AI configuration
The chat assistant, AI summaries, and daily briefs run on a model that is managed by the platform. The provider, model, and API keys are set by whoever operates your UnitOps deployment through environment configuration, so there is nothing to choose, connect, or paste in the app.
What admins control in the app is the operating envelope: a connection check and spending budgets.
Page: Admin -> AI Config (/admin/ai-config) · Role required: admin
What you can change
Admin -> AI Config has a Summary tab and a Chat tab. AI summaries (dashboard narratives and daily briefs) use the Summary configuration, and the assistant uses the Chat configuration. On each tab you can:
- Test connection - run a live probe against the platform AI configuration. Success shows the round-trip time; a failure shows the provider error so you can escalate to your operator.
- Monthly spending budget - cap spend in US dollars for that feature. Leave it blank for unlimited. The app soft-warns at 80% and hard-blocks new requests at 100% of the budget (a blocked request returns a 429). Click Save to apply.
The provider, model, and API key are shown as managed by the platform and cannot be edited here.
Watch spend
Admin -> AI Usage (/admin/ai-usage) shows request and token volume for the organization against the budgets in force. Review it weekly for the first month after rollout, then set budgets that match your plan.
Quality controls
| Page | Use |
|---|---|
| Admin -> AI Conversations | Read real Slack and Teams transcripts; useful for spotting bad answers and missing data. Direct messages with the assistant are visible only to an owner or an admin |
| Admin -> AI Coaching | Adjust the guidance the assistant follows when advising operators |
How answers stay grounded
- Questions about metrics are answered from your KPI catalog and warehouse data, not from model memory.
- Generated SQL is parsed and validated as read-only before execution, and it must be scoped to your organization.
- Questions about policy or procedure are answered from your indexed documents, see Documents.
- Answers respect the asking user's location permissions.
Troubleshooting
| Symptom | Fix |
|---|---|
| "Test connection" fails | The platform AI configuration is unavailable or misconfigured; contact whoever operates your UnitOps deployment |
| Chat or summaries stop with a budget error | The monthly budget was reached; raise it in Admin -> AI Config or wait for the next month |
| Answers ignore your documents | Document search needs an embedding provider, which is also platform-managed, see Document storage and embeddings |
| Answers are slow | Reduce the amount of grounding data on the dashboard, or ask your operator about the configured model tier |