An AI assistant inside the ticket: how it should actually work

There are two ways to put AI into customer support. One is to stand a chatbot at the front door and hope it turns customers away politely. The other is to put an assistant next to the agent and let it do the reading.

The second one is less visible and far more useful, and it is the one most teams should start with — because it makes every agent faster on day one without ever letting the AI speak to a customer unsupervised.

What "has already read the ticket" means

When an agent opens a ticket, they do the same four things every time: read the customer's message, look up the booking or order, check the history, and find the policy that applies. Five minutes on a good day, fifteen on a bad one. Then they write.

An assistant that lives inside the ticket does that reading before the agent arrives. By the time the ticket is open, the rail beside it already shows:

• A summary — two lines: what the customer wants and what has happened so far.
• The record — the booking, the payment status, the previous tickets from this customer.
• The relevant knowledge — the policy section or FAQ entry that applies, cited.
• A drafted reply — written from all of the above, in your tone, with a confidence score.

The agent reads the draft, edits or accepts it, and sends. The reply is still theirs. The reading was not.

Why the citation matters more than the draft

A drafted reply is only as good as what it was drafted from. If the assistant is working from a general model's idea of a refund policy, the draft will be plausible and occasionally wrong — which is the worst combination, because plausible-and-wrong gets sent.

A draft grounded in your own documents is different in kind. When the rail shows "Refund policy §3 · Gateway settlement log" under the draft, the agent can check the source in one click, and the model has nowhere to invent. Every answer traces back to something you wrote.

This is why the knowledge base has to come before the assistant. Load the policies, the SOPs, the FAQ pages; extract and test them; publish. The assistant is only as good as the shelf it reads from.

Confidence, and what to do with it

Every draft should carry a confidence score, and the score should change what the agent sees.

• High confidence (say, above 85%) — draft shown ready to send. One click.
• Medium (60–85%) — draft shown with the uncertain part highlighted and the sources listed. Read before sending.
• Low (below 60%) — no draft, or a draft clearly marked as a starting point, plus the relevant sources. The agent writes.

The thresholds are yours. They should be different for a refund desk and a general enquiries desk, and they should move as the correction rate falls.

The correction rate is the only metric that matters early

When agents edit drafts, record what changed. Over the first month you will see three groups:

1. Drafts sent as-is — the assistant is right and the agent trusts it.
2. Drafts lightly edited — tone, a name, a sentence. The assistant is right; the voice needs tuning.
3. Drafts rewritten — the assistant is wrong, or the knowledge base is missing something.

Group three is your to-do list. Each rewritten draft usually points at a document that should exist and does not. Fix those, and the rewrite rate falls; when it is low enough, you have the evidence to let the same knowledge answer customers directly through a bot.

Beyond the reply

Once the assistant is reading the ticket, drafting is the first job, not the only one. The same reading produces:

• Triage on arrival — category, priority, intent and a summary before anyone opens it, so routing rules have something to route on.
• Skill and gap insights — which agents handle which categories well, from the replies that went out and the CSAT that came back.
• A digest — "reply quality on refund tickets up since the new policy doc; reschedule requests spiking on the 11:00 sailing". The things a team lead would notice if they read every ticket, which they cannot.

What to insist on

If you are evaluating an agent assistant, three requirements are non-negotiable:

• It drafts from your documents, with citations shown. Ask to see the source under a draft.
• The agent sends, not the model. Unsupervised sending is a separate decision, made later, on evidence.
• Every AI action is logged with its reasoning. For audit, and so a manager can see what the assistant did on a ticket that went wrong.

Request a demo of Service CRM →

Start here, not with the bot

Teams that start with an inside-the-ticket assistant get a faster desk in the first week and a tested knowledge base in the first month. Teams that start with a customer-facing bot get a deflection number and a list of complaints. The first path leads to the second one anyway — with the confidence, and the evidence, to do it properly.