Should you cut support headcount because of AI?

Gartner's answer, published in February 2026: probably not yet, and if you already did, there's a good chance you'll be hiring back. The firm predicts that by 2027, half of the companies that attributed a customer service headcount cut to AI will rehire people to do similar work — often under a different job title. A Forrester survey cited across trade press this summer puts the regret rate even higher: 55% of leaders who cut jobs citing AI now say the decision was wrong.

This isn't a fringe finding. Through July and August 2026, CNBC, Forbes and a string of HR and CX outlets have covered the same reversal from different angles, all pointing at the same two public examples: Klarna and Salesforce.

What actually happened at Klarna and Salesforce?

Klarna cut roughly 700 support roles in favor of an AI system and, within about a year, started rehiring. CEO Sebastian Siemiatkowski's own explanation was specific: the AI handled volume but not complexity. Satisfaction scores dropped as edge cases, emotionally charged interactions and multi-step problems piled up on a system built for routine queries — and repeat contact climbed, meaning customers came back two or three times for the same unresolved issue.

Salesforce's move ran the other way in scale — from roughly 9,000 support staff to about 5,000, with its own Agentforce product handling close to half of conversations — and is a live example of the same tradeoff still playing out, not a settled counter-case.

Neither story says AI support doesn't work. Both say the same thing: the part of your ticket volume that AI can't safely resolve doesn't shrink just because you announced a headcount cut. If nobody is covering it, it shows up as worse service, then as a rehire.

Why does the regret keep happening?

Because headcount and automation rate get treated as the same decision, and they aren't. A support system can be genuinely good — resolving a real, measured share of tickets — while the unresolved share still needs exactly as many skilled people as before, sometimes more, because the queue that's left is the hard part by definition: the complaints, the exceptions, the accounts where being wrong is expensive.

Cutting on the promise of the automation rate, before you've measured what's actually left over, is the step every reversed decision has in common. Cutting after measuring it — and sizing the team to what genuinely remains — is a different decision, and it's the one that doesn't make the news a year later.

What to look at before you touch headcount

Three numbers, not one:

• Resolution rate on your real tickets. Not the deflection number from a vendor demo — the share of your actual queue, half-written emails and all, that gets solved without a person. This is the number that tells you how much repeatable volume you genuinely have.
• Misroute and escalation-agreement rate. How often does a person have to fix what the system decided? A high number here means the "resolved" figure is optimistic.
• Repeat contact on the same issue. The earliest warning sign in both public reversals. If customers are coming back for the same problem, the queue you think you automated is quietly growing again — just with a delay before it reaches a person.

None of these is a number we invent for you. They come from your own ticket history, over a real measurement window, before any commitment is made about staffing.

What does a headcount-safe automation project look like instead?

The architecture that avoids this trap is the one built to escalate on purpose: confidence thresholds, complaint language, money above a threshold and repeat contact all route to a person by design, not by failure. On an industrial components distributor running this pattern, 85% of tickets are created and classified automatically from inbound email, and the team got back roughly 58 hours a month — without a claim about headcount, because that was never the metric we were selling.

The team doesn't shrink because the system decided it could; it gets to spend its time on the part of the queue that was always going to need a person, with less of the repeatable work sitting on top of it.

What would your numbers say?

The free AI Audit takes about five minutes and gives you a straight read on what share of your support queue is realistically automatable today — the number to look at before any staffing conversation, not after. Want to walk through your ticket data with us first? Book a call. The packaged build is Customer Support Automation; the resolution-rate math behind these numbers is in what resolution rate is realistic.