When customers recoil from an AI agent, the usual fix is to make the AI better. That is the wrong fix. The recoil is mostly about where you placed the bot, and a smarter bot in the same wrong place makes it worse.
Michael Howlett, who writes the Customer Experience Decoded newsletter, gave a name to a feeling most customers know well. He calls it the AI ick: the small recoil you feel when you realise the thing handling your problem is a machine, and your problem was one you wanted a person for. The ick is real. It shows up in abandoned chats and demands to speak to a human, and it costs money.
The common reaction is to blame the model. The AI misunderstood, it looped, it gave a canned reply, so make the model better and the ick goes away. That reaction sends companies to fix the wrong thing. This article explains why the ick is mostly about the moment you used the bot in, and why a better model in the wrong moment makes the recoil sharper.
Customers welcome AI in one kind of moment and reject it in another
The ick comes and goes. The same customer who recoils from a bot in one chat will happily use one in the next. The difference is the moment. Howlett's framing splits moments into two kinds.
In the first kind, the customer just wants a fast, correct answer and does not care what gives it. Tracking a parcel, resetting a password, checking a balance, moving an appointment. A bot here is welcome, because a human would only add a wait. Call this the opt moment: the customer would choose the machine if you asked them.
In the second kind, the customer is dealing with a loss, a mistake, an unexpected charge, a service that failed them at a bad time. They want to be heard, and they know a machine cannot do that. A bot here triggers the ick. The complaint is not about the bot's accuracy. It is about being handed to a machine in a moment that called for a person.
Using one rule for both moments is the mistake
Most companies decide where to use AI by type of task, or by channel, or by cost. They automate billing questions, or automate chat, or automate the cheapest 40 percent of contacts. None of those choices tells the opt moment apart from the ick moment, because the very same billing question can be a routine query or the first sign of a customer's worst week.
When the decision ignores the moment, the bot lands in both kinds. It does fine in the opt moments and collects the ick in the others. The average satisfaction score then blends the two into one number that hides the problem. The company sees a calm average and decides the bot mostly works. The customers from the ick moments are already telling other people a different story.
A better model in the wrong moment makes the ick worse
This is the part the usual diagnosis gets backwards. A smarter model would fix an accuracy problem. The ick in the second kind of moment is something else: the customer feels the company decided their moment was worth a machine rather than a person.
A more fluent, more human-sounding model makes that feeling stronger. The upset customer is now talking to something that performs a warmth it does not have, and the gap between the performance and the moment is exactly what causes the recoil. The company spent its budget making the bot better at being the wrong choice. The fair objection is that customers cannot always tell. Often they cannot. In the moments where the ick lives, they can, and they do.
How to tell which moment you are in
The question worth asking is which moments to automate, not which tasks. The test is simple: in this moment, does the customer mainly want an answer, or mainly want to be heard? A task category cannot answer that. The signals you already hold can: the words the customer uses, how serious the issue is, whether the contact follows a mistake the company made, whether the customer already tried once and came back.
Where this leaves the roadmap
A roadmap built this way looks different from the usual one. It does not measure progress by the share of contacts automated, because that number rewards pushing the bot into ick moments. It measures progress by two things: the opt moments are handled well, and the ick moments reach a person quickly. It treats the human channel as a deliberate choice for a certain kind of moment, rather than as leftover cost.
The AI ick is useful information. When a customer recoils, they are telling you the bot was placed in a moment that needed a person. The fix is not a better bot. It is sending the easy moments to the bot and keeping a person for the hard ones.