The Headcount Deck Is Still Due at Nine.

The awkward case is the person who wrote the support runbooks. If the assistant’s good results depend on their corrections, removing their role changes the conditions under which those results were measured. The spreadsheet has real people in column A. Some of them may be dependencies as well as expenses.

Two drafts of tomorrow’s deck

The first draft starts with payroll. Current ticket volume, expected AI capacity, a smaller team to handle the rest. That can be a reasonable destination. It says little about how the business gets there.

The second draft starts with the work behind the capacity estimate. Which questions can the assistant resolve unaided? Who corrects it when an answer is plausible but wrong? Who decides that a billing dispute warrants an exception, then explains the decision to a customer?

Consider an assistant that drafts replies while experienced staff review every answer. Its performance includes that review. A plan that budgets for autonomous service has to establish what happens when the reviewer steps away. Otherwise it counts the benefit of their judgment and the saving from removing it at the same time.

The employee can run that calculation too. If their role is disappearing, they have reason to consider leaving before the transition is complete. You can schedule a reorganization; you can’t assume everyone will stay to implement it. If the person whose judgment made the replacement look feasible leaves before that judgment has been transferred, you’ve automated around a capability you no longer have.

That’s a conditional risk, not a prediction about how every team will react. It matters wherever the plan depends on knowledge or cooperation it hasn’t secured.

A gain with people still in the system

There is evidence that AI can spread expertise rather than simply consume it. In Generative AI at Work, published in 2025, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a staggered rollout of AI assistance among 5,172 customer-support agents serving a single software company. Access increased issues resolved per hour by 15% on average. Less experienced and lower-skilled workers improved in both speed and quality; the most experienced and highest-skilled workers saw small speed gains and small quality declines.

The design matters. The system learned from past support conversations, giving extra weight to those of top performers. It suggested responses and relevant documentation. Human agents retained authority to accept, edit, or ignore the suggestions.

Our inference for the headcount deck is that some of the team’s expertise can become useful to other employees through software. With demand and service quality held constant, higher productivity could support fewer staffed hours. That is a route to savings with people still using the tool; the measured gain doesn’t establish what an unattended assistant can resolve. The researchers explicitly limit their findings to medium-run effects at one firm, without estimating aggregate employment or wage effects.

That gives the second draft an alternative to immediate substitution: use the assistant to distribute knowledge, then evaluate what work still needs its current owner. It also challenges the comforting claim that every experienced person’s judgment is permanently irreplaceable. For the runbook writer, the distinction is between knowledge the assistant already carries and corrections it still needs them to supply. The transition plan has to establish how much of that ongoing work can move, and to whom.

Klarna’s numbers need their context

In its February 2024 launch announcement, Klarna said its AI assistant 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 the company’s own first-month measures, and the announcement explicitly kept the option to speak to a person.

The staffing claim needs care too. In a March 2024 interview with CBS News, CEO Sebastian Siemiatkowski explained that these customer-service agents worked for outsourced providers. He distinguished the reduction in purchased support capacity from Klarna’s earlier employee layoffs.

By May 2025, Bloomberg reported that Siemiatkowski believed the pursuit of cost-cutting in customer service had gone too far and was planning recruitment to preserve access to human support. Business Insider later corrected its reporting to clarify that his criticism concerned cost-cutting broadly, rather than AI implementation specifically.

Klarna’s public numbers describe a change in service capacity and a later concern about staffing and quality. They don’t identify the share of the quality problem caused by AI, nor measure whether the recruitment plan fixed it. Using either the 700-agent figure or the hiring announcement as a staffing model for your team would skip the work your deck has to do.

What belongs beside the savings

The useful distinction is between a job’s current shape and the capabilities the business needs from it. A support role might combine repeatable answers with product diagnosis and authority to make exceptions. Automating the answers changes the role. Whether it eliminates the role depends on who or what can do the remaining work.

Strategy without a transition plan is a press release. For a comparable volume and mix of cases, compare the human time today’s service consumes with the time the proposed service would still need. Include corrections and review, as well as the cases that reach a person. Separate initial training and handover from recurring work such as updating answers when the product changes. A temporary transition cost can end; a recurring dependency needs an owner after launch.

A reduction in those hours is human capacity released. To turn it into payroll savings, show how the remaining work fits into the smaller team’s actual shifts and responsibilities, then include the tool’s costs in the financial model. If minutes are saved across several roles but the same people must remain available for exceptions, counting those minutes as an eliminated position gets ahead of the evidence.

That comparison needs a service standard, not just a timer. Define acceptable resolutions and which mistakes require intervention before the trial. Compare the proposed arrangement with the current service on a similar mix of cases, counting repeat contacts and the seriousness of errors. Where it can be tried safely, use the level of review the future staffing plan actually funds.

Record every correction and escalation, including who handled it and how long it took. That makes review answerable to evidence too: if the assistant meets the agreed standard without a reviewer on certain cases, those cases can release reviewer time. If the trial’s success depends on an expert stepping in, budget for that dependency or test a replacement for it. Keeping expert review throughout a trial establishes assisted performance; removing it remains a separate claim.

The support lead and the people handling difficult cases can identify those dependencies with the technology and finance leads. When those people have the expertise to evaluate the work and the authority to change it, they can own the transition internally. Give them time and a mandate.

If the business needs immediate savings, the deck can still state the unresolved assumptions and the service risks being accepted. When it can afford to validate the transition first, the results have to be allowed to change the staffing plan in either direction. Neither choice earns a promise that everyone’s role will survive.

The revised deck may still propose a smaller team. It should be able to name who handles the first case the automation can’t resolve, and show that person has agreed to do it.