Response Methods

What is a Response Method?

A response method is how far the AI is allowed to go on its own with a given reply. There are three, and they mean the same thing everywhere they appear:

• Draft 📝 Writes the reply and stops. A human sends it.
• Semi-Automated 📨 Sends the reply and leaves the conversation open.
• Fully Automated ✅ Sends the reply and closes the conversation.

It is worth being precise about what this setting does not change: the answer itself. The AI searches the same knowledge, follows the same policies and writes the same text in all three modes. What changes is who presses send, and what happens to the ticket afterwards.

So the question a response method answers is not "is the AI good enough yet". It is "what does it cost if this particular reply is wrong, and who should be the one to catch it".

Draft 📝

The AI does all the work and then hands it over. The reply appears on the conversation in the helpdesk, an agent reads it, edits it if needed, and sends it.

• Nothing reaches the customer by itself. Not on a delay, not out of hours.
• The conversation is not closed. A draft is a proposal, so the ticket stays on the board until a human does something with it.
• The agent keeps the time saving. Most drafts need a glance and a send, not a rewrite, which is where the minutes actually come from early on.

This is the right setting for anything new. It is also the only sensible setting for topics where a wrong answer is expensive rather than merely annoying: goodwill decisions, complaints, anything with a legal edge.

Note: Mirror mode gives you the same thing before you have moved anything over. If you are still on Zendesk or Gorgias, your tickets are copied into Zuzana and the AI writes an answer next to each one, and no customer is ever contacted.

Semi-Automated 📨

The reply goes out on its own, and the conversation stays open.

That second half is the point. The ticket remains on the board, so it still shows up in your counts, it can still be picked up, and nobody has to reopen anything if the customer comes back. Use it when:

• The answer is safe but the exchange probably is not over. "Your parcel is in transit, here is the tracking link" is often followed by a second question.
• You want the speed of auto-send while a human still lays eyes on the thread afterwards.
• You are moving a policy up from Draft and want one step, not two.

Semi-Automated is the honest middle ground, and it is where a lot of good policies end up permanently. There is no prize for closing tickets the customer was going to reply to anyway.

Fully Automated ✅

The reply goes out and the conversation is closed. The AI handled the case end to end.

This belongs to the questions that genuinely are one message long and have one correct answer: opening hours, where the size guide is, what your return window is, whether you ship to Norway. The test is not "can the AI answer this" but "is there anything left to say once it has".

Note: Fully Automated is not a target to be reached. A support inbox where every policy is fully automated is usually a sign that the awkward topics were never written down, not that the AI got better.

Where the setting lives 🧭

The same three values are set in more than one place, at different scopes.

• On a policy. In the policy editor, under the Advanced tab, as the Response method. It applies to replies the AI writes from that policy.
• On a workflow's reply block. Each Respond node in a workflow carries its own response method, so a single workflow can be mixed: the branch that just returns a tracking link can be fully automated while the branch that has concluded the parcel is lost stays on draft.
• On your team and agent configuration, which can cap automation independently of what the individual policies say.

New reply blocks in a workflow start as drafts. That default is deliberate.

The strictest setting wins 🔒

A response method is a ceiling, not a guarantee. Every layer can lower it and none can raise it, so the most restrictive setting in play is the one that happens.

Two consequences worth knowing:

• A Fully Automated policy running under a drafts-only setup produces a draft, and no side effects are fired.
• A tag with a risk level of 50 or above means drafts on conversations carrying that tag are never auto-sent. So a fully automated policy on a ticket tagged Complaint still leaves a person to press send. This is the cheapest safety net in the product, and it works per conversation rather than per policy.

The upshot: you cannot accidentally open the gates from a single screen, and you cannot always answer "why did this not send" from a single screen either. The Runs button on a policy is where you go to find out.

Note: tool approvals are a separate gate. Response method decides what happens to the reply; a Use tool node's approval requirement decides whether the AI may take an action. A fully automated branch can still be required to wait for a human before it opens a carrier trace or cancels an order.

Moving up the ladder

The reliable sequence is boring on purpose.

1. Launch on Draft. Every new policy and every new workflow branch.
2. Read ten real drafts. Not test cases. Real conversations, on the topic, from your own inbox.
3. Fix the source, not the reply. A draft that was wrong means a missing article, a vague policy description or a condition nobody wrote down. Editing the one reply teaches the system nothing.
4. Raise to Semi-Automated once the drafts are boringly correct, and watch what the customers reply.
5. Raise to Fully Automated only for the topics where the follow-up rate is genuinely near zero.

Move one policy at a time. Automation that goes wrong is much easier to find when you changed one thing.

Good to know

• The Runs button on a policy lists every AI run that used it, which is the fastest way to see what a policy is really doing before you raise its response method.
• Unticking Active on a policy takes it out of the AI's index immediately. That is a faster correction than dropping it back to Draft when a policy is actively producing wrong answers.
• The field sits on the policy itself, so a policy has exactly one response method. If one market should be more cautious than another, that is a job for ticket rules or a second policy.
• A high risk level on a handful of tags (Complaint, Legal, Chargeback, Press) buys you more safety than keeping every policy on Draft, and costs you far less throughput.
• Semi-Automated and Fully Automated differ only in what happens to the conversation after sending. If you are unsure which one a topic wants, pick Semi-Automated: the cost of a needlessly open ticket is much lower than the cost of a prematurely closed one.