The AI could not generate a reply

You added an AI step to your flow, the person wrote in, and no reply came — or the step was quietly skipped. An AI step can fail for half a dozen different reasons, and each one is fixed in a different place. The good news is that the panel records exactly which one it hit on every run, so start there.

💡 Tip: This guide is for an AI step that does not work. If replies do arrive but you dislike the wording, that is not a fault: review the instruction text and the examples in the step.

You may be seeing one of these situations:

• The log says the AI could not answer.
• The log says the AI was postponed and the run is waiting.
• No reply was sent and the flow moved on to the next step.
• The step does not appear at all, or shows a plan lock.

A concrete example: a furniture shop answers questions about dimensions with an AI step. One day the answers start getting cut off. The log says the reply hit the length limit: customers now ask about several products in one message and the model runs out of room. Raising the maximum length on the step fixes it. Without reading the log, this would have been written off as "the AI is broken".

Read the reason in the run log

Before guessing, read what the panel wrote. Every run is logged step by step.

1. Open the automation from Automations.
2. Switch to the Logs tab at the top.
3. Tap the run for the person who had the problem.
4. Find the line that says the AI could not answer, or that it was postponed.
5. Read the text on the Reason line and jump to the matching section below.

Key and credit problems

These are the two most common reasons, and both are fixed in settings.

• "No AI key is set." The portfolio has no key of its own and your plan does not include system credits either. A portfolio owner or admin adds one under Portfolio settings › AI keys.
• "This month's AI credits are used up." Your plan credits are exhausted. See my AI credits have run out.
• "The AI key is invalid." The key was copied wrong or revoked at the provider. Edit the key, paste it again, and press Test connection. That test does not consume credits.
• "The AI key has run out of quota or credit." The problem is in your provider account; top it up there.
• "The AI feature is off on this portfolio's plan." The step is not included in your plan. The portfolio owner can upgrade.

A portfolio can hold several keys. On a given account, keys picked for that account are tried first, then keys open to all accounts; when one runs out of quota, the next is used. So the list must contain at least one active key attached to the right account.

Model and response problems

If the key is fine, the problem may be with the model itself.

• "Model name not recognised." When adding the key you chose Another model… and typed a name that is wrong. Edit the key and go back to Default model, or type the name exactly as it appears in the provider's documentation.
• "The reply hit the length limit." The model was cut off before it finished. Raise the maximum length on the step, or ask for a shorter answer in your instruction.
• "The AI returned an empty reply." The model produced nothing. Your instruction may be too restrictive; add a couple of example sentences.
• "The AI refused to answer this content." The provider's content rules kicked in. The person's message or your own instruction may be triggering them; soften the instruction.
• "The AI service could not be reached." A temporary network or provider problem. It clears up on its own shortly.
• "The AI service is busy; it will be retried shortly." The run is postponed and retried automatically. It shows in the log as the AI being postponed. If postponements pass a certain count, the run ends with an error.

Image-reading problems

If the step has the boxes for reading the person's image or the post's media switched on, the step stops on purpose when the file cannot be read.

• "The post's photo or video could not be read from Meta." The step stops so the model never describes a post it has not seen. Clear the media boxes to run text-only.
• "No image in the message", "The image URL is invalid", "The image is too large", "The image format is not supported": these are about the file the person sent. Add an "incoming attachment" condition to your flow and run a different branch when there is no file.
• "An image arrived but images are off on the step": the person sent a file but the matching box is off. Turn it on, or accept the behaviour.

⚠️ Note: Meta forbids storing incoming media. The file is read only at request time and then dropped from memory, so an image from an older message may no longer be readable later.

"The contact did not provide the requested details"

If you use the AI step to collect information, it ends with that reason when the person does not supply what you asked for. This is not an error; your flow has to handle that case.

1. Open the flow and connect a message step to the AI step's "not collected" output.
2. In that message, ask for the information again as a plain question.
3. If you need a strict format, use the information-collection step instead.

Other known issues and limits

• It does not run in the simulator the same way. Some steps make no real call in a test run; to see real behaviour, publish the draft and try it with a real message.
• Your own key still costs credits. The model cost is billed to your provider, but a fixed platform fee is charged in credits on every call. Test connection is free.
• Thinking depends on the key. The "allow thinking" box on the step only matters when the key's own setting allows it.
• There is rate protection. A flow making many calls in a short time is postponed; that is not a fault, it protects the provider's quota.

If you have followed every step and the problem continues, contact our Support team.