An AI support agent that cites its sources

The question nobody in this category can answer

Your AI agent just told a customer that their climbing harness is returnable. Is that right?

For most tools the honest answer is nobody knows. The reply was generated from a model that read some documents, and the interface shows you the reply. If it's wrong, you find out when the customer does — and you reconstruct what happened by guessing which help-centre article the model might have been looking at.

That's the normal state of this product category. It's also completely avoidable.

What a cited answer looks like

Every Glassdesk reply carries three things alongside the text:

• The sources it used. The specific knowledge-base chunks retrieved to write that answer — not "your help centre", but the actual passages, identified.
• A confidence score. How well-supported the answer was by those sources.
• A decision. Answered, or escalated — and if escalated, a bounded reason code.

So the harness question has a real answer: you open the ticket, see that the reply cited returns-policy at 94% confidence, read that passage, and confirm in seconds whether the policy actually says what the customer was told. If the policy is wrong, you fix the document — not the prompt.

Low confidence escalates. It never guesses.

The corollary of scoring confidence is being willing to act on it. When retrieval doesn't surface good support for an answer, Glassdesk hands the ticket to a human with the context attached, rather than producing a fluent, plausible, unsupported reply.

Certain categories escalate unconditionally regardless of confidence — refund exceptions, legal threats, anything safety-related, and abusive messages. Those are configuration, not a model judgment call, because they're too consequential to leave to a score.

This is why our published deflection rate is 65% rather than 90%. Escalating correctly is the feature.

Proof before it's live, not after

Citations tell you why a specific answer happened. Two other things tell you whether to trust the agent at all:

• A published eval suite. A fixed set of cases — including ones that must escalate, and adversarial attempts like prompt injection and policy forgery — re-run before any prompt or retrieval change ships. Our results and case mix are public.
• Shadow mode. Point it at your live Gorgias tickets and it drafts replies without sending any of them. You read what it would have said, on your real tickets, before a single customer sees it. Turn sending on when you're convinced — not before.

Who this is actually for

• You sell something where a wrong answer costs real money — safety equipment, regulated goods, high-value items, anything with a complicated returns exception.
• Someone will eventually ask you to justify an answer — a chargeback, a complaint, a compliance review.
• You've been burned by a confidently wrong bot and want the escalation path to be a designed behaviour rather than a hope.

If none of those apply and you just want the cheapest deflection at low volume, we'd genuinely point you at the alternatives — several are cheaper below about 110 resolutions a month.

And the bill doesn't move

Glassdesk is flat: $49 / $99 / $199 a month, a disclosed $0.15 overage, and no per-ticket fee — so a resolved ticket is never billed twice. Most of this category charges $0.90–$1.50 per resolution, which bills hardest in your busiest month. The break-even math is here.