You correct a Bedones agent by writing your note in plain language, straight from the conversation that bothered you. The agent reads the feedback, works out what went wrong and applies the fix from the next exchange of the same kind onward. There is no settings menu to dig through and no scenario to rebuild.
Correcting an agent is not the same as reconfiguring a tool
Most automation tools run on fixed rules. You describe every situation, every reply and every exception up front. When the result is not what you wanted, there is no “do it differently” button: you go back into the configuration and rework the rule that produced the bad answer. It takes time, and it wears you down.
A Bedones agent is corrected differently, with a sentence. You talk to it the way you would talk to a new hire, telling it what was wrong and what you expect instead. The agent is not matching your words against a table of parameters; it understands what you meant and carries that intent over to comparable situations.
The gap in method between a Bedones agent and a rule-based tool is the same gap that separates a chatbot from an AI agent: one runs a script, the other reasons over a context. The full comparison is laid out in chatbot or AI agent for an online shop.
Giving your agent feedback, step by step
Feedback always starts from a real conversation. Take a common case: your agent replied to a customer in a stiff, administrative tone, while your shop talks to people casually. You are not going to hunt for a “tone” setting in a list of options — you open the conversation in question and write your note to the agent.
- Open the conversation that went wrong.
- Tell the agent what was off and what you want instead: “That reply was too formal. Next time, be more relaxed and use emojis.”
- Let the agent take the feedback in: it applies to the next similar question.
The correction does not rewrite the message already sent to that customer; it governs the replies still to come. As soon as another customer asks something of the same kind, the agent answers in the tone you asked for. Nothing was approved, republished or redeployed — the note was enough.
The same move works for anything that counts as sales behaviour: how much detail a reply carries, when the agent offers an alternative, how it handles someone in a hurry. Tone is only one example among many, it is simply the most visible one.
The feedback that actually changes the agent’s behaviour
Effective feedback names a specific situation and the behaviour you expect in it. “Be better” gives the agent nothing to work with: there is no trigger and no rule to extract. “When a customer asks for the price, always include the delivery fee” translates straight into behaviour, because the condition and the action are both spelled out.
- “When a customer asks for the price, always include the delivery fee.”
- “Don’t offer a discount if the customer hasn’t tried to negotiate yet.”
- “Look at this conversation: the customer wanted size L, not M. Pay attention to details.”
- “Be warmer with customers coming back for a second time.”
These four notes share one trait: each describes an observable case rather than a vague intention. Two of them set an explicit condition, another points to a factual mistake in a specific exchange, and the last targets a customer segment you can identify. In all four, the agent knows exactly when to trigger the behaviour you asked for.
One note at a time beats a paragraph of instructions. If you correct the tone, the stock check and the discount policy in a single message, you will not know which of the three failed to stick the day something bothers you again.
Showing a conversation beats explaining a rule
The most powerful feedback is showing your agent an existing conversation instead of describing a rule to it. You open an exchange that went well and say: “Look at how I handled this hesitant customer, do the same next time.” The agent reads the whole thing — how the messages built on each other, the objections raised, the way the sale was closed — and turns it into a model.
The reverse works just as well. On a conversation that went badly, you point at the mistake: “Here you offered an item that was out of stock, always check stock before replying.” The agent ties the error to its context instead of receiving a floating instruction, and the rule it keeps stays attached to a case it genuinely saw happen.
Reading conversations this way assumes the agent already knows your catalogue and your references. That is the case for a Bedones agent, which recognises your products even from a screenshot: without that grounding, a note like “the customer wanted size L” would stay a sentence rather than a usable correction.
An agent that improves while you sell
Feedback compounds: every note stacks on the previous ones and pulls the agent closer to the way you sell. In the first days you read its replies and correct often. After a few days of regular feedback, it answers the way you would yourself on the cases you have already shown it.
What separates it from a human employee is availability. The agent does not sleep, does not take breaks, and applies the same instruction at three in the morning as it does at noon. What you taught it on a Monday evening is in force from the first question that comes in overnight, on any of your sales channels.
Keeping a human tone is still your job, and that is exactly what feedback exists to pass on. Automating your messages without losing the human touch covers the guardrails to set alongside it: what the agent handles alone, and what has to come back to you.