Draft Ticket Responses for Agent Review (Human-in-the-Loop AI)

That second pattern has a name in the industry — human-in-the-loop (HITL) — and it's how most teams sensibly start with AI before they ever turn on autopilot. This post shows how to build it with Macha: an AI agent that, the moment a ticket lands, reads the conversation, searches your real documentation, and drops a complete draft into the ticket as an internal note for an agent to edit, approve, or rewrite. No customer-facing message goes out without a human.

A quick framing note, because it matters: Macha is not a helpdesk. It's an AI agent layer that sits on top of the helpdesk you already run — Zendesk, Freshdesk, Gorgias, or Front — and connects it to your knowledge tools (Notion, Confluence, Google Workspace) and commerce systems. You keep Zendesk. Macha just gives it agents that can read, reason, and draft.

What "draft for review" actually means

In an autonomous setup, an agent reads a ticket and replies to the customer. In a draft-for-review setup, the agent does almost exactly the same work but stops one step short — it writes the reply and parks it somewhere only your team can see, then hands control back to a human.

The trade is deliberate. You give up a few seconds of agent time per ticket; in return you keep total control over what reaches the customer. As human-in-the-loop practitioners describe the workflow, the review step exists to catch three specific failure modes before they ship: factual errors (a wrong policy, price, or deadline), tone errors (technically correct but cold or dismissive), and context misses (a reply that ignores the customer's history). A reviewer typically clears a good draft in 10–30 seconds — far faster than writing from scratch, but with a real human signing off.

And that human step isn't ceremony — how you run it changes the outcome. A peer-reviewed study in Cognitive Research: Principles and Implications found that the order of the human–AI hand-off swings reviewer accuracy hard: people who formed their own judgment before seeing a flawed AI suggestion were correct 66.2% of the time, versus just 36.8% for those who saw the AI's answer first and reviewed afterward (Springer, 2024). The takeaway for support is concrete: a draft is a genuine accelerator, but a reviewer who rubber-stamps it without first reading the ticket is measurably worse than one who reads the customer's message and treats the draft as a second opinion. Design the workflow — and your agent's instructions — so the human still engages with the actual ticket, not just the model's guess at it.

The stakes here aren't hypothetical. In Moffatt v. Air Canada (2024), British Columbia's Civil Resolution Tribunal held the airline liable for a refund its support chatbot wrongly promised; Air Canada argued the bot was "responsible for its own actions," and the tribunal flatly rejected that, ruling the company owns whatever its automated tools tell customers (American Bar Association). That's the precise risk draft-for-review removes: a human stays between the model and the customer, so a bad answer gets caught in an internal note instead of in a courtroom.

This isn't a fringe idea. Zendesk's own Copilot generates suggested first replies as editable drafts an agent reviews before sending (if you're weighing the native option, see our Zendesk Copilot explainer); tools like My AskAI and Assembled work the same way. Macha's version is differentiated by what the draft is built from — your actual Notion docs and live ticket fields, via an agent you configure — and by an explicit confirmation gate on the action that writes back.

The workflow, end to end

Here's the agent we're building, taken straight from Macha's use-case library:

1. Trigger — Ticket Created. A new Zendesk ticket arrives needing a response. (You can also trigger on Customer Reply if you want drafts on follow-ups, not just first contact.)
2. Read the ticket. The agent uses Zendesk's Get Ticket action to pull the full conversation, requester, tags, and brand — so it's drafting from the actual message, not a summary.
3. Find the answer. It runs Search Pages against your Notion workspace, opens the most relevant page with Get Page, and grounds the reply in your documented policy rather than guessing.
4. Write the draft. The agent composes a complete, on-brand reply.
5. Hand it to a human. It posts that reply as a Zendesk internal note — visible to agents, invisible to the customer — and stops. Your agent edits, approves, or rewrites, then sends.

That confirmation card is the whole philosophy in one screen: any action that writes to your helpdesk can require a human to confirm it. The draft is generated by AI; the decision to place it — and later, to send it — is yours.

Why drafts beat suggestions

A lot of "AI assist" features give agents a vague suggestion or a canned-macro pick. A draft is different: it's a finished reply the agent can ship as-is or tweak. The difference in practice is between "here's a hint" and "here's 90% of the work, you do the last 10%."

Because Macha agents read the real ticket fields and full thread — and can even read attachments, extracting text from PDFs and using AI vision on screenshots and photos — the draft is specific to this customer, not a generic template. A UK customer asking about returning a defective item gets a draft that already knows it's an international order and a defective-item case, with the right policy pulled in.

Set it up in a few minutes

You build the agent in Macha's configuration view: give it instructions, pick a trigger, and switch on the tools it's allowed to use.

The instructions are where you set the rules of the draft — voice, length, what to do when it can't find an answer (a good instruction: "if the docs don't cover it, draft a short holding reply and flag for a senior agent"). The trigger decides when it runs. And the tools decide what it can touch.

Keep the agent's hands read-only

This is the single most important setup choice for a draft-for-review workflow. When you connect Notion, Macha shows you each tool tagged Read or Write:

For a drafting agent you want read-only knowledge tools (Search Pages, Get Page) so it can pull policy without ever editing your wiki. Combine that with the internal-note action on the Zendesk side and the agent's only "write" is a note your team reviews — it literally cannot send a customer-facing reply on its own. That constraint is the safety model: you've designed an agent that can think and draft but can't speak to customers unsupervised.

Where the draft lands

The agent posts into the ticket itself, as an internal note. Your agent opens the ticket, sees the draft sitting in the conversation, and works from there — public reply, edit, or discard.

Because the draft lives natively in Zendesk, there's nothing new for agents to learn and no second tool to babysit. They keep working in the workspace they already know; the AI just got there first and did the typing.

Read-only vs. autopilot: a sane rollout

Draft-for-review isn't the destination — it's the on-ramp. The pattern most teams use:

Start in stage one. You get the speed-up and a free quality signal: every edit an agent makes is feedback on where the agent's instructions or knowledge need work. When approval rates climb and edits shrink, you've earned the right to let specific, low-risk categories go straight to the customer — and Macha supports that too, on the same agent, by relaxing the confirmation gate for the topics you trust.

Watch-outs — when not to reach for this

Draft-for-review is the safe default, but it isn't free of trade-offs. Be honest about these:

• It still costs an agent's attention. Drafting saves typing, not triage. If a human has to open and approve every draft, your throughput ceiling is still human. For genuinely high-volume, low-risk queries (order status, store hours), a reviewed-but-automated reply will beat draft-for-review on cost-per-ticket — don't keep a person in the loop out of habit once you trust the topic.
• A bad draft can anchor a tired agent. A plausible-but-wrong draft is more dangerous than a blank box, because a busy reviewer may approve it on autopilot. Mitigate with instructions that make the agent cite its source ("Source: International Returns Policy") and say so when it's unsure, so reviewers know what to double-check.
• Garbage in, garbage out. The draft is only as good as the Notion/Confluence docs behind it. If your knowledge base is stale, fix that first — or the agent will confidently draft last year's policy.
• It's per-action metered. Reading a ticket, searching docs, and drafting are AI actions, and each consumes credits. That's usually trivial next to agent time, but it's real — see the cost note below.

What it costs

Macha is credit-based, and credits are charged per AI action, not per resolution or per draft — reading the ticket, searching Notion, and writing the reply are the actions that meter. The rate depends on the model you pick: 0.5–9 credits by model, with the default GPT-5.4 Mini at 1 credit. A drafting agent on a lightweight model is one of the cheaper things you can run, because it's a handful of read actions plus one generation per ticket. You can start on a 7-day free trial, no credit card required and see your real per-ticket cost before committing — full tiers are on the pricing page.

FAQ

Does the customer ever see the AI draft? No. The draft posts as a Zendesk internal note, visible only to your team. Nothing reaches the customer until an agent reviews it and sends a public reply.

Can the agent edit the draft, or is it all-or-nothing? Fully editable. The draft lands in the ticket and your agent can send it as-is, tweak a line, rewrite it, or discard it. The point is to save the typing, not to remove your judgment.

Which helpdesks does this work with? Macha is an agent layer on top of your existing helpdesk — Zendesk, Freshdesk, Gorgias, or Front. This use case uses Zendesk's Get Ticket and internal-note actions; the same pattern maps to the others.

Where does the draft's information come from? From your real knowledge — in this build, Notion pages via read-only Search and Get Page. You can also wire in Confluence, Google Workspace docs, or uploaded sources. The agent grounds the reply in documented policy instead of guessing.

Can I later let it reply on its own for some topics? Yes. Draft-for-review and autopilot are the same agent with a different confidence setting. Most teams start in draft mode, watch approval rates, then relax the confirmation gate for specific low-risk categories.

How is it priced? Per AI action in credits (0.5–9 by model, default GPT-5.4 Mini = 1), billed only when the agent actually runs. See pricing.

Start with the safe version

If AI in your support queue feels like a leap, draft-for-review is the small first step that de-risks the whole thing: real speed, zero loss of control. Build the agent, point it at your Notion docs, keep its tools read-only, and let it draft into internal notes for a week. Read the edits your team makes — that's your roadmap to trusting it with more.

Spin up a 7-day free trial, no credit card required, connect Zendesk and Notion, and ship your first drafting agent. The Macha docs have the full walkthrough, and the use-case library has more agents to build next.

Written by Abbas (Customer Support & AI, Macha) · Reviewed by Ankeet Guha (Co-founder & CTO) · Published 2026-06-24 · Last updated 2026-06-24.