO template gallery offers number of engine templates which can fast track the implementation. Templates consist of D-A-S-E components ([D] Data Source and Data Preparator, [A] Algorithm, [S] Serving, [E] Evaluation Metrics) and are fully customizable for given use cases.

What next

Einstein has emerged as the key and focused offering from Salesforce.com stack for obvious reasons. This offers AI capabilities in an integrated and user-friendly manner. As next steps:

Visit the web sessions and demos at our YouTube channel. (Coming soon!!!)

Earn some badges (includes hands-on) on Trailhead.

Explore Salesforce Einstein documentation