Training AI customer service: correct once, right every time after
Published on October 29, 2026

Every new temp in support asks the same three questions in their first week. Up to what amount can I refund without asking. What do I write when the return window ended two days ago. What do we reply when someone threatens to call a lawyer. You explain it, they forget half of it, you explain it again. After three months they move on, and the next one asks the same three questions.
With an AI agent it starts exactly the same way. It does not know your rules, it only knows general politeness. The difference is in what happens to your correction. With most tools it disappears the moment you hit send. With Humane it becomes a rule that applies to every similar request from then on. This article shows how that works in practice, what you should set up beforehand, and where the limits of this learning are.
Why the correction matters more than the first draft
A language model writes polite, grammatically clean replies out of the box. That is not the problem. The problem is that these replies follow general rules and not yours. A model does not know that you still exchange items for regular customers after 40 days, that you never apologise for the carrier but for the wait, and that “Dear Sir or Madam” is banned in your shop.
A new employee learns these things through correction. They write a draft, you read it, you say “not like that, like this”. With an AI agent, that exact moment is the most valuable one, because it holds what separates you from a generic bot. If a tool does not capture that correction, you are still correcting the same thing in month six as in week one. That is why ratings of chatbots are so mixed: according to a Bitkom survey from 2025, only half of those who have used a chatbot in customer service were satisfied with it, while 86 percent were satisfied with a human contact. The human knows the shop’s rules. The bot often does not.
What you set up once: hard policies and dynamic guidelines
Before learning from corrections begins, you give the agent a framework. At Humane this happens in the Training Advisor, a chat where you describe your rules the way you would explain them to a new colleague. The advisor sorts them into two categories:
- Hard policies are rules without exceptions. Return window, refund limits, countries you do not ship to, statements that must never be made (“we do not guarantee delivery by a specific date”). They are checked on every reply. The business rules engine makes sure a draft that violates a hard policy does not go out.
- Dynamic guidelines are instructions about tone and approach. “For delays we always apologise and name a concrete date.” “We use first names unless the customer is formal first.” “For complaints we ask for a photo before offering a replacement.” They shape how the agent writes, not whether.
On top of that come escalation steps: what happens when a customer writes a second time, when a lawyer is mentioned, when an amount exceeds your goodwill limit. This is not prompt engineering. You describe your support, the advisor structures it.
How to put together a usable knowledge base for the agent in an hour is covered in What an AI agent needs to know about your shop.
What happens when you edit a draft
Now for the actual mechanism. In the first week the agent only writes drafts. You read them in the dashboard and have three options: send, edit and send, or discard.
Editing is where the learning happens. You correct the text directly in the draft and choose “Train on this message”. Two things are recorded:
- The corrected version. The agent sees what it wrote and what you turned it into.
- The reason. One sentence from you on why you changed it. “For delays we always name a concrete date.” “This is a regular customer, we refund those without a return.”
From these two parts a guideline is created. From then on it sits alongside the rules you set up in the advisor and is taken into account for every similar request. The reason matters more than the correction itself. A changed wording without an explanation could be taste. A reason turns it into a rule that transfers to other cases.
An example
A customer writes that their parcel has been “out for delivery” for five days. The agent drafts:
We are sorry that there has been a delay. We have informed the carrier and will get back to you as soon as we know more.
You change that to:
We're sorry, five days is too long. We filed a trace request with the carrier today. If nothing has arrived by Thursday, we'll send you a replacement, without you having to do anything.
And as the reason you enter: “After more than four days without movement we file a trace immediately and name a date for the replacement.” From now on the agent knows what to do with stalled parcels, in your words. The next request of this kind gets a draft that already sounds like that.
How you know it is working
The simplest measure is the number of your corrections per week. During the training week you change a lot; that is normal and intended. After that the curve should fall. If in week four you are still making the same kind of change as in week one, either the reason was not clearly worded or two of your rules contradict each other.
Three things help the curve fall quickly:
- Correct the rule, not just the text. “Please be friendlier” does little for the agent. “We never open with an apology, we open with the solution” does.
- Correct early. Ten corrections in the first week are worth more than ten spread over three months, because in the meantime the agent does not repeat the same mistakes in drafts.
- Look at the guidelines. In the dashboard you can see what your corrections have become. Some can be merged, some were one-offs and should be deleted.
When the drafts for a category go through unchanged for a while, that is the moment to approve it for automatic replies. How that works, and which tickets still stay with you, is covered in Auto-send: which tickets Humane answers without approval.
What this learning cannot do
Honestly, there are three limits that no training removes.
Contradictory rules stay contradictory. If your website says a 14-day return window but in practice you give 30, someone has to decide. The agent follows what is stored. Two corrections in opposite directions cancel each other out.
One-offs are not rules. If you grant a customer something as a goodwill gesture that you would not normally do, say so in the reason (“exception, do not generalise”). Otherwise a gesture becomes a guideline.
The agent learns from what you do, not from what you mean. If you send a draft even though it was not perfect, because it had to be quick, that was a confirmation for the agent. That is not a reproach; it is just important to know: sending without changes is a signal too.
How an AI comes to sound like your brand in general, and how you can tell when it does not, is described in Tone of voice: how an AI sounds like your brand. For the bigger picture of what AI in customer service can and cannot do today, see AI in customer service: what really works in 2026.
What you can do now
- Write down the three questions every new temp asks in their first week. Those are your first hard policies.
- Set aside twenty minutes a day for drafts during the training week. Correct with a reason, not just a red pen.
- After two weeks: review the guidelines, merge duplicates, delete one-offs.
- Approve category by category, not everything at once.
If you want to see how a draft turns into a rule: try 100 customer requests for free, no credit card. The first week is the training week; nothing goes out automatically until you switch it on.