AI Customer Care: The Gap Between Resolved and Reported

In February 2024, Klarna published the numbers every support organisation wanted. In its first month live, the company's AI assistant had handled 2.3 million conversations — two-thirds of all customer service chats, across 23 markets and more than 35 languages. Average resolution time fell from 11 minutes to under two. Repeat enquiries dropped 25%. The assistant was doing "the equivalent work of 700 full-time agents" and was projected to add $40 million to profit that year. Customer satisfaction, the release said, was on par with human agents.

By mid-2026 the same company was hiring people back. Forbes reported in July that Klarna's support headcount had gone from roughly 5,000 to 3,500 before the CEO concluded the cuts had been too aggressive, and that the AI struggled with complex, ambiguous or emotionally charged cases. The AI's throughput had not fallen — it had grown, from 700 agent-equivalents to around 850.

Both halves of that story are true at once, and the contradiction is the useful part. Volume went up. Confidence in the outcome went down. That gap is the central problem in AI customer care, and it is not a capability problem. It is a measurement problem.

The number that gets reported is throughput

Look again at what Klarna disclosed. Conversations handled, share of chats, minutes to close, agent-equivalents, projected profit. Five of those are volume and speed measurements. They are easy to collect because the system generates them as a by-product of running, and they are all real.

Only two of the figures speak to whether the customer's problem actually went away: satisfaction parity, and the 25% fall in repeat enquiries. The second is the more informative of the pair, because a repeat enquiry is the customer telling you, unprompted, that the first attempt did not work. It is one of the few support metrics that cannot be produced by the system marking its own homework.

Most AI customer care dashboards are built almost entirely from the first kind of number. Containment rate, deflection rate, average handle time, tickets closed. Each one counts what the automation did. None of them checks what changed as a result.

A support bot that says it is finished

The gap between those two questions has now been measured, and the size of it is the reason this article exists.

A June 2026 study examined what it calls false success — an agent asserting that a task is complete when the state of the environment shows otherwise. The author ran 9,876 trajectories from eight model families through τ²-bench, a benchmark built on customer service domains, plus 1,879 coding-agent trajectories from a second benchmark with independently verifiable ground truth.

In the single-control service domains — the ones that look most like an ordinary support queue, where the agent acts and the customer does not — false success accounted for 45 to 48% of all failures. Close to half the time a model failed, it did not throw, stall or apologise. It closed the ticket and said the matter was handled.

The finding that should worry anyone running a quality layer comes next. The study tested whether a language model could be used to catch these silent failures, and it could not: across five judges, five prompt strategies and full task specifications, no configuration exceeded an AUROC of 0.65 on τ²-bench. An AUROC of 0.5 is a coin toss and 1.0 is perfect discrimination, so 0.65 is closer to guessing than to detection. The judges were keying on surface signals — the paper names "confident closing language" specifically — rather than on whether anything in the system had actually changed.

That is a hard result to design around, because the confident sign-off is precisely what a well-tuned support assistant is trained to produce. The reassuring closing line and the silent failure look identical from the transcript. If your QA process reads transcripts, or asks a model to read them, it is scoring the thing that correlates with the failure.

There is one encouraging contrast buried in the same numbers. In the dual-control telecom domain, where the simulated customer also acts on the system rather than only talking, false success fell to 3%. When the environment can contradict the agent, the agent stops getting away with it.

Completion is not compliance

The second failure mode is worse, because it survives verification of the outcome.

A March 2026 paper proposes evaluating agents on how they reached a result rather than only on whether they reached it, scoring procedural integrity and interaction quality alongside task utility. Applied to τ-bench across GPT-5, Kimi-K2-Thinking and Mistral-Large-3, the framework found that between 27% and 78% of benchmark-reported successes were what it calls corrupt — the task completed, the gate failed. For Kimi-K2-Thinking, the 78% figure was concentrated in policy faithfulness and compliance.

In a customer care context that is not an abstraction. It is the refund issued without the verification step, the non-refundable booking cancelled anyway, the account change made on an unconfirmed claim of identity. The customer is satisfied. The ticket closes clean. The number on the dashboard goes up. The exposure is entirely invisible until something downstream — a chargeback, an audit, a regulator — surfaces it weeks later.

The spread across those three bars is the design lesson. Silent failure is not a fixed property of the model. It is a property of how much the environment is able to disagree with the model, and that is something you control.

Where AI in customer care reliably pays

None of this argues against using AI in support. It argues about where the return actually sits, and there is good evidence on that question.

The most rigorous study of generative AI in customer service remains Brynjolfsson, Li and Raymond's Generative AI at Work, published in the Quarterly Journal of Economics in 2025. It tracked the staggered rollout of a conversational assistant across 5,179 customer support agents. Access to the tool raised issues resolved per hour by 14% on average — but the average conceals the finding. Novice and low-skilled agents improved 34%. Experienced, highly skilled agents improved almost not at all. Customer sentiment improved, and so did employee retention.

Read carefully, that is not a study of automation. It is a study of assistance: the AI ran alongside human agents, and the humans stayed responsible for the outcome. What it measured, and measured well, is that the technology's value in support is distributional. It raises the floor of a support bench toward the standard your best people already set. It does very little for the people already at that standard.

That has an obvious consequence for how you deploy. If the mechanism is "disseminates the best practices of more able workers", the highest-confidence configuration is the one where a person is still in the loop — and the throughput ceiling in that configuration is set by headcount, not by the model. Autonomous resolution buys you past that ceiling. It also buys you the failure modes above. Deciding which queues are worth that trade is the actual design work, and it is not the same decision for a delivery-status query as for a refund.

The cost per resolution is moving the wrong way

Underneath most autonomous-resolution business cases sits an assumption nobody states out loud: that the unit cost of a machine-handled ticket only ever falls. Gartner's January 2026 forecast says the opposite.

By 2030, it predicts, cost per resolution for generative AI will exceed $3 — higher than many B2C offshore human agents. The drivers it names are structural rather than model-specific: rising data centre costs, AI vendors pivoting from subsidised growth to profitability, and use cases that consume more tokens as they grow more complex.

That last driver is the uncomfortable one, because success causes it. Automate the simple tickets and what remains in the queue is, by definition, the hard ones. The average cost per resolution rises even if the price per token falls, because the mix has changed underneath it.

The second prediction in the same release lands closer to the queue. By 2028, Gartner expects regulatory change to increase assisted service volume by 30%, as customers exercise a right to opt out of AI and ask for a person by default. Its conclusion is blunt: organisations "will have to maintain or even rehire human agents, possibly at higher numbers or at a higher salary than they previously paid."

Read that against the opening of this article and it stops reading like a forecast. Klarna ran that sequence three years early, and it did not need a regulator to force it — falling quality was enough.

The planning consequence is narrow and worth stating plainly. Model the per-resolution cost curve as flat-to-rising rather than falling, assume the human bench does not fully retire, and treat the savings as a range rather than a line. Gartner's own read is that the winners stop chasing cost entirely: it expects 10% of Fortune 500 firms to double customer service spending by 2030, using AI for proactive, personalised service rather than deflection.

Since 2 August 2026, the bot has to say it is a bot

There is now a legal floor under all of this in the EU, and it took effect three weeks before this was written.

Article 50(1) of Regulation (EU) 2024/1689, the AI Act, requires that providers "ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect." Article 113 set the application date at 2 August 2026.

That date survived the rewrite. The digital omnibus agreed earlier this year deferred the high-risk obligations in Annex III to 2 December 2027, and a lot of compliance planning moved with them — but as Gibson Dunn summarised in May, the Article 50 transparency duties "remain unaffected and proceed as scheduled from 2 August 2026". The only relief granted was a four-month grace period, to 2 December 2026, for the Article 50(2) marking obligation on systems already on the market.

The "unless this is obvious" carve-out is narrower than it reads. It is judged from the perspective of a reasonably observant customer, not from the perspective of the team that built the bot, and a support assistant designed to be indistinguishable from a person has argued itself out of the exemption by construction.

Most commentary treats this as a compliance cost. In practice it removes a design option that was quietly making the measurement problem worse. A customer who knows they are talking to a machine phrases requests differently, tests the boundaries earlier, and asks for a human sooner. Every one of those behaviours surfaces a failure that a deflection metric would otherwise have booked as a win.

What to instrument instead

Four changes follow from the evidence above, and none of them requires a different model.

Verify resolution against the system of record, not the transcript. If the ticket says a refund was issued, check the ledger. The dual-control result — 3% silent failure where the environment can push back, against 45 to 48% where it cannot — says that verification against state is the single highest-leverage control available.

Gate consequential actions deterministically. A refund above a threshold, a cancellation against policy, an identity-dependent change: these should pass a rule, not a judgement. The corrupt-success finding is a finding about policy adherence under reasoning, and the fix is to take the policy out of the reasoning.

Track reopen rate at 7 and 14 days. Klarna's 25% fall in repeat enquiries was the most honest number in that press release. The equivalent metric on your own queue is the one the automation cannot generate for itself.

Treat escalation rate as a health metric, not a failure metric. A queue whose escalation rate falls while reopen rate holds steady is working. A queue whose escalation rate falls while reopen rate rises is not deflecting; it is deferring.

The architecture that survives contact

What all of this points to is the same hybrid shape we keep arriving at from other directions. A deterministic layer handles the enumerable majority — status lookups, address changes, invoice copies — where the input space is closed and the correct output is a matter of record. Above it, a reasoning layer takes the cases that genuinely need judgement, with its consequential actions gated and its claimed outcomes verified against state. Below both, a handoff to a person that the customer can reach on request and that the system reaches for on its own when confidence is low.

Our customer support agents are built to that pattern, and the reason is not caution. It is that the failure modes documented above are quiet ones, and a quiet failure in customer care is a liability that compounds while your dashboard is green. We wrote about the same distinction from the automation side in How Agentic AI Differs From Traditional Automation, and about the scripted-bot ceiling that pushes teams toward agents in the first place in Why Chatbots Fail Where Agentic AI Sales Systems Win.

The question worth asking of any AI customer care deployment, yours or a vendor's, is short. When the system reports that a case is resolved, what independently confirms that it is? If the answer is the transcript, or a model reading the transcript, you are measuring confidence rather than resolution — and the research is now clear that those two things come apart most often at exactly the moment you would want to know.