People argue ‘AI will remove jobs’ against ‘AI will only augment people’ as if one of them has to win. Inside a company the debate can wait. The early change is to capacity, which is one of the hardest numbers to pin down, and a lot has to happen before any of it reaches headcount.
A model cuts a report from four hours to one, and that’s a real change to the task, which is also the only thing most people measure. The workflow around the task may not speed up at all, because the report waits two days for a review either way, and the bottleneck was never the writing. A workflow gets more capacity when its slowest step changes, and that’s usually not the step the model touched. Speeding up the writing while the queue is at review just grows the pile in front of the reviewer.
The analyst who writes that report also fields requests, checks a feed, covers a colleague and holds context nobody wrote down, and her role is that whole bundle. AI usually takes over a few of those tasks and leaves the rest. That’s why saved minutes don’t add up to positions you could actually remove: a position comes out when a whole bundle disappears from one person, and AI mostly shaves a slice off many bundles, at least for now.
The same tool in two sets of hands produces different results, and the gap between functions is even wider. Drafting and summarizing speed up a lot; the jobs that are mostly meetings and judgment calls barely change. Two teams next to each other on the org chart can be a year apart in what the tools do for them, and a finance team that closes faster says nothing about what the sales team got. An average across the company describes nobody, so capacity planning has to happen team by team, where you can see the spread.
Teams keep queues that never appear in any system, like the requests they decline and the analysis they’d do with more time, and that’s where they put the freed hours, before anyone counts them. Some of that is the best possible use of the hours, because it’s demand the team could never get to before. The rest is harder to spot: extra polish, and extra iterations the decision never needed. From the outside it all looks the same. Everyone’s week is still full.
Someone reads the model’s output, and the strange cases still go to a person. When mistakes are expensive, that reading takes back a lot of the gain. A team that used to write forty answers a week and now checks a hundred is spending the same week, on reading instead of writing, and task time won’t tell you any of it happened.
A team that was cleared to hire and covered its growth without the hires made a saving workforce planning can book, and that’s the easiest kind to prove. Most claims are looser than that, because most of the hires being avoided were never approved to begin with. The claims without paper behind them are guesses.
Once somebody has counted the spare hours, a team can shrink by attrition, or pick up work the company approved and never started, like the integrations that kept slipping and the documentation nobody had time for. It can cover the customers who’d been getting the short version of everything, or run with room to spare on purpose while the checks and standards are still new. If nobody makes the choice, the team still spends the hours, just not on anything anyone picked.
Before anyone touches a headcount assumption, sample the work and check what the team actually spends its week on now. By sampling I mean sitting with the work for a day, or pulling a week of output and walking through where it came from, which is slower than a survey and harder to fool. Look at what the queues nobody writes down were holding, whether errors and rework went up or down, how many hours the checking uses, and which teams changed their output and which just changed their pace. Workforce planning and finance need that picture more than they need a forecast, because next year’s plan is being built on last year’s capacity right now.
Teams watch what happens the first time someone reports spare capacity. If it turns into a layoff list, nobody files a second one, and suddenly no team has spare capacity on paper. That’s self protection, and it works, which is the problem. From then on the company plans blind. Saying in advance what happens to freed capacity, and meaning it, is the only way I know to keep the information coming.
‘AI will remove jobs’ assumes the minutes add up into positions on their own. They don’t, not until somebody redesigns the workflow around the new capacity. And the roles do change inside specific workflows, when a company redoes the bundle on purpose, which is what ‘AI will only augment people’ misses. I think the slogans describe where things might end up. Until then there’s a company to run.
Argue about the spare hours instead. That goes better when somebody’s planned for it, and much worse in the week a budget is due. A company that has sampled the work can pick from real options. Without the sample, the argument falls back to the slogans.