Who gets the human?
AI will make support cheaper but the harder decision is which moments still deserve a person.
Years ago, a young customer support agent in our startup digital bank took a taxi across Dublin to hand deliver rugby tickets to a customer. The tickets had gone to an old address, the match was that afternoon and, technically, it was the customer’s fault. Most companies would have pointed to the terms, closed the case and enjoyed the quiet satisfaction of being correct.
She did not.
She found two replacement tickets, paid for the taxi herself and set off across the city. Nobody approved it. No policy allowed an agent to leave the office and spend money saving someone’s Saturday. She understood what we were trying to be as a company and made the call. We lost money on the customer that day. When she returned, we were delighted.
Had we hired someone unusually good, or built a culture where that kind of judgement felt safe? Probably both.
The story came back to me during a recent meeting about AI in customer support. Everyone was discussing how much work bots could take. I wondered what, exactly, we were handing over. For routine support, AI makes sense. I am happy asking a bot for a statement, payment status or password reset. I do not need hold music while somebody reads me the help centre.
A 2026 Gartner survey found that 87% of customers considered access to a human essential. People can accept automation but for the stressful key moments they still want a person.
The valuable part of what that agent did in Dublin was her judgement. She understood the process, why it existed and why this afternoon mattered more. She was not ignoring a fraud control or a legal requirement. She knew which rule could bend without putting anyone at risk.
That is much harder to automate than finding an answer in a help centre.
A DPD customer once tried to use its chatbot to locate a missing parcel. He could not get the information or reach a person. Frustrated, he persuaded the bot to swear and write a poem about how useless DPD was. The poem went viral. The actual failure was worse. The customer needed help and the system had nowhere useful to send him. DPD disabled the AI part of the chatbot.
That leaves judgement with a smaller group of people. Their time becomes scarce, and scarce service gets rationed. Klarna offers a blunt preview. In 2024, it said its AI assistant was doing work equivalent to 700 full time agents. Its chief executive later described human customer service as a VIP thing, while also promising customers a human connection.
The human will remain. Reaching one with authority to act will carry a far higher price.
So who gets one?
Most companies will begin with status. Platinum customers, large accounts and people who pay for priority support move to the front. Every company already knows how to rank customers by revenue.
Status can set the normal service level. It should not be the only route to human judgement. Certain moments need an override. A customer who spends a fortune and wants a copy of a statement may need less help than someone whose salary payment has vanished. Most systems know what a customer is worth to the company. They are worse at recognising what the moment is worth to the customer.
A better routing system would look at what is at stake, how reversible a mistake is, whether delay increases the harm and whether someone needs authority to make an exception. Sentiment alone is not enough. Serious problems often arrive in calm language.
When an account is frozen, a payment has disappeared or somebody suspects a scam, they want a competent person who can understand what happened, make a decision and tell them they have got it. The bot may record that conversation as handled and the customer may remember it as the day they stopped trusting you.
Us humans are not automatically better. We can be slow, inconsistent and deeply committed to the sentence there is nothing I can do. A person reading the same script is a more expensive chatbot.
What customers value is a person with context, authority and enough common sense to make a judgement. The job is deciding where that judgement belongs. Give routine requests to AI. Keep regulated decisions within clear controls. Route urgent, ambiguous and hard to reverse cases to people before the damage grows. Customer value can still matter. It cannot be the only signal.
Over the next few years, human agents will become harder to reach and more carefully allocated. Some companies will treat them as a paid privilege, a form of luxury prduct.
Better companies will preserve a moment based override, so a serious case reaches a person before the customer has to plead with a bot.
See you out there
Martin
P.S. This week’s soundtrack is Human by The Killers.