Draft and hold is the right default for new support agents

A new support agent should start at depth 2: it drafts each reply or action, and a person approves and sends it. Draft and hold costs some speed at first, but every approved, edited or rejected draft becomes evidence about where the agent is ready to act alone. An agent that starts by sending alone has none.

What does draft and hold mean in support?

The agent does the work and a person releases it. For a reply, the agent reads the customer's message, gathers what it needs from the account and the knowledge base, and writes the response. The draft waits in a held queue. A support colleague reads it, edits it if needed, and sends it, or rejects it and writes their own.

For actions, the pattern is the same. The agent prepares a refund, an address change or a ticket reassignment, and a person approves it. The detailed treatment of money-related actions is in refunds, credits and exceptions: draft and hold.

The customer sees a reply from the firm. The difference is that a person stood behind every message while the agent was being proven.

Why not start at depth 3 for simple questions?

Because "simple" is a guess until you have evidence. Teams launching an agent usually have a list of question types they believe are easy: delivery status, opening hours, password resets. Some of them are. Others turn out to have edge cases nobody had written down: the delivery status question from a customer whose order was split, the password reset for an account that is locked for fraud reasons.

Starting at depth 3 on those questions means discovering the edge cases through customers. Starting at depth 2 means discovering them through the held queue, where a colleague catches the problem before it is sent. The cost of the second route is a few weeks of review work. The cost of the first is borne by customers.

There is also a measurement reason. An agent that sends on its own gives you only indirect evidence of quality: complaints, repeat contacts, survey scores. An agent in draft and hold gives you direct evidence on every single case: did a person accept this reply as written?

What does the held queue teach you?

More than any other source in the early weeks. For each draft you learn whether it was sent unchanged, edited or rejected, and, if reviewers are asked to pick a reason, why. Grouped by request type, that becomes a map of the agent's readiness:

• Request types where drafts are almost always sent unchanged are candidates for depth 3.
• Types with frequent small edits suggest the brief or tone needs work.
• Types with frequent rejections point to missing sources, wrong bounds or work the agent should not attempt.
• Types where reviewers disagree with each other show that the process itself is unclear.

Routing affects how useful this is. If drafts are sent to the right specialist team for review, the feedback is more expert and more consistent. The approach in how to route support tickets to the right team with an AI agent applies to held drafts as well as to tickets.

How does an agent move from draft to send?

Through certification. The held-queue data shows which request types look ready. Certification then tests the agent properly, on a labelled set of real cases, gate by gate. If the agent clears every gate at depth 3, it earns depth 3, up to the ceiling its role allows.

Everything else stays a draft. An agent might send delivery updates on its own while still drafting every reply about billing disputes. That is the point of certifying the agent and setting the depth each action needs: the agent's authority matches its evidence.

Which actions should stay held regardless?

Some actions should stay at depth 2 however good the evidence looks, because their consequences call for a person's decision:

The customer can enforce these through its own limits, so they are held for a named person whatever depth the agent has earned.

What does draft and hold cost, and how do you keep it manageable?

It costs reviewer time, and that cost is real. If the agent drafts hundreds of replies a day and each needs reading, someone has to read them. A few things help. Show reviewers the sources the agent used next to the draft, so checking is quick. Make the right data available at the moment of drafting, which is what getting AI agents the right data at the right time is about. Batch similar drafts. And set an exit plan from the start: which request types you expect to move to depth 3, and what evidence will trigger the move.

Without an exit plan, draft and hold becomes permanent, and reviewers become a bottleneck who stop reading carefully. With one, it is a deliberate phase that ends with an agent proven on your own cases.

Where Pryme fits

In Pryme, new agents start with depth set by evidence, so a support agent typically begins at depth 2 with every draft held for a named person. Held-queue decisions feed into Attest, which certifies each agent on the customer's cases before it can earn depth 3. Actions the customer wants held, such as exceptions, stay held through the customer's own limits. This is the stance behind the AI support agent that knows when to say no.