Secured AI access: AI that receives only the context the user is allowed to use

Resolve the authenticated tenant user, active account, DataRoles, locations, and AI-read capabilities.

Use secured-record functions to produce only records that the user may provide to AI.

Pass approved location identifiers and the all-locations flag into semantic and structured retrieval.

Exclude or separately authorize encrypted and restricted fields.

Build prompts, retrieved chunks, and structured record summaries from the filtered set.

Embedding inventory, chunking strategies, vector dimensions, query embeddings, and reranking support controlled knowledge retrieval.

Insights are tied to source modules and records, include confidence, severity, evidence, recommended actions, and lifecycle status.

AI can propose structured actions, but application authorization and workflow rules govern execution.

Interactive chat and agent sessions can use CRM contacts and record context only when the owning user has AI-read authority.

Worker lanes route embedding, reranking, insights, and other workloads according to capability, health, load, and hardware profile.

AI queries and resulting actions can be recorded with tenant, user, source record, model, and operational context.

Insights preserve evidence summaries and source-context hashes for review.

Suggestions, acceptance, modification, rejection, execution, and output can be tracked.

AI calls through isolated appliances can be logged with request and response hashes, duration, endpoint, and disclosure information.

The client can request an AI operation, but it cannot widen the backend security envelope.

Ordinary read access and AI-read access are separate decisions.

Generated output cannot perform a protected write, export, billing action, or approval without the required application permission.