AI agent overview

Ochre ships with an AI agent that reads every incoming conversation, classifies it, looks up the right context from your knowledge base, and either drafts a reply, sends one, or stays out of the way. It is not a separate product. It is a teammate that lives inside the same inbox your humans use.

This article explains what the AI does end to end, what it costs, and how to think about the controls.

What it does on every new message

1. Reads the message and the prior thread.
2. Classifies topic and priority, and writes those to the conversation as labels. See Auto-labeling: topic, priority, confidence.
3. Pulls the most relevant articles from your indexed knowledge base. See The brain: KB graph.
4. Drafts a reply in your workspace voice and length settings.
5. Decides what to do with that draft based on your reply mode and confidence floor.

Every step is logged. See AI receipts.

Three reply modes

You pick how visible the AI is:

• Auto-send. AI replies on its own when confidence clears your floor.
• Draft. AI puts the reply into the agent's editor. One click to send, edit freely first.
• Suggest. AI shows its draft beside a blank composer. Agent writes their own reply, with the AI as a reference.

You can mix modes per channel. Read Auto-send vs draft vs suggest.

Confidence is the steering wheel

Every AI reply carries a confidence score from 0 to 100. You set the floor. Below the floor, auto-send falls back to draft. Below an even lower floor, the AI stays silent and the ticket goes to a human untouched. See Confidence thresholds.

Bring your own model

Ochre is BYOK, on Anthropic and OpenAI. You connect an Anthropic key, an OpenAI key, or both. You pay the model provider directly. We do not mark up tokens. See BYOK: bring your own Anthropic or OpenAI key and Choosing a model.

If a provider is rate limited or has an outage, traffic fails over to your other key automatically. See Multiple keys + automatic fallback.

Gemini, Mistral, and self-hosted models are not supported today.

What it costs

Ochre's seat price is flat. AI usage is separate, paid to your model provider.

Per-resolution cost is roughly $0.005 to $0.01 on GPT-4o-mini or Claude Haiku 4.5, and proportionally more on the flagships. Compare that to the $1.50 to $2.00 per resolution that incumbent helpdesks bill for AI. Every reply shows tokens in, tokens out, and the dollar amount in its receipt. Set a monthly cap and alerts in Spend caps and alerts.

Voice, length, and guardrails

Three knobs shape the writing:

• Voice. Pick a tone and feed it sample replies. See Voice and tone.
• Length. Cap how long replies get and how the AI handles long threads. See Reply length.
• Guardrails. Bypass labels (bug, abuse, spam), per-channel overrides, confidence floors, quarantine. See Guardrails and bypass labels.

Autopilot

When you flip Autopilot mode on, new conversations route AI-first and the reply mode jumps to auto-send. It is the difference between the AI assisting your team and the AI being the first responder with humans on review.

How the AI gets smarter

Two systems do the work of improvement:

• Learned Q&A library. High-confidence Q+A pairs are promoted from your past conversations and reviewer feedback. They go through a quarantine queue before the AI uses them on new tickets. See Learned Q&A library.
• Quality assurance review. Reviewers grade past AI replies. Their notes feed the learned library. See Quality assurance review.

Try before you trust

The Playground lets you paste any sample message and see exactly what the AI would draft, which articles it would cite, and what confidence score it would assign. Use it to tune voice, length, and guardrails before you turn anything on.

What humans still own

The AI does not:

• Process refunds or change account state on its own.
• Reply to anything labeled bug, abuse, or spam (default bypass list).
• Send messages below your confidence floor.
• Send to VIP segments unless you explicitly opt them in.

A human is always one click away, and the AI receipts trail lets you reconstruct exactly what happened on any conversation.

Recommended starting setup

If you are new, this is a sane default:

1. Connect both Anthropic and OpenAI keys for fallback.
2. Default model: Claude Sonnet 4.6.
3. Reply mode: draft.
4. Confidence floor for auto-send: 85%.
5. Bypass labels enabled for bug, abuse, spam.
6. Spend cap set to twice your expected first-month usage, with alerts at 50, 75, and 100%.

Flip Autopilot on once you have a week of receipts you are happy with.