An AI that knows when to shut up

Every AI support tool sells you the same number. Seventy percent deflection. Eighty percent resolution. Pick a figure, put it on a slide.

Nobody sells you the other number: how often it's confidently wrong in front of a paying customer.

That number is the one that matters, because the costs aren't symmetrical. A question the AI declines to answer costs you a few minutes of someone's attention. A question it answers wrong can cost you a refund you never agreed to, a compliance problem, or a customer who screenshots your bot promising something you don't offer and posts it.

We built the agent around that asymmetry.

Guessing is the default behaviour

Language models don't have a mode where they don't know something. Ask one about your return policy and it will produce a return policy. Not yours — a plausible one, assembled from every return policy on the internet. Thirty days, original packaging, minus shipping.

It reads perfectly. That's the problem. There's no tell.

So the question isn't "how do we make the model smarter". It's "how do we know when to not use the answer it produced".

Three things, none of them clever

It reads your stuff, not the internet's. When a message arrives, we search your knowledge base — your help docs, your uploaded PDFs, the answers you pasted in — and put the matching passages in front of the model. If nothing matches, the model is told, plainly, that nothing matched.

It grades its own homework. The reply doesn't come back as text. It comes back as a structure: the reply, a confidence score, a short reason, an explicit "a human should take this" flag, and the ids of the knowledge passages it actually leaned on. That last field is the interesting one — you can check whether the answer came from your docs or from the model's general sense of how businesses work.

Low confidence doesn't send. Below your threshold, it escalates instead. The conversation lands in your inbox with the thread, the draft it would have sent, and the reason it stopped.

Is self-reported confidence perfect? No. A model can be confidently wrong about its own confidence. But paired with grounded retrieval, it's a strong enough signal to be worth acting on, and it's a great deal better than the alternative, which is no signal at all.

The escalation is the feature

Most tools treat handoff as failure. A number to drive down. Something to apologise for in the release notes.

We think handoff is the product working. If a customer asks something your docs genuinely don't cover, there is no good automated answer. There's a wrong one and a stall. Getting it to a person quickly, with the whole conversation attached, is the correct outcome.

Which is why the handoff carries context. The agent's attempt, its reasoning, the passages it found, the full thread. You open the conversation already knowing what happened. Nobody asks the customer to start over — which is the single most infuriating thing a support system can do.

Start it where it can't hurt you

The agent has three settings and we think you should use them in order.

Off. It's a shared inbox. Email and chat in one list, assigned to people, answered by people. A completely legitimate place to stay.

Suggest a draft. The agent writes a reply and leaves it in the thread. You read it, edit it, send it, or bin it. Nothing reaches a customer without a person deciding it should. This is where you find out what it's actually like — not from our marketing, from a fortnight of watching what it wanted to say to your customers.

Auto-reply. Confident answers go out on their own. Everything else still waits for you.

Most people should spend a couple of weeks in draft mode. Not because we're cautious by temperament, but because that's how you build a calibrated sense of when it's good — and calibration is the whole game here.

What this costs you

An agent that escalates readily resolves fewer conversations on its own. If you're shopping on deflection rate, ours will lose to a tool that answers everything with a straight face.

We're fine with that trade. Slow is recoverable. Wrong, in writing, to a customer, is not.