12.3 Applying Predictive & Generative AI Tooling to Customer Scenarios

12.3 Applying Predictive & Generative AI Tooling to Customer Scenarios

Quick Answer: Section 1.4 covered why Data 360 underpins AI. This section covers what you configure. AI Models — the workspace formerly called Einstein Studio — is where a consultant creates predictive models from scratch, connects external models with bring your own model (BYOM), connects external foundation models with BYOLLM, and builds retrievers over search indexes. Predictions come back into Data 360 as data, which makes them usable in segmentation, calculated insights, Flow, and prompt templates. The exam tests scenario matching: given a business requirement, which tool, and what has to be true about the data first.

This is the "Apply predictive and generative AI tooling capabilities of Data 360 to customer scenarios" objective inside Data Enhancements, Sharing, and Analysis (18%).

---

AI Models: The Workspace

AI Models is the single place inside Data 360 for model management, regardless of model type or where the model runs. What a consultant configures there:

The zero-copy principle carries through to BYOM: the external platform trains against Data 360 data without ETL, and once deployed, predictions update continuously and land back in Data 360.

---

The Predictive Workflow End to End

Step 1 — Unify First

A predictive model trained on Individual__dlm learns from duplicates. Train on the unified profile. This is the single most consequential prerequisite, and it is why identity resolution (Chapters 8 and 9) precedes AI in every sensible implementation sequence.

Step 2 — Engineer Features Inside Data 360

The documented pattern is to flatten behaviour and history into a one-row-per-customer feature set using calculated insights and data transforms: spend in the last 90 days, order count in the last 12 months, sessions in the last 30 days, days since last case, average resolution time. Keeping this work inside Data 360 avoids external joins, keeps training fast and scalable, and keeps the data under in-platform governance.

Step 3 — Train

Either build from scratch in AI Models — select the source object, select the outcome field, let Data 360 train and score — or connect the external model with BYOM.

Step 4 — Operationalize the Prediction

This is the step scenarios actually test. Once a score is in Data 360, it is just an attribute:

• Segmentation. Filter the canvas on Churn Propensity > 0.7 to build a retention audience, then activate it to Marketing Cloud Engagement.
• Calculated insights. Reference the score in a further aggregation — average churn risk by region, revenue at risk by loyalty tier.
• Flow. A Data 360-triggered flow reads the score and creates a proactive task for the account executive.
• Prompt templates and agents. A grounded prompt can reference the score so an agent's suggested next step reflects the customer's risk.
• Enrichment. Copy the score onto the Account or Contact so ordinary CRM reports, list views, and validation rules can use it.

Scoring runs in batch or streaming mode; streaming scores records as data changes, which matters when the decision is being made during a live interaction.

---

The Generative Workflow End to End

Structured Grounding

For attributes that already live in the model, grounding is declarative: ground the prompt template on DMO merge fields using the two supported relationships — copy fields and related lists of DMOs on Salesforce objects (section 12.1).

Unstructured Grounding

For content that is not tabular — knowledge articles, contracts, manuals, call transcripts, and also audio, video, and images — the pipeline is:

1. Ingest the unstructured content into Data 360.
2. Map it to standard DMOs or unstructured data model objects (UDMOs).
3. Chunk it into meaningful passages.
4. Embed the chunks as vectors.
5. Index them as a vector, hybrid, or enriched search index.
6. Create a retriever in the AI Models tab over that index.
7. Reference the retriever from a prompt template or an Agentforce agent for RAG.

Choosing the index type is a real consultant decision. Pure semantic recall over conceptual questions suits a vector index. Content full of exact tokens — SKUs, part numbers, policy codes — suits a hybrid index, because pure semantic matching will drift off an exact identifier. Where recall quality is paramount, an enriched index generates three chunks per passage: the original text, a metadata chunk, and a chunk of questions the passage can answer.

---

Scenario-to-Tool Matching

---

Governance Obligations That Survive the Move to AI

A consultant who presents AI tooling without the governance frame has given an incomplete answer, and the exam rewards the complete one:

• Consent. A prediction about a person is personal data. A profile that opted out of processing through the Consent API processing action should not be feeding a model or an agent response, and a profile deleted with shouldforget must not survive as a model input.
• Data spaces. Models, retrievers, and search indexes are scoped. A brand's model does not silently read another brand's data.
• The Einstein Trust Layer sits between the grounded prompt and the model, applying protections such as masking and zero data retention with the model provider. It requires Data 360 enabled.
• Explainability and drift. A score that was accurate at launch degrades as behaviour shifts. Schedule retraining and monitor the score distribution the same way you monitor consolidation rate.

---

Exam Traps & Consultant Pitfalls

Trap 1: Modelling on fragmented data. Training on source-level individuals rather than unified individuals reproduces duplicates as features. Unify first, always.

Trap 2: Treating a calculated insight as a prediction. A calculated insight summarizes what happened. A model estimates what will happen. "Customers who have not purchased in 90 days" is an insight; "customers likely to churn in 90 days" is a model.

Trap 3: Using a retriever for structured attributes. If the value is a field on a DMO, ground with merge fields. Retrievers exist for chunked, embedded, unstructured content.

Trap 4: Forgetting where the score lives. A prediction is written back into Data 360 and is then usable in segmentation, insights, flows, and enrichments. Answers that propose exporting scores to a file and re-ingesting them are recreating a pipeline the platform already removed.

Trap 5: Choosing a vector index for exact-token content. Part numbers and policy codes need lexical matching too — that is what the hybrid index is for.