For the owner of a product business between GBP 1M and 10M who is about to hire, or about to not hire, and wants to know what AI actually changes about that decision.
The short answer. Fewer than you think, and not for the reason you think. AI is not mainly taking over the work. It is taking away the reason you hire people to coordinate other people. So the useful question is not how many people you need. It is what is the smallest group who can own a whole result, and what should the machines do so those people have time to think.
Everybody is arguing about whether AI takes jobs. That is the less interesting question. The interesting one is what it does to the cost of organising, because the cost of organising is what your org chart is actually made of.
Ronald Coase asked in 1937 why firms exist at all. If markets are so efficient, why not just buy everything from outside?
His answer was transaction costs. Searching, negotiating, agreeing, checking, chasing. When doing all that inside is cheaper than buying it outside, you build a company. And it grows exactly as far as that stays true.
So what is a firm? A machine for making coordination cheaper.
Now think about what you actually pay for. Not only the work. You pay for somebody to know who is doing what, somebody to chase it, somebody to write it up, somebody to check it, and somebody to decide when two of those people disagree.
Hierarchy was the answer to a bandwidth problem. Luis Garicano's model from 2000 puts it plainly: the common problems get solved at the bottom, the rare ones get escalated up to whoever knows more, so nobody has to know everything. One person cannot hold 500 people's questions in their head. So you build layers. Ten to a manager, ten managers to a director, and up it goes.
That is the tree. It was never really about status. It was about how much one head can hold.
Took me years to see this. There are lot of things in a company that look like management and are actually just information moving between two people who could not see the same screen.
Here is the part I did not expect.
A study published in Management Science in March 2026 looked at 3,017 US public firms from 2010 to 2019, then a matched sample of 622 treated firms and 941 controls. The treatment was boring. Adopting collaborative work management software. Jira, Smartsheet, that kind of thing.
| After adoption | Change |
|---|---|
| Managerial intensity | down 3.2% |
| Delegation to non-managers | up 5.2% |
| Lateral coordination between peers | up 7.2% |
Small numbers. But it is the same direction, and it happened before generative AI existed. Software took one slice of the manager's job, which was moving information around, and the organisation changed shape around the hole.
And the authors make the distinction that matters more than the percentages. The technology can either make workers better at running themselves, which pushes decisions down, or make managers better at watching people, which does the opposite. Same tool. Two very different companies.
Hold that thought. It comes back later.
The strongest result I have found is a field experiment run at Procter and Gamble with 776 professionals.
A single person working with AI produced work rated at roughly the same level as a team of two people without it. It also softened the wall between the R&D people and the commercial people. The ones using AI produced more balanced solutions regardless of which side of the business they came from.
So think about why you add somebody to a team. Usually it is this: I need you because you know the thing I do not know. If AI supplies enough of that missing thing, the reason to add a person gets weaker. Not gone. Weaker.
And before anybody tells you the optimal team is five people. It is not. A 2023 meta-analysis pooled 208 samples and roughly 21,435 teams and found the corrected correlation between team size and task performance was zero. Not weak. Zero.
Which does not mean size is irrelevant. It means the right size depends on how much different human knowledge you need, against how much coordination those humans then cost you. AI moves that ratio. It does not hand you a number.
Not the people making the product. The people moving information about the product.
| What a manager does in a week | How much of it AI can take |
|---|---|
| Collecting information | Nearly all |
| Writing the report | Nearly all |
| Tracking whether the work happened | Most |
| Preparing the priority call | Most |
| Passing expert knowledge down | Most |
| Making the call | Varies |
| Resolving a fight between two people | Little |
| Coaching | Little |
| Building trust | None |
| Carrying the blame | Still a person |
Look at the top half of that table. Then ask what your middle layer actually spends its week on. If the honest answer is the top half, that job is about to change shape whether anybody plans it or not.
Amazon told its organisation in 2024 to increase the ratio of individual contributors to managers by at least 15%. Jassy's stated reason was not cost. It was that the layers had produced pre-meetings, approval chains and less ownership. Shopify went further in 2025 and made AI the default: before a team can ask for a new headcount, it has to explain why the work cannot be done with AI instead.
Neither of those is a prediction. Both already happened.
I will use my own, because it is the only one I can put real numbers on.
Rutherford Wren is a consumer product business doing roughly GBP 3 million a year across own site and Amazon, in UK and US. Three people.
Eleven years ago that shape would have needed a buying assistant, a stock controller, a marketing coordinator, a customer service person and a bookkeeper. Five salaries, plus somebody to manage the five, before a single product had been designed.
What replaced them is not one clever robot. It is a set of very unglamorous agents doing the shovel work. Pulling the numbers, reconciling them, flagging the SKU whose margin has slipped, drafting the listing, watching the stock cover, writing the weekly summary nobody wants to write. One of them exists entirely to notice that a supplier invoice does not match the purchase order. It has saved more money than the last three marketing ideas.
The judgement did not move anywhere. What to buy, at what price, from whom, and when to stop. That is still a person, and it is still the part that decides the year.
I would love to say this was a plan. It was not. We ran out of money to hire and had to find another way, and then it turned out the other way was better. That is not a strategy anybody should copy on purpose.
This is the part missing from almost everything you will read on this.
The same technology builds two opposite companies.
If AI helps your team find the answer, make the call and coordinate their own work, then you need fewer managers and the decisions move down. That is the version being sold in every keynote.
But if AI helps one person at the top watch, measure and direct an enormous amount of work, then you also need fewer managers. And the power goes up, not down. Fewer people with more control over more of it, and now in real time.
Same headcount reduction. Opposite company.
So when somebody tells you AI will flatten organisations, ask which of the two they mean. A company can remove half its managers and hand what is left of the leadership a live view over everything that moves. The hierarchy did not go anywhere.
It just got a login.
The distance between the slide deck and the Tuesday morning is enormous.
Deloitte surveyed 501 senior executives between April and June 2026, all of them already piloting agents. 74% expect nearly half their business processes to be redesigned or rebuilt around agents within four years. 5% said their processes are highly prepared for agents today. 15% have reached scaled, cross-functional multi-agent use.
74% believe. 5% are ready.
And then there is Meta. Reuters reported on 26 August 2026 on an internal programme called Project OT, short for Organization Transformation, hatched at a leadership retreat in January, which aimed at replacing up to 60% of roles with agents overseen by small groups of human staff. A first wave of layoffs went ahead in May at 10% of staff. The November wave was cancelled. AI-assisted code went up sharply, product improvements actually reaching users did not follow, and the internal employee sentiment score fell from 74% favourable to 55%.
That is the most expensive version of this experiment anybody has run in public. Worth reading before you plan your own.
If AI does the junior work, then where does the next senior come from?
Stanford's Digital Economy Lab, using ADP payroll data through June 2026, found employment for 22 to 25 year olds in highly AI-exposed occupations sitting about 19% below where it would be if it had tracked the less exposed roles. That gap was 15% in July 2025. The authors are careful and say plainly they cannot yet prove AI caused it.
The detail I keep coming back to is which jobs. The decline sits in codified knowledge work, the documented, standardised, written-down kind. Roles built on tacit knowledge, the sort you only get by doing it badly for a few years with somebody watching, held up better, and for experienced people they held up best.
Which is exactly the work juniors used to do in order to become seniors.
I do not have a good answer to that one.
Three things, none of which need software.
One. Write down what your last three hires actually do all week. Not the job description. The week. Then split it into work that makes the product better and work that moves information about the product. If the second pile is more than half, you have an information-routing problem, not a headcount problem, and hiring will not fix it.
Two. Before the next hire, write the sentence. The one Shopify makes its teams write. Why can this work not be done with AI? If you cannot answer that in three lines, you are not ready to hire either way. And a hire is a cash decision before it is anything else.
Three. Pick one result, not one function, and give it to two people. Not "marketing". Something like "the autumn launch lands on time and at the margin we modelled". Give them the numbers and the tools and leave the reporting to the machines. Then watch how much of the coordination you thought was essential simply does not happen.
No. It means stop hiring for coordination and keep hiring for judgement, taste and relationships. The roles that shrink are the ones whose week is mostly collecting, formatting and passing on information. The roles that get more valuable are the ones deciding what to buy, what to charge and who to trust.
No, and the research is blunt about it. Across 208 samples and roughly 21,435 teams the corrected correlation between team size and task performance is zero. Small is not a strategy on its own. Small, with the right people, and less coordination needed between them, is.
This is the real risk and it does not get enough air. Work published in Science Advances found generative AI raised the average quality of individual creative output while making those outputs more similar to each other. Everybody gets better and everybody converges. If your advantage is seeing the market differently from your competitors, count that cost before you count the salaries.
The recurring report nobody wants to produce. Low risk, easy to check, and if it goes wrong you find out in an hour instead of a quarter. More on that in what AI is actually for in a product business.
Only 5% of organisations in Deloitte's survey said their processes are highly prepared for agents today, so waiting for readiness is a long wait. But the Meta story says planning a 60% cut on a slide is not the alternative either. Start with one process, one number, and a date in the diary when you check whether it actually worked.
Anyway. The old question was how many people do we need.
The better one is which people should be in the room together, and what should the machines do so that those people have time to think, decide and talk to somebody.
I do not know the right answer for a business of ten. I am fairly sure it is not fifty.
Back to the spreadsheet.
Ronald Coase, "The Nature of the Firm", Economica, 1937. Luis Garicano, "Hierarchies and the Organization of Knowledge in Production", Journal of Political Economy, 2000. Collaborative work management technologies and managerial intensity: Management Science, published online 20 March 2026, 3,017 US public firms 2010 to 2019, matched sample of 622 treated firms and 941 controls. Procter and Gamble field experiment with 776 professionals: "The Cybernetic Teammate", NBER working paper 33641, 2025, since published in Organization Science. Team size meta-analysis: Journal of Organizational Behavior, 2023, 208 samples and approximately 21,435 teams, corrected correlation .00. Amazon manager ratio: Amazon company news, Andy Jassy, September 2024. Shopify AI-before-headcount memo: reported April 2025. Deloitte agentic AI readiness survey: 501 senior executives, fielded April to June 2026, released August 2026. Meta Project OT: Reuters special report, 26 August 2026. Young worker employment gap: Stanford Digital Economy Lab, ADP payroll data through June 2026. Creativity and homogenisation: Science Advances, 2024. All retrieved 29 August 2026.