Salesforce Einstein spans multiple CRM AI features rather than a single freely interchangeable chatbot. The most important buyer question is what data the particular CRM feature actually reads and what its security controls cover.
Product identity and supported workflows
Salesforce documentation describes the Einstein Trust Layer for generative AI with permission-grounded retrieval, audit trails and configurable masking. Critically, the official limitations page states that LLM data masking is disabled for agents, a detail absent from many high-level promises of secure AI.
Who it is for and the decision it addresses
A support supervisor might use an embedded summary draft or a sales manager may request a short account-context write-up. Such summaries need the source record checked for permission, date and factual accuracy before they are sent to a customer.
Step-by-step first evaluation
Work with a sandbox and controlled mock CRM fields. Enable only a supported feature, inspect the user’s field-level permissions and run a prompt that includes known dummy sensitive fields. Compare the masked input and output in documented auditing tools; repeat a test against a different feature rather than assuming one trust control applies universally.
Failure modes and limitations
Automatic masking can miss uncommon names, local data patterns and information held in unstructured fields. Some features use different processing paths. Regulatory compliance depends on the full organization configuration and contract, not simply turning on a security toggle.
Privacy, permissions, licensing and governance
Audit trails themselves may contain sensitive information, including hydrated prompts and results. Restrict administrator access and define a retention policy. The vendor explicitly warns that pattern masking is not perfect and cannot be generalized to every Agentforce function.
Exports, retention and avoiding lock-in
Preserve CRM records and approved reports in the configured system of record. Exporting the generated text does not preserve its security context, versioned prompt configuration or permissions; document those separately when auditing decisions.
Alternatives and team boundaries
monday AI offers board-centered assistance; Intercom Fin is customer-facing service. Detailed Agentforce automation belongs to the agent-security and orchestration guide, not this general product registry.
What remains unverified for each country, version and plan
This October 25, 2026 profile reflects dated first-party product documentation, not independent hands-on testing or an endorsement. Exact country availability, model versions, accessibility provisions, seat entitlements, enterprise residency, retention settings and the whole supported language matrix vary by configuration. Where a plan/region/date combination was not verified with its exact original source, it is left unknown, not assumed worldwide. The product may have changed since the documented feature appeared.
Pricing and evidence boundary
No subscription price, usage allowance, API token rate, special promotional plan, success benchmark or commercial license is asserted here. A dated pricing reference covers independently time-stamped pricing snapshots. Product brochures are supplier claims; human reviewers must check linked evidence against the actual source and retain their own decision trail. No screenshots, borrowed brand assets or pretend product tests are included. Specialist model, media, coding and autonomous-agent deep dives belong to their designated teams.
Sources & Further Reading
Primary vendor documentation accessed 2026-10-25; market-specific terms and feature rollout require additional verification. These sources support stated vendor-product claims, not an independent quality score.