AI Adoption

AI Won't Shrink Your Team. It Will Change Who You Hire.

August 31, 2026

Every leadership team arrives at the same sentence eventually, usually in a budget meeting, usually said more carefully than it is meant. How many of these open roles do we still need to fill, now that we have this?

Fair question. Answering it first is how businesses cut in the wrong place, feel clever for two quarters, and then quietly conclude that AI did not work for them.

The headcount question is the wrong first question

The instinct behind it is sound enough. You bought a capability, the capability does work that people used to do, so the saving should turn up in payroll. That logic held for machinery. It holds badly for this.

PwC’s advisory framing puts the split at about 20% of an AI initiative’s value coming from the technology and the other 80% from redesigning the work around it. Read that as guidance rather than as a measured finding, because that is what it is. The direction it points, though, is the direction the survey data has been pointing for two years running. In McKinsey’s 2026 global State of AI survey, nearly three-quarters of the high performers, the small group who can attribute real EBIT impact to AI, report fundamentally redesigning workflows because of it. Among everyone else, about one quarter say the same.

Both groups licensed the same models. The models are not where they diverged.

Redesigning a workflow is unglamorous work. Someone decides what the process should actually produce, at what standard, with which handoffs, checked by whom, against what definition of wrong. None of that is technical. All of it is senior. So the first question worth answering is where the effort of the people you already employ is about to move, because that answer determines whether the tool pays for itself or becomes another line item nobody will defend at renewal.

Run the sequence the other way and you get the outcome that has been showing up in post-mortems since 2025. Headcount comes out, volume holds up for a while because the tools really do produce volume, and the quality slides slowly enough that no single person is ever accountable for noticing.

The sequencing matters more at your size than it does at enterprise scale, and for an unfair reason. A company with four thousand people can absorb a bad cut in one function and carry on. At eighty people, the person you let go was the only one who knew how the renewals process actually worked, and the documentation they left behind describes a version of it from 2023. The redesign work still has to happen. It now has to happen without the one person who could have done it quickly.

The work relocates upstream

Watch a team put a model into a real workflow and the same two things happen every time.

The producing work compresses hard. First drafts, first passes, the raw volume that used to eat a week now takes an afternoon. That compression is real, and it is almost always the part that gets measured, because it is the only part with a before-and-after number attached.

The defining work expands, and it expands quietly. Someone has to say what good looks like in enough detail that a machine can hit it, which turns out to be much harder than saying it to a person who already has taste. Someone has to build and maintain the example set, because a curated set of your own shipped work teaches a model your voice far better than any description of it. Someone has to draw the constraints the system runs inside. And someone has to keep watching the output, because an unattended workflow drifts toward the middle whether or not anybody is looking, and it drifts without a warning light.

Add those up and you have a job. It did not appear on an org chart three years ago. It exists now in every function that has put AI into production, and in most growing businesses it has been handed to nobody in particular, which means it is being done badly by everybody in rotation.

This is the shape of it: the work does not shrink, it relocates upstream. A junior used to spend Tuesday producing. Now a senior spends Tuesday specifying. The hours did not leave the business. They moved to a more expensive person, doing a harder thing, with less to show for it on any dashboard you currently run.

That last part is what makes this hard to see from the budget meeting. Specifying work produces no artifact. There is no deliverable, no ticket closed, no asset shipped. There is only output that came out right, which looks exactly like output that came out right by luck.

Who you actually need

Two wrong turns show up at this point, and both are expensive.

The first is hiring an AI person. A business doing $5M to $100M with thirty to four hundred people cannot justify a dedicated AI hire, cannot usually attract one, and would struggle to keep one busy on work that matters. This is the shape of the bind most of our clients arrive in: nobody in house really knows this stuff, there is no budget for someone who does, and the leaders who would have to learn it are the ones with the least spare time in the building. Hiring a specialist looks like the way out of that. It mostly relocates the problem, because the specialist knows models and your team knows the work, and the value sits in the join between them.

The second wrong turn is worse, and it wears the costume of good financial discipline. You look at a function where output has doubled, notice that the expensive people there are no longer visibly producing, and take the saving. Those expensive people were the standard. They are the reason the doubled output is still recognizably yours. Remove them and you keep the volume and lose the thing that made the volume worth having, and the loss will not show up in any number you currently track until a customer tells you.

What the work actually asks for is narrower and more ordinary than an AI hire. The person needs real command of a function and the ability to put what good looks like into words that someone else can apply, which is a much rarer skill than having the taste in the first place. Plenty of excellent marketers cannot explain why a piece of copy is wrong. They can only tell you that it is. That gap used to be tolerable, because the work went to another human who had absorbed the standard by sitting near them for two years. A model has not sat near anyone.

The second half of the job is the patience to read output that is nearly right and keep noticing the one in eight that is confidently, plausibly wrong. That is duller than it sounds and most people are bad at it after the first month.

And the standard needs consequences attached, which in practice means this person owns a number that moves when the work is good and sags when it is not. Without that, the specifying becomes an opinion that other people route around when they are busy.

Taste is the scarce input now. Tool operation is not, and it gets less scarce every quarter. Hire accordingly.

Audit your next three hires

Take whatever is open or planned right now and run two passes over it.

First pass, per role: does this person add production capacity that the tools already cover cheaply, or judgment that you will need more of as volume climbs? A role that exists to produce more of something is worth re-specifying before you post it. A role that exists to decide what good is, and to be answerable for it, is worth more than it was two years ago and should probably be paid accordingly.

Second pass, per function: name the individual who owns what good looks like here. Not the team. The person. Marketing, sales, ops, finance, support, whichever functions have live AI workflows. If a name comes back immediately for each one, you are in better shape than most businesses your size. If you get a pause, or a tool name where a person should be, you have found the gap, and hiring will not close it. Somebody already in the room needs the job, with the time to do it written into what else they are no longer expected to deliver.

Worth doing the same pass over the roles you have already cut or quietly left unfilled this year. If the reasoning at the time was that the tools now cover it, check what has happened to the output since. The honest version of this audit turns up at least one place where the saving was real and one where it was borrowed against quality that has not been paid back yet.

Growing businesses build this judgment bench on purpose or they find out it is missing by reading their own output six months later and not recognizing it. The shift is happening in your company either way, at whatever pace your team adopts these tools, with or without a plan behind it. The only real decision is whether somebody in the building is watching it happen.

Sources

1. PwC, 2026 AI Business Predictions: https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html
2. McKinsey & Company, The State of AI: Global Survey (2026 edition): https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai