The downturn playbook for AI-native firms: cut the fat, not the capability

If revenue drops and you need to cut costs, I would look very carefully at what you are about to lose before deciding that AI lets you keep the same business with fewer people.

You may need fewer hours to produce the work, that part can be true. But somebody still has to look after the client, catch the mistake, improve the service and do the work that makes the next renewal easier, and those hours need to be in your calculation too.

My recommendation is to find the waste first, protect the parts of delivery that clients depend on, and use the time AI actually frees to keep those clients and win more of the right work. Have a fallback plan as well, because telling an owner to keep investing without asking how long the cash will last is not useful advice.

I don't have franchise-network data showing how this plays out through a recession. This is the approach I would take, drawing on what previous downturns tell us and what we already know about AI in delivery, not a result I have measured across firms.

And you don't need to agree that a recession is coming to do this work. A client can leave, a renewal can get smaller, a strong pipeline can stop turning into signed work, your own firm can come under pressure while the wider economy is doing fine.

The hard part is that the quick cut usually looks very clear. You can see a salary leaving the monthly bill, while the cost of losing that person's judgment or their knowledge of the client shows up somewhere else, sometimes much later.

That doesn't make every job untouchable. It means you need to follow the saving through the business instead of stopping at the line where you found it.

There is a useful piece of recession history here. In Roaring Out of Recession, published in Harvard Business Review in 2010, the researchers looked at 4,700 public companies across three recessions and compared how they cut and invested.

The firms most focused on cutting costs had a 21 percent probability of pulling ahead of competitors after the recession. For the group that combined selective cuts with continued investment, that figure was 37 percent.

Those are results from groups of public companies, not the odds that your AI plan will work. The study predates this kind of AI, and it does not prove that keeping a particular tool or employee causes a better recovery.

What it gives you is a reason to question the idea that the deepest cut is the safest choice. The firms that did better were cutting too, they were making different decisions about what to protect.

Sony's experience in that study is worth staying with. Around the 2000 downturn, it cut its workforce by 11 percent, research and development by 12 percent, and capital spending by 23 percent.

Its margin improved from 8 percent to 12 percent, which would look like a successful response if that were the only number you watched. But its average annual sales growth went from 11 percent before the recession to 1 percent afterward.

Now look at the other approach. Bain's 2019 recession analysis describes Samsung continuing to invest in research and marketing through the 2009 downturn, while competitors pulled back.

That is a large technology company, not a small consultancy, so I would not copy its spending plan. The useful comparison is what each company was trying to preserve, one was protecting the immediate margin, the other was also protecting what it would have to sell when demand returned.

Research spending sounds remote from the week of a service-firm owner, but think about the work that makes your AI useful. Someone has to learn the tools, write down how a good job is done, test the checks, fix the failures and help the team use the new workflow.

That work costs money before it gives you a dependable saving, which makes it very easy to remove when you need the bill to come down quickly.

An NBER summary of research on financially stressed firms during the 2008 crisis found that cuts to research persisted beyond the crisis. The authors discuss adjustment costs as an explanation for why firms don't simply restore those teams when conditions improve.

My comparison is that learning and workflow redesign deserve the same care in an AI budget. I am not calling your training afternoon basic scientific research, I am saying you can cut the work that creates a future capability and then discover that buying it back is harder than keeping it going.

We also have examples of the cost-first AI approach before any future recession arrives. Klarna said its assistant was doing work equivalent to 700 customer-service agents, and later moved employees back into customer support after quality concerns.

The 700 figure was a claim about work capacity, not a count of 700 people laid off in one action. And the CEO's criticism of going too far concerned cost-cutting broadly, so I would not turn that story into proof that every use of AI in support failed.

It is a warning about what you count as success. Handling the volume is one test, taking proper care of the person asking for help is another, and you need both before removing the people who handle the difficult cases.

In Gartner's February 2026 release, only 20 percent of surveyed customer-service leaders said they had reduced agent staffing because of AI. Most reported steady staffing while supporting more customers.

Gartner also predicted that half of companies attributing customer-service cuts to AI would rehire for similar work under different titles by 2027. That second part is a forecast, not a rehiring result already on record.

The first part matters more for the choice in front of you. Keeping the team and serving more customers is a real option, it just doesn't make the same headline as a large headcount reduction.

And a cost target is not a saving yet. In Bain's 2026 survey of 951 companies, 37 percent targeted cost reductions of 11 to 20 percent, while nearly 40 percent of those measuring outcomes reported reductions in the 0 to 10 percent range.

Those figures describe the target and result distributions, they do not tell us that every firm aiming for the first range landed in the second. They are still a good reason to keep checking the actual result rather than treating the budget proposal as money already recovered.

Now, AI does create useful capacity. The NBER working paper on 5,179 customer-support agents reported a 14 percent average increase in issues resolved per hour, with a larger gain among novice and lower-skilled workers.

That is people doing the work with an assistant, not a test of removing the people. If you use the finding to plan layoffs, you have added a decision the experiment did not test.

There is a similar boundary in the consulting research. In Ethan Mollick's account of the BCG consultant experiment, consultants using AI did more work, faster and at higher quality on tasks within the model's abilities, but performance got worse on a task outside them.

Somebody still needs to know which kind of work your service contains. You cannot take the gain from one part of that study and leave the failure condition out of your staffing plan.

There is another reason I wouldn't start with headcount. You can make a task faster without getting useful time back across the whole job, and you can get time back without deciding what it should be used for.

Those are separate problems. If review and rework consume the saving, the hours aren't available yet, and telling the team to take on more clients would give them work the new process cannot support.

If the hours really are available, leadership has to assign them. That is the problem I describe in the extra time AI buys you, a tool improves what people can do, but management has not changed what they are expected to work on.

My hypothesis is that this helps explain some disappointing AI returns, and that those disappointments could feed wider pressure to cut. The OECD's March 2026 interim outlook names lower-than-expected AI investment returns as a possible financial-market risk, but it does not establish that unused employee hours are causing that risk.

For your firm, the useful part is much smaller than the economic argument. Before counting a saving, decide how you will find out whether it exists and what work will receive it.

I would start with your profit-and-loss statement and the people who know how delivery actually happens. Look at tools nobody uses, vendors doing overlapping jobs, work you keep selling despite its poor margin, and steps everyone follows without knowing why.

There is no honest percentage I can promise you will find there. But you should know what the waste is before cutting your ability to deliver good work, because removing an unused subscription and removing the only person who can fix a troubled account are very different savings.

Then write down the cutting order while you can still have a calm conversation about it. Put customer experience, service quality and the technology that improves them near the end, and name what evidence would make you change that order.

An AI project doesn't get protection just because it has AI in the name. If it adds cost and still has no credible path into delivery, you need to question it like anything else, while distinguishing it from a working system that clients already depend on.

For an automation that saves time, put the next assignment beside it. If, for example, you establish that it frees 20 hours a month, decide who will use those hours and which clients or delivery problems they will work on.

In a downturn, my first choice would be retention. Give the manager time to catch a service problem earlier, help the client get more use from what they already buy, or improve the work that keeps coming back for correction.

That can support new business too, but choose the work deliberately. Sending every client more AI-generated material is not the same as giving them better service, especially if they now have to spend their time finding the useful part.

I would keep looking for good opportunities while other firms pull back. Bain's research on the early-1990s recession found that more than a fifth of bottom-quartile companies moved into the top quartile during the downturn, and many gains persisted through the recovery.

Again, that is historical evidence that positions can change, not permission to spend money you don't have. Taking a useful risk needs a floor underneath it.

So set the fallback plan as carefully as the growth plan. Decide which revenue or cash conditions trigger another cut, what gets paused, and who can make that call, then look at the effect on delivery before you approve the list.

If cash is running out, survival can require cuts to capability too. This playbook cannot remove that tradeoff, its purpose is to help you avoid arriving there after cutting the things that might have kept your clients.

Sit down with the person who owns delivery and write those decisions before the next difficult budget meeting. You need to know what you will protect, what you will stop, and where any real saved time will go.