Untrained mode

What it is

Untrained mode is what visitors get when a chatbot is activated but has never finished a training run. The page loads properly — your colours, your avatar, your welcome message — and then tells the visitor honestly that it is not ready to answer yet.

It is a deliberate design choice rather than an error state. The alternatives were both worse: a not-found page makes it look like your business is broken, and letting an empty chatbot improvise means it invents answers about your refund policy.

Draft and inactive chatbots never reach this state. They return a not-found page instead, because from the outside they do not exist at all. Telling those two symptoms apart is the fastest diagnosis in the product:

• Not-found page → activation problem.
• "Not ready yet" reply → training problem.

When you would use it

• Your hosted page loads but will not answer from your documents.
• Analytics says the chatbot has never been trained and you want to know what visitors are currently seeing.
• You removed some knowledge and the chatbot went quiet.

Where to find it

Open Dashboard → Chatbots → your chatbot → Analytics, then Knowledge.

How a chatbot ends up here

Three routes:

1. Activated too early. Someone activated it before adding knowledge — common when a team wants the link to exist before the content is ready.
2. Training never completed. Sources were added but every one is still pending, still processing, or failed. See Knowledge stuck processing.
3. The last source was removed. Delete everything and the chatbot has nothing to search, so it returns to untrained. See Remove knowledge.

Which training states count as "not ready"

State	Ready?	What visitors get
Not trained	No	The not-ready reply
Pending	No	The not-ready reply
Training	No	The not-ready reply
Failed	No	The not-ready reply
Trained	Yes	Real answers
Updating	Yes	The previous good answers, uninterrupted
Needs retraining	Yes	The previous good answers, possibly missing the newest content

The bottom two rows matter. Once a chatbot has trained successfully at least once, a later update never drops it back into untrained mode — it keeps serving what it already knows while the new content syncs. You do not need to deactivate anything, and you do not need to warn anyone.

Steps

1. Open Analytics. If the health card says not trained, that is exactly what visitors are being told.
2. Open Knowledge and add at least one source — Upload files, Add text knowledge, or Crawl a website.
3. Wait until at least one source shows as trained. Uploading is not the same as finished — see Dataset status.
4. Ask a real question through Test Chatbot in the chatbot header and confirm you get a grounded answer.
5. If the chatbot is still draft, activate it — Activate or pause a chatbot.
6. Open the hosted link in a private window and confirm you get a real answer rather than the not-ready line.

Should I activate before I have knowledge?

You can, and there are honest reasons to — you might need the link to exist for a review, or want the widget in place before content lands.

But be clear about the cost: every visitor who arrives before training finishes gets the not-ready reply, and that turn can still spend message credits even though nothing useful was said. On the Free plan's 50 credits, a busy afternoon of not-ready replies is expensive for nothing.

The safer order is train, then activate, then share. If you must publish early, add one small document first so the chatbot has something rather than nothing, and finish the real knowledge afterwards.

What you will see

Analytics explains the never-trained state in plain language on the health card. The public page still renders the chatbot's full appearance and welcome message — visitors see a working, branded chat that simply cannot answer yet, not a broken page. Neither the hosted page nor the widget errors out for this reason, provided the chatbot is active.

As the owner you may see a slightly different, setup-oriented experience inside Test Chatbot. That is for you; it is not what visitors get.

Limits and plan notes

Untrained mode is not plan-related — a paid chatbot with no trained knowledge behaves identically to a Free one.

Visitor turns can consume message credits even when the reply is the not-ready message. See Message credits.

Adding knowledge needs permission to manage knowledge; activating needs permission to edit chatbots.

Common problems

Visitors get a not-found page, not the not-ready message.

Then this is not untrained mode. The chatbot is draft or inactive, or the link uses a stale handle after a rename. See Hosted page 404.

I uploaded a file and it still says untrained.

Uploading and training are different steps. Wait for the trained badge on the source. Large files and website crawls take longer.

One source failed and now nothing works.

If it was your only source, the chatbot is untrained. If others trained successfully, the chatbot answers from those and simply cannot use the failed one. Retry the failed source.

Test Chatbot answers, but the public page says not ready.

Check that you are testing the same chatbot, then hard-refresh the public page — it caches its configuration briefly. If the chatbot genuinely has a trained source, give the cache a minute.

It was working, then went back to untrained.

Someone removed the last knowledge source. Check the Knowledge tab and re-add something.