Why is customer service so bad now?

It isn't your imagination, and it isn't that people stopped caring. Service got worse because of five decisions that each looked sensible on their own spreadsheet. Here's what actually happened inside those companies — and what the ones that avoided it did differently.

It isn't only you noticing

The industry's own research says the same thing its customers do. Surveys of contact centre leaders through 2026 report that customer conversations have become more difficult, not less, even as spending on automation rose sharply — one widely cited figure puts AI budgets growing around 38% against roughly flat overall service budgets. Consumer research continues to find around half of people frustrated by chatbot interactions.

Treat the exact percentages carefully: most of this research is published by companies selling contact centre software, and the numbers move depending on who is asking. The direction is the useful part, because it is agreed on from both ends. The people running support say it is getting harder. The people contacting support say it is getting worse. Those are the same fact seen from opposite sides of the queue.

1. Support was reclassified as a cost centre

Almost every other department is measured on what it produces. Support is one of the few measured almost entirely on what it consumes. Once cost per contact becomes the headline number on the slide, the logic runs itself: anything that lowers it reads as progress, and anything that raises it — more time per customer, a more experienced person on the queue, a phone line that somebody answers — reads as waste. The cost question itself is perfectly fair; what a support team actually costs to build or to buy is worth knowing precisely. The error is letting it become the only number on the slide.

That single reclassification is the root. The four below are all consequences of it.

2. Deflection got mistaken for resolution

Deflection counts the ticket that was never created. It does not count whether the customer's problem went away.

This is the measurement error at the centre of the whole decline. A customer who found the answer in your help centre and a customer who gave up in disgust produce identical data: one contact avoided. One of them is a success and the other is churn with a delay on it, and the dashboard cannot tell them apart. Teams then optimised extremely hard against that number, because it was the number they were judged on. A good version of deflection does exist — a help centre that genuinely answers the question — but it only counts if you measure whether the answer landed.

3. Automation took the easy half, and staffing was cut as if the work hadn't changed

This is the one that catches almost everybody, because it looks like a failure of the team rather than a change in the work.

Simple, repetitive questions are the easiest to automate, so they are the first to leave the human queue. What stays behind is the complicated, the ambiguous, the emotionally charged and the multi-step. The mix changes even when the headcount doesn't.

The trap is that the numbers used to justify the next round of cuts are the numbers this change produced. Average handle time went up, so the team looks less efficient, so the pressure increases. Measured properly, it should have gone up. That was the plan working. It is often the moment a team discovers it has outgrown its support setup without having added a single customer.

4. The people who knew things left

Support has always had turnover. Sustained cost pressure makes it considerably worse, and what leaves with experienced agents is almost never written down: the workaround for the bug that was never fixed, which account manager actually replies, the three questions that reveal what a customer really means when they say the app is broken.

That knowledge is the difference between a resolution and a script. When average tenure falls, the new agent follows the script, because the script is genuinely all they have. The customer experiences this as being talked at by somebody who doesn't understand the problem — which is exactly what is happening, and it isn't the agent's fault.

5. The escape hatch was closed

Removing the phone number, burying the contact form four clicks deep, making a bot the only door in. Each one lowers cost per contact, and each is defensible on its own.

What they do together is remove the release valve. A five-minute annoyance that a person could have settled becomes forty minutes of looping, and the customer who would have been satisfied becomes the customer writing the review. The cost didn't disappear. It moved somewhere the support budget doesn't get charged for it. If somebody else answers your tickets, how they handle this is worth settling before you sign — it belongs near the top of what to ask a support provider.

The clearest public example

Klarna is worth knowing about because the company said both halves out loud, which almost nobody does.

In February 2024 it publicly reported that its AI assistant was handling around two-thirds of customer service chats in its first month, doing work it described as equivalent to 700 full-time agents. In May 2025 its chief executive publicly said that cost had become too dominant a factor in the decision, that quality had suffered as a result, and that the company was hiring human agents again — remote and flexible, rather than a return to traditional call centres.

Two things are worth taking from it, and much of the commentary gets both wrong. The first is that this was a correction, not a reversal: Klarna still describes itself as AI-first, and its stated position is that automation handles the simple work while people handle the complex. The second is the uncomfortable one. Klarna had the volume for its models to learn from, in-house engineers, and budget. It still had to walk the decision back in public. If it went wrong with those advantages, the odds are not better for a thirty-person company that bought a chatbot subscription and switched off its contact form.

What the companies that avoided it did differently

None of this is an argument against automation. It is an argument against the five decisions above, each of which is avoidable:

• Measure resolution, not avoidance. Count whether the problem went away and stayed away. Deflection rate on its own will reward you for frustrating people — the same trap as leaning on Average Handle Time.
• Keep a real escape hatch, and make it easy to find. One click to a person, with context carried across so nobody repeats themselves. How that handoff is built is most of whether automation helps or harms — see designing the AI-to-human handoff.
• Automate retrieval, not judgement. Finding the right answer is a good job for software. Deciding what to do about an angry customer with a half-broken order is not, yet. Most of the value sits in documentation good enough to automate from, not in a bot improvising.
• Staff for the hard half. If automation removes the easy tickets, the remaining work is slower per ticket by design. Budget for that rather than treating it as underperformance.
• Keep people long enough to know things. Tenure is an operational asset, not an HR statistic. The cheapest knowledge base is an agent who has been there two years.
• Write down what you promise, and report the misses. A published support SLA turns service quality into something with a number attached, which is the only version of it that survives a budget meeting.

The short version

Customer service got worse because it was optimised, competently, against the wrong number. Cost per contact went down. The thing customers actually wanted — their problem solved, once, by somebody who understood it — was never on the dashboard at all.

The companies quietly winning on service right now are not the ones spending most on automation. They are the ones who kept a human reachable and measured whether the problem went away.