What does the dashboard check for you?

The Overview tab tracks five steps for the support agent: approve knowledge, configure the agent, add a website origin, complete a test call and switch on customer calls. Each agent’s calls switch can’t be turned on until its own knowledge, settings, origin and test are in place.

It can’t tell whether the answers are right, whether your pages allow the microphone or who will read transcripts. The REST API and MCP don’t apply its checks either, so the list below covers the rest.

Copy the voice agent launch checklist.

Paste it into your tracker as it is, or trim the lines that don’t apply. The sections after it explain the checks that catch teams out.

Launch checklist in Markdown
## Knowledge and behaviour
- [ ] Every source each agent needs is approved
- [ ] No outdated source is still approved
- [ ] The test set passes on each agent, including uncovered questions
- [ ] Greetings say what each agent helps with
- [ ] Instructions name your real contact routes
- [ ] No instruction asks for bookings, account checks or callbacks
- [ ] No instruction offers a person unless handover is on and tested
- [ ] Sales notes match the live pricing page

## Website
- [ ] Every production origin is listed, www and bare domain included
- [ ] Unused localhost, staging and tunnel origins are removed
- [ ] The script tag is on the chosen pages, before </body>
- [ ] Content Security Policy and Permissions-Policy allow the widget
- [ ] Usual contact options are visible on the same pages
- [ ] Your privacy notice covers voice conversations, transcripts and any handover

## Launch day
- [ ] Customer calls switched on and saved for each agent
- [ ] One real call on desktop and one on a phone, transcripts read
- [ ] A page on an unlisted address is refused
- [ ] If handover is on, one handover tested end to end for each agent
- [ ] Support team told what the agent covers and what it can't do
- [ ] Free conversations remaining noted

## First week
- [ ] New conversations read daily, each miss fixed and retested
- [ ] Feedback notes read and passed to the product team
- [ ] A plan agreed for when the free conversations run out

## Way back
- [ ] A named person can switch calls off
- [ ] Everyone knows removing the script tag hides the button
- [ ] Token register up to date, with revoke steps written down

Why test the knowledge again before launch?

Approval controls what the agent can use; it doesn’t prove the answers are right. The dashboard’s test step needs one completed call, so re-run your full test set on each agent after your last change to knowledge or instructions. The voice agent testing guide covers which questions to put in it.

Check sales notes against the live pricing page on launch day. Website import is switched off, so nothing updates them for you.

What goes wrong on the real site?

These are the website problems a dashboard test can’t show.

  • A missing origin. https://example.com and https://www.example.com are different origins, and a visitor on the one you didn’t list sees “This domain is not allowed to use this support line.” after allowing the microphone.
  • Blocking headers. A Content Security Policy must allow the widget’s origin for scripts, connections and frames, and a Permissions-Policy must allow its microphone. The widget embed reference gives the exact values.
  • A module script. The widget reads document.currentScript, so it does nothing in a tag with type="module".
  • Leftover test origins. Your workspace ID is public in your HTML, so a localhost origin left on the list lets anyone start calls from their own computer that use your allowance.
  • No visible way to reach you. Callers the agent can’t help need your contact options on the page, even if you’ve turned on human handover, because it’s an early release and customer routing is still being tested.

Does your privacy notice need updating?

Check it. Before each call, the widget tells callers that starting means microphone use and a saved transcript, and that they’re speaking with AI. OpenAI processes the voice conversation, and after calls it also analyses what callers said to build automatic insights. If you turn on human handover, an early release, Twilio connects the caller to your team’s phone, and your team first hears the caller’s name and a short AI-written summary of the issue. The security page lists where Kawl Now stores data and who processes it.

Kawl Now doesn’t yet offer product terms or a data processing agreement for pilot workspaces, so make sure your own notice covers voice conversations on your site. If you use handover, it should also cover the caller’s name, the private briefing and the phone call to your team.

Who does what on launch day?

Name one owner. They switch on Accept customer calls or Accept sales calls, make one real call on desktop and one on a phone, read both transcripts and check that an unlisted address is refused. If handover is on, they also complete one handover to your team’s phone. Calls start from your listed origins as soon as the switch is saved.

Tell your support team the same day. Customers may mention the agent, and the team should know what it covers and what it tells callers when it can’t help.

What happens in the first week?

Read every new conversation daily and fix the source behind each miss. Read the feedback the agent noted too, because nobody is emailed when a caller leaves it.

You’re emailed when 40 of the 50 free conversations have started. After that, watch the count of free conversations left in the dashboard, because you may not be emailed again when the last one is used. Paid calls aren’t available yet, so once all 50 are used, new calls are refused and callers are told the free conversations have been used. Decide in advance whether to switch calls off before then.

How do you switch the agent off quickly?

Switching off Accept customer calls or Accept sales calls stops new calls straight away. The button stays on your pages, and the panel says “This line is not accepting calls yet.” Removing the script tag hides the button on your next deploy. If a token may have leaked, revoke it under Developers, and it stops working at once.

Common questions

Can you turn a voice agent off after launch?

Yes. Switch off Accept customer calls, or Accept sales calls, in the Kawl Now dashboard and new calls stop straight away. The button stays on your pages and tells visitors the line isn’t accepting calls, so remove the script tag if you want it gone.

Who should read transcripts after a voice agent launches?

Someone who can change the knowledge base, every day for the first week. They can add a missing answer the same day, and read the feedback the agent noted, because nobody is emailed when a caller leaves it.

How do you know a voice agent is ready to launch?

Every test question passes on each agent you’ll launch, including questions your sources don’t cover, the widget works on your production pages in desktop and mobile browsers, and someone is booked to read transcripts daily. The dashboard’s checklist covers setup rather than answer quality, so your test set decides.