The Hanover lounger case has produced its inevitable sequel question: would an AI have made it better?

It is a reasonable thing to ask. The case has the precise texture of the problem that agentic AI is meant to solve. A customer is unhappy. The cause is operational. The intermediary has the data, the supplier has the levers, and somewhere between them sits a resolution that nobody had to drag through a district court. If ever there was a moment for the small army of customer-service AI vendors to step forward with a clean counterfactual, this is it.

Let me sketch the optimistic version. Our German guest, towel-less and aggrieved at six in the morning, opens an app and tells it so. An agent is dispatched on his behalf. It pings TUI's agent. TUI's agent pings the hotel's agent. The three of them confer at machine speed — bookings, contracts, occupancy, policy. A package of options is assembled within seconds: a relocation by the quieter pool, a partial refund, a credit toward dinner, a complimentary mid-tier Sauvignon, and a quiet recommendation that towels deposited before eight in the morning be removed by staff for the remainder of the stay. The guest accepts. Everybody goes back to the buffet. €986.70 stays in TUI's accounts. The dog never bites the man, the man never sues anyone, and a trial judge in Hanover doesn't have to write a paragraph about loungers.

It is a tidy picture. It is also, I think, mostly wrong.

The problem the AI was not asked to solve

Begin with the obvious. The Hanover case is not, at its core, an execution problem. The hotel had a policy against pre-dawn towel deployment and chose not to enforce it. TUI had a brochure promising a holiday and chose not to underwrite it. The dispute is over who is responsible for that pair of decisions, not over whether the right form got routed to the right desk.

An agentic AI is a magnificent instrument for executing decisions you have already made. It is a poor instrument for making them. If the hotel has not decided whether to enforce its lounger policy, no quantity of inter-agent negotiation will manufacture a position out of thin air. The AI will instead do what AI does well, which is to surface an answer that looks reasonable, present it with confidence, and then, when challenged, fold quickly into whatever the operator's legal team prefers. The substantive question — what are we actually selling? — was never on the table.

This is the part the vendors do not want to dwell on. Their pitch deck implicitly assumes that the bottleneck in customer service is response speed, or coverage, or the cost of a human agent in Manila. In many transactional cases that is true. In a case with the moral character of Hanover — ambiguous policy, competing intermediaries, an aggrieved customer with both reasonable expectations and a slightly absurd cultural ritual at stake — it is not. The bottleneck is upstream of the agent, and no software runs upstream of a decision the business has declined to make.

What the studies actually say

It is worth being honest about the empirical picture, because the marketing material is unusually noisy. The more serious 2025 work on AI in service recovery shows a real efficiency advantage and a less flattering satisfaction picture. One controlled study found that an empathic AI agent resolved more cases (around 62%) and offered more compensations than a neutral counterpart, but participants rated the neutral one higher on satisfaction and organizational reputation. A separate study found between 79% and 93% of customers preferred a human agent in service-recovery contexts, citing better understanding (61%), clearer explanations (53%), and less frustration (52%). A useful term has begun to appear in the literature — "parametric reductionism," which has the additional advantage of looking serious in a paper title — for the AI tendency to compress emotion into quantifiable parameters and then respond at the level of the parameters rather than the emotion.

You could think of it as the chatbot's particular tragedy. The thing that makes AI fast and consistent — its willingness to reduce a problem to its solvable parts — is precisely the thing that makes it unsuitable for the parts of a complaint that are not, properly, problems at all. The guest who has paid €7,000 to be told at six in the morning that he is the seventeenth person to attempt the maneuver does not want a parametric resolution. He wants, in some non-trivial sense, to be heard. Whether that is best provided by real empathy or convincingly performed empathy is an open question, and a more interesting one than it looks. It is not obvious that performed empathy from a machine — the polite "I understand how frustrating this must be" that everyone has now heard several hundred times — will outperform indifferent empathy from an underpaid duty manager. The data so far suggests it usually doesn't. Fake empathy delivered fluently can de-escalate, but when the customer detects the seam — and increasingly, they do — the escalation curve gets steeper, not shallower.

The case for AI subjectivity-elimination is similarly less clean than the deck suggests. AI is not subjective in the human sense; it is biased in the data sense. A model trained on years of TUI complaint resolutions will be exquisitely calibrated to whatever historical settlement amount has, on average, kept the matter out of court. That is not neutrality. That is the firm's accumulated preference, expressed without the friction of a human exercising judgment. It is more consistent. Whether it is more right is a separate question, and one most vendors prefer to leave unasked.

The cow path problem

There is a longer story here that the AI industry would prefer the audience forget. The contact-center business has, for forty years, been replumbed by successive waves of technology — IVRs, offshoring, web chat, chatbots, now agentic agents — each sold as a service improvement and each delivered as a cost reduction. The American Customer Satisfaction Index has held essentially flat since 2017, and has fallen in three consecutive quarters into early 2026. Tier-one IVR resolution rates have hovered in the low double digits for a generation, with caller frustration above 80%. The promise has been consistent, the outcome has been consistent, and the consistency is not in the customer's favor.

A useful phrase among the more sober practitioners is that we are paving cow paths. The metaphor is a fair one. The path predates the engineering. The cows decided where it goes. The engineers arrived later and, finding the cows uninterested in renegotiation, are now pouring concrete. Agentic AI, applied to a customer-service operation whose actual problem is that no one has decided what the policy is, will do what each previous wave did. It will pave the path. The path will go where the cows have always gone. The reduction in cost will be real. The reduction in service will be real. The two will be presented as the same thing.

It is worth saying that the technology itself is not at fault here. The fault, such as it is, is in the business logic that keeps pulling the technology in to solve a problem it was not designed for. Each generation of customer-service automation has been used, primarily, as a permission slip to cut a budget. The headline KPI on the rollout deck is cost per contact. Satisfaction shows up several pages later, in a smaller font, with a careful narrative wrapped around an unmistakably flat line.

The unhelpful conclusion

So: would an AI have improved the outcome at TUI? It is a non-obvious answer, and I think the honest one is "probably not, and possibly worse." A well-built agent could have surfaced a faster resolution at lower cost. It could not have manufactured a policy the hotel had declined to enforce, an underwriting commitment TUI had quietly withdrawn, or a moment of recognition that the customer was paying for something his intermediary had stopped standing behind. Those are decisions taken in the building, by people who are paid to take them. The Hanover case is, fundamentally, a failure of decisions, not of execution. AI does not fix that. In some configurations it makes it worse, by giving the executives who have not made the decisions a faster, cheaper, more emotionally fluent way to avoid them.

The lesson generalizes. The most valuable use of customer AI is in places where the underlying business has already decided what it is selling and just needs the apparatus to deliver it faithfully at scale. The least valuable use is in places where the business is reaching for AI as the next cost-cutting instrument in a sequence of cost-cutting instruments, to disguise the fact that it never quite committed to what it sold in the first place. The first is genuinely useful work. The second is a more polite, lower-cost route to the same district judge in Hanover, who already knows what €986.70 looks like.

I'm Richard Owen, founder and CEO of OCX Cognition. We build predictive customer analytics for companies who'd prefer to know which customers are at risk before those customers have already decided to leave.