twimco.Let’s talk ↗︎
Menu
Field notes / AI in practice

Zendesk's AI agent launchpad writes procedures from your ticket history. Red-pen them before customers meet them

Your ticket history knows what customers ask. It also remembers every workaround, apology refund and policy you've since changed. Zendesk's new launchpad turns that history into procedures and articles for your AI agent, so the review step now decides what the agent will promise.

Editorial engraving of a bright support team room where three colleagues mark up printed procedures together around a long table.

Zendesk announced AI agent launchpad on November 5, 2026, rolling out through November 30 to customers on its latest AI agent infrastructure. It analyzes ticket history, generates and proposes procedures and help center articles, tests the AI agent against high-impact scenarios, and adds an agent overview and tailored recommendations. Before accepting generated content, check each procedure against current policy, remove past exceptions, confirm handoffs and write your own test cases.

What Zendesk announced

On November 5, Zendesk announced AI agent launchpad, a guided setup for its AI agents, rolling out between November 5 and November 30. It's available to customers on Zendesk's latest AI agent infrastructure.

There are three parts. The first is AI-guided onboarding. Instead of a blank configuration, launchpad analyzes your account's ticket history to detect common customer topics, then generates and proposes procedures and help center articles. It then tests your AI agent against high-impact scenarios to check readiness before go-live.

The second is an agent overview page for each AI agent, showing automated resolutions, total conversation volume and understood conversations, with links to tone of voice, channel configuration and knowledge sources.

The third is tailored recommendations. The AI agent keeps evaluating ticket trends and suggests new generative procedures and help center articles. Admins can generate the content in one click, save a suggestion for later, or skip it with feedback that shapes future suggestions.

For new messaging or email AI agents, the guided flow runs when you create the agent. For existing ones, it's under the AI agents workspace: select an agent, click Optimize in the sidebar, then Agent overview or Recommendations.

Sequence of work

Launchpad rollout

  1. Nov 5, 2026

    AI agent launchpad announced; rollout starts.

  2. Through November

    Accounts on the latest AI agent infrastructure gain guided onboarding, agent overview and recommendations.

  3. Nov 30, 2026

    Rollout scheduled to end.

Dates from Zendesk's announcement, read November 9, 2026.

Why the review step is the decision

Ticket history is a record of what you did, not a statement of what you'll do.

That's the thing to keep in mind with any tool that learns procedures from past tickets. Your history holds the questions customers actually ask, which is exactly what a new AI agent needs. It also holds every refund an agent granted to calm down an upset regular, every workaround for a bug you've since fixed, the shipping promise you made before you changed carriers, and the return window you shortened last spring.

A person reading old tickets knows which of those were exceptions. A generated procedure may not. If it learns that customers who complain about late delivery usually got a refund, it may write a procedure that offers one. Then your AI agent offers it to everyone, politely and consistently.

None of this is a criticism of launchpad. Drafting from real tickets beats a blank page by a long way. It just moves the important work from writing to reviewing. The one-click generate button makes the writing fast. The review has to stay deliberate.

Ticket history is a record of what you did, not a statement of what you'll do.

Red-pen each procedure against today's policy

Before you accept a generated procedure, print it or open it next to your current policies, and read it like an editor.

Check every promise against today's terms. Refund windows, return conditions, shipping times, warranty rules and loyalty perks change, and the tickets that taught the draft may predate the change.

Strike out exceptions. If a step only happened because a senior agent made a judgment call, it doesn't belong in a procedure an AI agent will follow for every customer.

Mark the approval points. Anything that moves money, changes an account, cancels an order or can't be undone should name the point where a person decides.

Name the handoff. Each procedure should say where the conversation goes when the AI agent can't finish: which queue, during which hours, and what the customer is told while they wait.

Treat the generated help center articles the same way. Those are public. A wrong article misleads customers who never talk to your AI agent at all.

One practical tip: review in a small group, not alone. Put the support lead, someone who owns returns or billing, and one experienced frontline agent in a room for an hour with the first batch. Each catches different mistakes. The support lead sees tone and handoffs, the policy owner sees promises that are out of date, and the frontline agent recognises the exceptions that crept in from old tickets. An hour together usually beats a week of separate comments, and you finish with procedures that everyone has actually read.

How the work moves

Five red-pen questions for each generated procedure

  1. Is this still the policy?

    Compare against today's returns, refunds and shipping terms, not last year's tickets.

  2. Is this an exception?

    Remove one-off goodwill gestures that agents granted case by case.

  3. Who decides?

    Mark the step where a person must approve money, account changes or anything irreversible.

  4. Where does it hand off?

    Name the queue, the hours and what the customer is told.

  5. What would break it?

    Write one awkward test case of your own before go-live.

Decision framework: apply to every procedure before you accept it.

Write your own test cases too

Launchpad tests your AI agent against high-impact scenarios before go-live. Keep that, and add a handful of cases of your own. You know the awkward ones: the customer who ordered from two channels, the gift return without a receipt, the person asking about an order placed under a different email, the angry message that's really a cancellation request.

Write five or ten of those down in plain language, with the outcome you expect, and run them before switching the agent on. Then keep the list and run it again after you accept a batch of new recommendations. Our earlier note on how an AI agent audit tests behavior, not just answers, is a useful frame for what to look for.

Include at least one case where the right answer is "I can't do that, but here's who can," because an AI agent that never says no is usually promising too much. Who writes these matters. The best test cases come from the agents who handle your hardest tickets, not from whoever set up the integration.

Read the overview with the right questions

The new agent overview gives you automated resolutions, conversation volume and understood conversations. Those numbers are useful, and they're easy to misread.

A rising automated-resolution count is good news only if those customers actually got the right answer. Pair the dashboard with a small weekly sample: read twenty automated conversations and check them against policy. If understood conversations are high but customers come back on another channel a day later, something is resolving on paper and not in life.

It also helps to note which procedure handled each sampled conversation. If one generated procedure keeps showing up in the bad samples, fix or retire that procedure first instead of adjusting everything at once. Treat recommendations as a queue, not a stream. Have one owner look at them on a fixed day, red-pen them like the originals, and use the skip feedback honestly so the suggestions improve. We made a similar argument in our Gorgias AI agent review: the setup choices you make in week one decide what customers hear in month three.

If you want help turning your support policies into procedures an AI agent can follow safely, with handoffs, approval points and test cases, that's part of our business automation work. Talk to TWIMCO.

Related work: Sutton|Place

twimco.

We build alongside the people doing the work.
Bring us the problem you’re thinking about.

Related work

Sutton Place ↗︎