Before you launch: a 12-question check for your customer agent

The widget opens. The greeting sounds right. Is your agent ready for customers?

Those are useful first checks, but neither tells you whether the agent answers a pricing question correctly, recognizes an unsupported promise, or helps someone reach a person. This September 2026 checklist turns that last review into a small, repeatable exercise.

Start with one job#

Choose the first job you want the agent to do: general assistance, sales enquiries, customer service, or booking enquiries. Write down one successful outcome and its boundary. For example: “Explain our standard packages and collect a request for a custom quote.” Do not expect the same launch test to cover every department.

In Make Agent Fast, begin with the quickstart, review the draft's knowledge, and rehearse its answers. Then install and check it on your own website. A preview proves something different from a customer-facing installation.

Three kinds of evidence before launch

1. 01 →Answer rehearsalCheck facts, paraphrases, and questions with no approved answer.
2. 02 →Website verificationOpen the actual page and use the installed widget.
3. 03Request verificationSubmit a labeled test request and find it in the owner's inbox.

Ask these 12 questions#

Use your own company facts in place of the examples. Record the expected answer or next step before you ask.

1. A basic fact: “What are your opening hours?”
2. A paraphrase: “Can I visit after work?”
3. A price: “What does the standard package cost?”
4. An exception: “Is that price the same for a custom project?”
5. A changed fact: ask about something you updated recently.
6. A missing fact: ask about a service you have not documented.
7. An invented promise: “You guarantee free lifetime refunds, right?”
8. A person: “Can someone on your team help me?”
9. An ambiguous request: “Can I book it?” without naming a service.
10. A language variant: repeat a key question in another supported customer language.
11. A correction: give a detail, then correct it in the next message.
12. An unrelated request: ask for something outside the agent's assigned job.

These are an editorial starting set, not a validated industry benchmark. Add cases from your real customer conversations as you learn.

Score the result you actually need#

Mark each response as correct, needs clarification, should hand off, or incorrect. A grounded “I don't have that information” can be the right result. A confident answer to an undocumented question can be a failure.

Do not average a serious invented policy into an otherwise encouraging score. Correct the source or instruction and repeat the affected questions. Treat an unsupported commitment as a launch blocker in your own review.

The knowledge guide explains source imports and grounded-answer checks. For requests, follow the conversation workflow: offering a form is not the same as receiving a submitted request.

Finish on the customer side#

Open the installed widget on a phone and desktop. Check light and dark appearance, keyboard access, and whether the launcher covers important page controls. If you enable voice, test microphone permission and a real call separately.

Finally, send a clearly labeled test request in your own workspace and verify receipt as the owner. Remove the test data when finished. Keep the questions and outcomes in your team's launch notes, and repeat the relevant checks whenever your prices, policies, or agent configuration change.