87% of customers want a way to reach a human. Design for it

The most quoted objection to putting AI in front of customers is that people hate it. The most useful research on the question says something more precise and more actionable.

Gartner published a survey in August 2026 , 3,566 B2B and B2C customers, fielded in February and March, in which 87% said a company using generative AI for customer service must provide access to a human agent.

The same survey found that half of customers say generative AI makes interactions easier. Both things are true at once, and the combination is the whole design brief.

This is not a verdict on AI

If people hated the technology, the number saying it makes things easier would not be fifty per cent. What they object to is a specific experience, and everyone has had it: the loop. The assistant that cannot help and will not let go. The four attempts before a human becomes reachable. The sense that the exit has been deliberately hidden because a deflected contact is cheaper than an answered one.

The complaint is not about talking to software. It is about being trapped with software while a human is visibly available and deliberately out of reach.

That distinction matters commercially, because the two problems have opposite fixes. If people hated AI, the answer would be to use less of it. Since what they hate is being trapped, the answer is to build a better door.

Build the exit first

The practical version, in the order we would build it:

1. Make the exit visible from the first message. Not buried three menus deep. A customer who can see the way out is far less likely to take it.
2. Escalate on frustration, not just on failure. Repetition, rising terseness, a second attempt at the same question, those are signals, and a system that only escalates when it cannot parse the request will miss every one of them.
3. Hand over everything, not a transcript. What was established, what is still missing, what was promised. If the human has to read the conversation and start again, the escalation solved nothing.
4. Never deny being an AI. Asked directly, answer honestly. There is no version of this that ends well otherwise.

Some handovers are not optional

There is a category of moment where escalation is not a service decision at all. Taking consent for a medical procedure. Quoting a credit rate. Giving regulated cost information. In those cases the rule is not “escalate when it goes badly”; it is that information-giving can be automated and the regulated act cannot.

And even outside regulated work, there is the plain commercial risk of an agent that keeps going when it should stop. A tribunal has already held a company responsible for a discount its chatbot described but its policy did not offer. That was not a hostile customer or an obscure edge case. It was a chatbot answering a question confidently, which is what they are built to do.

The counter-intuitive part

Making the human easy to reach usually means fewer people reach for one. The instinct is the opposite, hide the button or everyone presses it, but a visible exit converts anxiety into patience. People will let an agent try when they know they are not stuck with it.

Which is a comfortable conclusion, so treat it with the appropriate suspicion: we have not measured it in our own product, because our product has not had enough traffic to measure anything yet. It is a design position, argued from someone else’s survey. When we have our own numbers, we will publish those instead.