The cost of cost reduction: An AI governance playbook for customer leaders

In the second article of her AI governance series, Sue Duris explains why governance is usually sold as risk management when its most underrated job is economic

Add bookmark

Sue Duris
11/05/2026

In 2024, Klarna told the world its AI assistant was doing the work of around 700 customer service agents.

It was the headline every board wanted to be able to say out loud: the same service, at a fraction of the cost.

A year later, the company was recruiting people back, with its CEO acknowledging that cutting humans to hit a cost target had come at the expense of quality.

If your board has asked you to take cost out of customer service with AI this year, that story is worth sitting with, because you are not an outlier. You are where the market is pointed.

This article addresses:

Why cutting human headcount backfired for Klarna – and what your organization can learn from the story.
How effective AI governance can prevent false savings.
The million-dollar question: If your board told you to cut service costs with AI tomorrow, what would you automate, what would you protect, and what would you refuse to compromise on I?

Don't miss any news, updates or insider tips from CX Network by getting them delivered to your inbox. Sign up to our newsletter and join our community of experts.

Klarna's AI customer service reversal: Why cutting agents backfired

In Gartner's 2025 survey of customer service and support leaders, more than 80 percent said they expect to reduce agent headcount within the next 18 months, through a mix of attrition, hiring freezes, redeployment into new roles, and some layoffs. The direction is unmistakable: AI is being pointed at taking cost and headcount out of frontline service. Klarna just did it the loud way, with cuts, so it made the news.

And here is the problem that story exposes, which has nothing to do with the technology. They cut real cost. A year later, they were hiring the people back.
A cut you must reverse is not a saving; it is a cost with a delay.

Klarna paid to take the people out, absorbed the drop in service quality while the AI ran alone, and then paid again to recruit, rehire, and rebuild the capability it had removed.

A cut you must buy back is rarely cheaper than the team you cut.

This usually gets called a technology problem, or a cost problem. It is really a governance problem, though not the kind governance is usually sold as.

The hidden costs of AI automation in customer service

When you automate a block of customer contacts, the saving shows up immediately and lands on a line the CFO can see: fewer agents, lower cost-to-serve, a number you can put in a deck this quarter.

The cost of that saving shows up later, and it lands somewhere else – in churn that creeps up two quarters out, in repeat contacts from customers the bot couldn't resolve, in the experienced people who leave because the work got worse, in the institutional judgment that walks out with them, and sometimes in liability when an automated decision goes wrong.

None of those arrives with a label that says, "this is what the cost cut cost you." They arrive scattered, delayed, and usually in a different part of the business from the one that booked the win.

So, the saving is attributable, and the cost is not. That asymmetry is the whole problem.

It is what lets an organization record a win it has not actually earned, and it is why the Klarna story is about counting, not about AI. They measured what they saved. They did not measure what it cost.

By the time the cost was visible in the quality of the service, the saving had already gone into the results.

Efficiency vs. deferral: How AI governance prevents false savings

This is the line I would put in front of any leader being pushed to cut: efficiency that is not measured against the cost it creates is not efficiency. It is deferral. You have not removed the cost. You have moved it off a line you are watching onto lines you are not and handed the bill to a later quarter and often to a different owner.

Governance is what would have caught that before it happened: a decision about where AI acts, who is accountable for the customer, and what gets measured made before the cut, not after the complaints.

The answer is not to stop automating. A great deal of customer contact is pure cost – a balance check, an order status, a password reset – and automating it well genuinely lowers cost-to-serve without touching anything the customer values. The answer is to stop treating all contact as the same kind of thing.

Because it is not. Some of what a service organization handles is cost, and some of it is relationship, and the two look identical on a volume report. The status check and the fraud call are both "a contact", but one is a transaction you want to make cheaper, and the other is a moment where a customer decides whether they still trust you.

The automate, protect, refuse framework for AI governance

I was on a panel where we were posed a question I think every leader should sit with: if your board told you to cut service costs with AI tomorrow, what would you automate, what would you protect, and what would you refuse to compromise on I?

It is the right question, because it forces the sort of thinking a blunt cost target skips and because those three decisions - automate, protect and refuse - are what AI governance is in plain commercial terms. Not a compliance checklist. A set of decision rights about where AI belongs.

Automate: Low-stakes, high-volume customer contacts

What you automate is the pure cost: the high-volume, low-stakes, repeatable work where speed is the whole value. Automating it well lowers cost-to-serve without touching anything the customer values.

Protect: Customer relationship moments that build trust

What you protect is the relationship, the moments where someone is distressed, exposed, out of money, or deciding on something they cannot easily undo.

A complaint, a fraud case, a hardship call, the conversation that decides whether a customer stays. These look like cost on a volume report, but they are where trust is won or lost, and trust does not show up on the cost line until it leaves, as churn.

Automate them to save money, and you are not cutting cost; you are spending trust to book a saving.

Refuse: The lines AI should never cross in customer service

And what you refuse is the part most cost programs never name: there are things you do not automate at any price. You do not trap a customer in a system with no way to reach a person.

You do not let AI make a decision the organization cannot explain or stand behind.

Those are not customer experience niceties; they are the line between a saving and a liability, and drawing that line is governance doing its clearest job.

How smart AI cost reduction protects customer retention and revenue

This is also where the efficiency case and the growth case stop being opposites. A cut that protects the relationship moments and automates the transactional ones does not just cost less to run; it keeps the customers and the judgment that generate revenue.

A cut that treats every contact as cost saves money on paper and loses it in the places no one was counting.

Same AI, same budget line. The entire difference is whether the organization knew which contacts were cost and which were relationship, and refused to confuse the two.

It is worth noting that Gartner, reporting that headcount figure, framed the smart path for 2026 as a move away from cost-cutting and toward value, the same point from the other direction.

The organizations that win with AI in service are the ones that cut the right things and could see what the cutting cost. Klarna, to its credit, noticed and reversed.

Most organizations will not get the clean before-and-after that a public company hands the press. They will get the slow version, a saving that looks fine for three quarters and a customer base that is quietly thinner by the fourth, with no single moment connecting the two.

The question every CFO, COO, and CRO should ask before cutting service costs with AI

So, before the next AI cost initiative, the most valuable question a CRO, COO, or CFO can ask is a simple one: what will this cut cost us, where will that cost land, and how will we see it coming?

If the only number on the table is the saving, you are not looking at an efficiency case yet. You are looking at a deferral, a governance gap with a delayed invoice.

Efficiency and effectiveness are not the same number. Governance is how you tell them apart. The job is not to stop cutting. It is to stop confusing the two.