Skip to content

Test and review an agent before publishing

Beside the agent editor sit four panes: Test, Coach, Review and Activity. This article covers Test and Review. Test is a chat with the draft on screen, saved or not, and Review reads the instructions back and suggests fixes. Neither one sends anything to a customer or changes the live agent.

On a phone they’re tabs at the top of the editor: Configure, Test, Coach, Review and Activity. On a wider screen they sit to the right of the form.

To try it out:

  1. Open the agent and choose Test.
  2. Type what a customer might say in the box at the bottom.
  3. Press Enter, or the send button.
  4. Read the reply, and the grey line under it that says what the agent did.
  5. Keep going as the customer, or press Reset chat to start again.

Open the info mark beside Test your AI agent and it says the same thing: it answers from the draft on screen and changes nothing live.

The grey line names each action it took. It might have moved the lead to a stage, recorded an email, left a comment or handed the conversation over. When the agent writes nothing and does nothing, the line reads It had nothing to say. A live conversation would come to you at that point.

Contact fields, beside Chat, is who the agent is talking to. Fill in Name, Phone, Email, Lifecycle, Pipeline column, other details, tags and What your team has written about them to try it on someone it already knows. What the agent changes shows up here too. Comments, Workflow and Outcome show what it did in this test, and read Not yet until it does something.

Edge cases, the tab after Contact fields, takes one message per line and sends each as the customer’s next message, so a dozen awkward openers can be tried in one go. Add the starter set fills the box with the usual ones. Each line is one model call, and nothing goes to WhatsApp.

The agent speaks for whoever it would speak for live: the advisor or business behind a number it answers, or the one on screen if it answers no number yet. Like a live reply, it reads the last 30 messages of the chat.

$contact.firstname uses the first word of Name on Contact fields. With no name there, it comes out as “there”. Quiet hours, Stop after this many replies and follow-ups depend on real time passing on a live number, so the test pane never shows them.

A message that fails keeps its place, marked Not answered. with Retry beside it.

To ask for fixes:

  1. Choose Review.
  2. Press Review now.
  3. Read each suggested fix, grouped under the part of the agent it changes.
  4. Untick any you don’t want, or use the menu on a fix to Apply this one or Dismiss it.
  5. Press the button at the foot of the pane, which counts the ticked fixes, such as Apply 3 fixes.
  6. Read each fix’s Before and After in the Apply fixes dialog, then press the same button there.

Review looks for concrete faults rather than style: placeholders left in, no way to end a conversation or bring a person in, an action switched on with no guideline, instructions that contradict each other, no idea what to do on the first message, and anything that invites the agent to invent prices or promises.

Every fix starts ticked, and Select all ticks or clears the lot. Stop cancels a review still running, and Recheck runs it again. With nothing to fix, the pane reads No fixes needed. The line at the foot of the pane is worth keeping in mind: Fixes are suggested by AI and can contain mistakes.

A fix is applied by finding its Before text exactly. If you’ve edited that part since the review, the fix is skipped and the message that follows says so.

The editor’s header carries the review state: Unreviewed, Reviewing, Reviewed or Needs attention. Editing the instructions, an action or the knowledge sets it back to Unreviewed. An agent left at Needs attention shows a badge on its card in the list as well, such as 2 AI review fixes open.

Does testing send a real WhatsApp?

No. The test chat talks to the draft and sends nothing to anybody.

Is the review saved?

Its result is. The badge and the number of fixes left open are saved with the agent when you press Save draft or Publish. The suggestions themselves are not kept, so press Review now again to see them.

Why did the test stop answering with a message about AI in one hour?

Tests, reviews and Optimize all count toward one limit of 200 AI requests an hour for each person. Wait a little, then press Retry on the line that wasn’t answered.

Does a fix change my live agent straight away?

No. It changes the draft on screen. Nothing reaches customers until you publish.