Why did the AI answer that way?

You can always find out

Every reply the AI writes comes from a run you can open and read. The run records what it understood the customer to be asking, which of your knowledge it actually used, what it did step by step, how long it took and what it cost.

So "why did it say that" is never a guess. It is a page you look at. This article is about how to read it and, more usefully, how to turn what you find into the right change.

Step 1: open the run

In the helpdesk, open the conversation and look in the timeline for an Agent run entry with a View run link. That takes you straight to the run behind that reply.

A conversation can have several runs, because the AI processes each new customer message. Pick the one sitting next to the reply you are asking about, not the newest one.

You can also browse them all under More, then Orchestrator Runs, which is the better route when you want to look at a pattern across many conversations rather than one case.

Step 2: read the sources

The Sources panel is the most useful thing on the page, and most people skip past it. It lists exactly what was used to build this reply, and only what was used:

• Policies 📋 The rules that applied.
• Articles 📚 The knowledge it answered from.
• Conversations 💬 Past cases it drew on as memory.
• URLs 🔗 Pages it pulled in.
• Workflows 🔀 A procedure it followed.

Each one links to the thing itself, so you can open the article or policy and read what the AI read.

This panel usually answers the question on its own. If a wrong article is listed, you have found your problem in ten seconds. If the right one is missing, the problem is that the AI never found it, which is a completely different fix from rewording anything.

Step 3: the rest of the run

• Intents 🎯 What it decided the customer wanted. When a reply is confidently about the wrong subject, this is where it went wrong, and everything after it was a reasonable answer to the wrong question.
• Status, duration and cost ⏱ Whether the run finished cleanly and what it took.
• View Logs 📜 The step-by-step execution. Go here when the sources look right but the behaviour still does not make sense, for example when a lookup failed or a tool returned nothing.

Step 4: match the symptom to the fix

This is the part that saves real time. Most wrong answers fall into one of six shapes, and each has a different home.

The facts are right but it sounds wrong. Too formal, too chatty, too long, the wrong greeting. That is Communication Style, or a Policy if it is about how one specific type of question should be worded.

The facts are wrong. Open the article listed under Sources. The AI said what your knowledge told it to say, and the knowledge is out of date or unclear. Fix it there, not by adding an instruction telling the AI to ignore it. See Articles and Knowledge Sources.

Sources are empty or thin. It had nothing to answer from and did its best. Adding more instructions will not help. This is a knowledge gap: write the article. This is by far the most common real cause, and also the one people are most likely to misdiagnose as a wording problem.

It knew the right things but skipped a step. It answered instead of checking the order first, or refunded without asking the question you always ask. That is a Workflow, which is what turns knowledge into a procedure.

It could not look something up. The logs show a tool that failed or returned nothing. That is an integration or a data problem rather than an AI problem. Check the connection under Settings and then Integrations.

Wrong language, wrong brand name, wrong opening hours. The conversation is in the wrong market, or the market is missing something. See Markets.

Step 5: rate it

Hover an AI message in the helpdesk and you get a thumbs up and a thumbs down. On a thumbs down you can add a comment. You can also rate from the run page itself.

Use it. It costs two seconds and it is what turns "the AI has been a bit off lately" into a list of specific runs someone can actually work through. A rated negative run with a one-line comment saying what should have happened is worth more than a long conversation about AI quality.

Once a run is rated, the rating shows on the message, so your team can see that a case has already been flagged and does not need reporting twice.

If no reply went out at all

A run that completed without sending anything is usually not a fault. The AI may have produced a draft for a human to approve rather than a message to send, which is a setting rather than a failure. See Response Methods.

It is also worth checking whether the conversation was closed by the Auto Closer before the AI ever got to it.

Good habits

Look at the run before you change anything. The instinct is to go straight to the instructions and add a sentence. Ninety seconds on the run page usually shows you that the instructions were never the problem, and saves you from an instruction set that slowly fills up with fixes for things that were not broken.

Fix the source, not the symptom. Telling the AI "never say we ship in 2 days" when an article says you ship in 2 days leaves you with two contradictory sources of truth and an answer that will drift back.

Look for the pattern before you fix the case. If one customer got a bad answer, ten probably did. The Orchestrator Runs list and your negative ratings are how you find out whether this is a one-off or a theme.

Keep the good examples too. A thumbs up on a reply that handled something difficult well is a useful record when you later want to know whether a change made things better or worse.