Skip to content

Usage, audit and feedback

AI usage, tool invocation, saved business proposals and product feedback are separate records. They answer different questions: what a model call consumed, which authorized tool was used, what a human reviewed and which unsupported capability was requested. None supplies a general permission to perform a new business operation.

Record or mechanism Recorded purpose Interpretation limit
ai_usage_logs Feature/provider/model, actor/company, token counts, duration and success/safe error code Usage observation, not a complete business-outcome receipt
MCP ai_events Tool/client, resource identity, output section names, response bytes, missing-info labels, contract versions and latency Audit metadata, not a copy of business results or proof of their completeness
Proposal revisions and decision result Reviewed content/version and confirming actor/outcome Family-specific approval evidence; no automatic customer acceptance
Product feedback/capability request Request reference, desired work and tools that fell short A received request, not an enabled feature or expanded authorization

The MCP audit helper intentionally records output sections, counts and labels instead of response values. Provider failures use enumerated safe diagnostics where implemented. Logging policy still applies to callers and error handling; a metadata-only helper does not certify every log line in the system. Information handling owns the broader content boundary.

The MCP business gate applies per-person daily and per-minute tool-call counts from that surface’s events. It does not count every in-app assistant/model action as the same quota. Counting and execution are separate operations; concurrent calls can race. Missing counts or counting errors fail open in this defense-in-depth check. Normal capability/resource checks remain required.

Model usage helpers expose a configurable daily token ceiling and usage summaries; these are not a universal atomic admission mechanism around every model call. The document-intake worker has its own per-company daily interpretation reservation and a precheck before decryption/model work. A budget pause can leave an eligible candidate waiting while manual review remains available. Do not diagnose a full daily budget as source corruption or bypass it with another technical job.

When no supported tool covers the requested outcome, request_capability records what the user wanted, why available tools were insufficient and related structured context. Its received/reference result does not mean the feature was built or that an agent may improvise a database write. The administrator capability-request view requires its own company permission (tenant:update on the reviewed path).

The product feedback workflow distinguishes bugs, improvements and view requests. An acute bug can create incident/notification work; the internal alert contains its reference rather than the report text. Notification delivery is separate from saving the feedback. Do not promise an alert was delivered merely because the feedback row exists.

If synthetic Tool Call A returns a payment section with missing coverage, the audit can record that section’s absence and response size. It must not copy bank/payment payloads into general logs. If synthetic Request B asks for a tool that does not exist, return the recorded request reference and the actual received status; do not report completion of the requested business action.

For a refused call, distinguish authentication, company selection, missing capability, quota and provider availability before deciding on another attempt. For a generated draft, inspect its source/revision and human confirmation rather than inferring approval from token usage. The MCP interface and proposal concept own those contracts.

The delivery PR names newly executed synthetic checks and reused evidence. This review does not send report text, archived mail or other private context to a provider, and does not certify complete metering or tamper-proof audit retention.