How to draft replies to support tickets for staff to send with an AI agent

Drafting at depth 2 is where most support agents should start. The agent reads the ticket, the customer's history and your approved sources, writes a reply, and shows the adviser which source backs each sentence. The adviser edits and sends. Their edits are the evidence that decides, category by category, whether the agent may later send on its own.

What is the process, and who owns it?

An adviser working a ticket reads it, looks up the account, finds the right article or policy, and writes a reply. The agent does the reading, looking up and first draft. The adviser keeps the judgement and the send button. This arrangement is the subject of draft and hold is the right default for new support agents: the agent proves itself on real tickets before it speaks to customers alone.

The head of support owns reply quality. Team leaders own their advisers' use of drafts. The content team owns the sources the agent drafts from. The agent's place is set out in our guide to the AI support agent that knows when to say no.

What are the steps, and who acts at each one?

1. Agent, depth 1: reads the ticket and identifies the request, the customer and the category, using the tagging described in how to tag and categorise support tickets with an AI agent.
2. Agent, depth 1: reads relevant history: recent tickets from the same customer and the account facts needed for this request. Can a support agent read a customer's previous tickets? covers how far back that should go.
3. Agent, depth 1: retrieves approved sources: help articles, policies, templates, account data.
4. Agent, depth 1: decides whether to draft. Formal complaints, legal letters, vulnerability signals and urgent tickets get no draft; they get a routing note.
5. Agent, depth 2: drafts the reply in your house style, with a side panel for the adviser listing the source behind each sentence and marking any sentence with no source.
6. Agent, depth 1: runs the promise check on its own draft, flagging commitments for the adviser, as in checking support replies for promises the business cannot keep.
7. Person, adviser: reviews, edits and sends, or discards the draft and writes their own.
8. Agent, depth 3: records the outcome: sent as drafted, edited, or discarded, with the final text.
9. Person, support lead: reviews the evidence by category and decides whether to put the agent forward for certification at depth 3, and for which categories.

Which controls matter, and what evidence does each leave?

Sources visible to the adviser, not the customer. The adviser can check each claim in seconds. Evidence: the source panel stored with each draft.

Unsourced sentences are marked. Evidence: the marks, and whether the adviser kept or removed those sentences.

Some tickets get no draft. Drafting a reply to a formal complaint invites an adviser to send it unchanged. Evidence: skipped tickets with the reason.

Edits are kept. The difference between draft and sent text is the most honest measure of quality. Evidence: draft and final text in the record.

Promotion is per category. Evidence from billing drafts says nothing about cancellation drafts. Evidence: the agent's role scope, by category.

What should you measure?

• Edit rate by category: sent unchanged, lightly edited, heavily edited, discarded.
• Errors found in sent replies, through quality review, traced to draft or adviser.
• Handling time per ticket, before and after drafts.
• Unsourced sentences kept by advisers. These deserve a look.
• Adviser-level patterns. Someone who sends every draft unchanged may not be reading.

Where does it usually go wrong?

Rubber-stamping. Good drafts train advisers to stop reading. Sample unedited sends, and keep sensitive categories out of drafting entirely.

Drafts are too long. The agent explains everything it found. Customers want the answer and the next step. Set a length guide per channel.

Drafts carry over the last customer. Context from a previous ticket leaks into a new one. Each draft must use only this ticket's customer.

Edit data is never used. Edits pile up but nobody reviews them by category, so the agent never earns depth and advisers never get better drafts.

Advisers stop writing. After months of editing drafts, some advisers lose the habit of composing a reply from scratch, which matters on exactly the tickets the agent skips. Keep a share of each adviser's work on undrafted tickets, and include writing in coaching.

Drafts arrive too late. If the draft appears after the adviser has started typing, it is ignored or pasted over their own text. Generate drafts when the ticket is assigned, not when it is opened.

Promotion happens by default. A team lead decides drafts "look good" and turns on sending. Use the evidence, then certify; should an AI agent send replies to customers without review? sets out what that takes.

Where Pryme fits

In Pryme, the drafting agent runs at depth 2, and advisers' edits feed the case set Attest replays when you put a category forward. A category moves to sending only when the agent's certificate clears every gate at depth 3, and the agent drops back to depth 1 the same day if a run clears no depth.