Ask AI

A chat over your decision record that answers with citations, shows its working, and runs on your own model account.

Bring your own model

Precedent supplies no AI accounts. A workspace admin connects one under Settings, Integrations: Anthropic, OpenAI, or OpenRouter, with the model you choose. The key is verified against the provider before it is stored, encrypted at rest, and never shown again. Your key, your bill, your data.

Open Settings, Integrations and choose your vendor.
Paste an API key and a model name. A wrong key or model is rejected on the spot, not discovered weeks later.
The same account powers the chat and the reading of captured Slack and Teams threads.

There is deliberately no deployment-wide key and no environment-variable fallback, so one workspace’s questions can never run on another’s account.

Answers you can check

Ask AI runs the same read tools an MCP agent gets, as the signed-in person, so it can never surface a decision you cannot see. Every claim in an answer cites the decisions it came from, and the tool-call trace (which tools ran, with what arguments) is shown alongside the answer rather than hidden.

With no account connected, the chat still works: a deterministic stand-in walks the record with the same tools and says plainly that it is a script, not a language model.

Cost control

Every model turn is metered: tokens, tool calls, duration, and whether it succeeded. What was asked and answered is never stored. Thread extraction and chat share one daily cap per workspace, so a busy Slack cannot quietly spend the allowance before anyone opens the chat. Usage is on the Ask AI page.