The Bolt-On Trap

Take a refund workflow in which an agent writes a reply and then waits for a manager’s approval. Suppose AI shortens the writing but leaves the manager’s review unchanged. If that manager is already working at capacity, faster drafts join the same queue; they can’t raise the rate of completed refunds beyond the manager’s limit. The saved writing time is real. Turning it into lower cost or more completed work requires another decision.

The constraint can move

That distinction sits at the center of Eliyahu Goldratt’s Theory of Constraints: improve the step that limits the whole process, then look again when the limit moves. The Theory of Constraints Institute’s explanation specifically warns against letting inertia become the next constraint. Old rules can outlive the reason you needed them.

Applied to the refund example, that gives you a hypothesis to test. If drafting used to limit completed work and AI has made it faster, approval may now be the limit. If approval was already the limit, the tool has improved a different part of the process. But suppose the tool also supplies the verified purchase details the manager used to chase down. That could increase the manager’s capacity without changing who approves a refund. You have to follow the work far enough to distinguish those cases.

An assistive tool can still earn its price. In Generative AI at Work, published in 2025, researchers studying 5,172 customer-support agents found that AI assistance increased issues resolved per hour by 15% on average. The gains varied: less experienced and lower-skilled workers improved speed and quality, while the most experienced and highest-skilled workers saw small speed gains and small quality declines. The average supports neither a uniform staffing cut nor a uniform rollout.

That study gives a reason to measure the value of helping people do their existing work. A reorganization has to earn its additional cost. The bolt-on trap begins when you treat a faster task as proof that the whole operation got cheaper, without following the work through review, exceptions, and the customer’s eventual result.

What 700 agents left unanswered

Klarna’s experience makes the staffing question harder than an automation headline suggests. In February 2024, the company said its assistant had handled two-thirds of customer-service chats in its first month, doing work equivalent to 700 full-time agents. It also reported customer satisfaction on par with human agents. Those were Klarna’s own measurements, and the same announcement said customers could still choose a live agent.

By May 2025, CEO Sebastian Siemiatkowski said his pursuit of AI-fueled cost-cutting in customer service had gone too far, according to his interview with Bloomberg. Klarna’s spokesperson told CX Dive the company was piloting a new human customer-service workforce.

Yet in a direct interview with Big Technology that month, Siemiatkowski said the AI assistant was taking on more complex work, while Klarna hired people for higher-end conversations it had previously outsourced. The company was expanding what the assistant could handle and reconsidering who should handle the rest.

Those accounts leave the net cost and quality effects of the revised arrangement unresolved. Our inference is narrower than either a victory lap or a retreat: a workload-equivalence figure leaves out the decision about which conversations need a person and what that person must be able to do. It can’t settle the staffing plan on its own.

You already made a design decision

When you approve the tooling budget and leave every approval rule intact, you make an organizational design decision. You make it by default. The company has chosen to preserve the old handoffs while changing how quickly work reaches them.

Keeping a handoff can be right. A refund approval may catch fraud or protect a contractual obligation. Removing it because the reply now takes less effort would confuse the cost of writing with the reason for review. The useful question is which cases still need that review, and what evidence would justify a different route for the others.

Changing that route has a cost a tooling budget doesn’t cover. Redesigning a workflow can mean renegotiating decision rights, redrawing team boundaries, and telling people that work they built a career around is now the part the machine does. That’s a political cost, paid by you, in rooms the vendor isn’t in. Someone still has to own the disputed refund after the subscription is renewed.

Follow the whole refund

Your support lead can start with existing ticket records and the people who review exceptions. Follow a defined class of refunds from the customer’s request to the completed refund, separating time spent working from time spent waiting. If requests wait for a manager who reviews them in batches, changing the review schedule may help. If the delay is missing purchase evidence, supply that evidence before removing an approval.

For a refund inside a clear policy, with the purchase verified and the amount within an authorized limit, a shorter route might mean automatic approval. A simple rules check may be enough; the model could handle the wording while the rule decides eligibility. Ambiguous eligibility would retain a human decision. A model declaring itself confident leaves the answer unchecked.

Test that rule against past requests without issuing refunds. Have the existing reviewers check proposed approvals and refusals against the policy, including cases that were disputed or reopened. A replay can expose a bad eligibility rule; it can’t establish how well the process will handle cases absent from those records. Keep the current route where evidence is thin or the consequences of a mistake warrant it.

Compare any bounded live trial with the current route on comparable requests, following both through corrections and repeat contacts. Count the full cost of resolving that mix of requests, including software, rule upkeep, and human work in escalations. If automation takes the easy cases, a rise in human handling time can simply mean the remaining queue is harder. Counting only the assistant’s work would miss that burden; judging only the human queue would miss the work automation removed.

If assistance meets your cost and service goals and review still earns its place, keep that arrangement. The proposed shorter route owes you the same proof as the old one. Expand it only where observed outcomes support doing so, before reorganizing around the capacity you expect it to free.

Then decide what to do with any capacity you’ve actually freed. A salaried agent finishing replies earlier doesn’t reduce the wage bill by itself. Absorbing more demand without another hire, reducing overtime, and improving service have different financial consequences. Put the outcome you chose, and its measured result, in the next board review. The adoption chart can stay in the appendix.