Conversational product discovery can reduce the work a shopper does to translate a goal into catalogue vocabulary. It can also produce confident but unusable recommendations when the catalogue, pricing rules, inventory, or eligibility data are weak.
Salesforce announced the agreement to acquire Cimulate in February 2026 and, according to the editor’s note on that announcement, completed the acquisition on March 3, 2026. Completion does not itself establish general availability, contractual entitlement, or behaviour in a specific tenant. Verify those points in current Salesforce documentation and your agreement before designing around them.
Start with the shopper decision
“Improve discovery” is too broad for an architecture decision. Choose a bounded decision such as:
- identify a compatible replacement part;
- narrow a product family by material, environment, and budget;
- compare products against named constraints;
- recover from a zero-result query;
- explain why an item is ineligible for a destination or customer type.
For that decision, define the expected output and the conditions under which the system asks a clarifying question, declines to recommend, or hands off.
Catalogue quality remains the foundation
A language model may interpret intent expressed in natural language. It does not create missing product truth.
The evaluation should identify the authoritative source for:
| Data | Typical risk if it is weak |
|---|---|
| Product description | Vague or invented compatibility |
| Structured attributes | Incorrect filtering or comparison |
| Taxonomy | Inconsistent retrieval across categories |
| Price and promotion | Stale or ineligible offer |
| Inventory and lead time | Recommendation that cannot be fulfilled |
| Customer entitlement | Exposure of an unavailable product or price |
| Compliance constraints | Unsafe recommendation or missing disclosure |
If those sources disagree, the conversational layer needs an explicit precedence rule. “Let the model decide” is not a governance policy.
Data 360 is conditional
A catalogue-only use case may not require a unified customer profile. Data 360 becomes relevant when the approved scenario depends on governed identity resolution, cross-system order history, segments, calculated signals, or activation across channels.
Before adding it, document:
- which cross-system facts change the recommendation;
- the freshness required for each fact;
- the legal basis and consent boundary for personalisation;
- how deletion, correction, and access controls propagate;
- what the experience does when identity resolution is uncertain.
This turns “use Data 360” into a testable architecture decision rather than a default dependency.
Keep deterministic controls outside free-form Instructions
Conversational reasoning should not become the source of truth for:
- price calculation;
- tax;
- promotion eligibility;
- inventory reservation;
- regulated product restrictions;
- payment authorisation;
- returns policy exceptions.
Expose those decisions through bounded Actions or services with explicit inputs, outputs, permissions, and errors. The agent can explain the result, but the deterministic system should calculate or enforce it.
Compare with the existing search experience
A replacement decision needs local evidence. Build a versioned evaluation set from real, privacy-safe query patterns and include:
- successful existing queries;
- zero-result queries;
- reformulations;
- ambiguous goals;
- conflicting constraints;
- unavailable products;
- policy-sensitive requests;
- multilingual variants where they are in scope.
Compare at least:
- relevance of the returned products;
- factual accuracy of explanations;
- constraint compliance;
- latency;
- fallback quality;
- escalation or handoff quality;
- conversion or task completion measured on the bounded flow.
Do not claim a conversion uplift before the controlled comparison exists. Search quality and commercial outcome are related, but they are not the same measure.
Availability checklist
Before committing architecture or budget, confirm in writing:
- the current product and feature name;
- availability status and region;
- compatible Commerce Cloud edition and storefront architecture;
- required Agentforce or Data 360 licences;
- catalogue connectors and supported Actions;
- data retention and logging behaviour;
- usage metric and price in the actual contract;
- rollback path if the evaluation fails.
Seasonal roadmap labels such as a past release’s beta or forecasted GA should not remain in a current decision document. Date every availability statement and replace it when the source changes.
Decision rule
Proceed when a bounded product-discovery decision has authoritative catalogue data, deterministic commerce controls, a test set, a measurable baseline, and a safe fallback. Expand only when the evidence improves for the next product family or channel.
Conversational discovery is not a reason to replace every existing search component. It is a hypothesis to test against the customer decision and the operational truth already in the commerce estate.
