The AI agent has stopped drafting

Symptom: New conversations land in the inbox but the AI agent is silent. No draft, no auto-send, no auto-label. Maybe it worked yesterday and stopped today.

The AI has more knobs than most parts of Ochre, which is good for control and bad for triage. Walk these in order.

Checklist

1. Confirm the mode

Open AI → Drafting. Look at the current mode.

• Off means no drafts, ever. Someone may have toggled it during an incident. See drafting modes.
• Suggest means drafts appear in the composer but never send.
• Autopilot means the AI sends without a human in the loop, subject to confidence thresholds. See autopilot.

If you expected autopilot and you are seeing suggest, that is your answer.

2. Check confidence thresholds

If autopilot is on but nothing is auto-sending, the confidence threshold may be set too high. The AI is drafting, then declining to send because it is not sure enough. Lower the threshold gradually and watch AI receipts for the cutoff.

3. Check spend caps

Open AI → Spend. If you have hit a daily, weekly or monthly hard cap, the AI will stop drafting until the window resets or you raise the cap. This is the single most common silent failure. See spend caps.

4. Check BYOK keys

Open AI → Keys. If you brought your own key (Anthropic, OpenAI, etc.) and that key is now invalid, expired, or out of credit, drafting halts. Look for a yellow or red status indicator on the key card. Rotate the key, or fall back to Ochre-managed credits via AI keys fallback.

5. Check guardrails

Open AI → Guardrails. Two things to verify:

• The conversation may be tagged with a bypass label that tells the AI to stay out. See guardrails.
• A topic guardrail may be matching the customer's message and silently declining to draft. The receipts panel will show this.

6. Check the receipts

Open the silent conversation and look at the AI receipts panel. Receipts always tell you why a draft did or did not happen. Common reasons:

• "No matching brain context" means the AI does not have anything to say. Improve the AI brain. Note: a previous bug returned zero KB chunks even when articles were ingested, which made the AI hallucinate or stay silent. Fixed in migrations 0127, 0130, and 0131. If you saw this before, retry now and check receipts.
• "Confidence below threshold" means lower the bar.
• "Bypass label active" means a teammate or rule applied a guardrail label.
• "Provider error" means a key or upstream model issue.

A 60-second sanity check

Open the AI playground and paste a representative customer message.

• If the playground produces a draft, the AI itself is working. The issue is in mode, threshold, caps, or guardrails on the live conversation.
• If the playground also fails, the issue is keys or upstream availability.

When to contact support

Contact us if:

• The receipts panel shows "Provider error" repeatedly even after rotating your key.
• Spend caps are not exhausted, mode is correct, no guardrails apply, and the playground also fails.
• The AI is drafting on some conversations but not others with no obvious pattern.

Send a conversation ID, a screenshot of the receipts, and your current mode and threshold. The fastest path is in support contact.

Related reading

• AI overview
• Drafting modes
• Confidence thresholds
• Spend caps
• AI receipts