Six months ago you stood in front of the team, or you wrote the Slack post, and announced the company was “rolling out AI.” ChatGPT seats for everyone. Maybe Copilot too. A channel where people could share their best prompts. It felt like leadership. It felt like proof you weren’t the founder who slept through the thing everyone else was scared of missing.
Last week you opened a job req.
Nobody in the building connected those two moments. I want you to.
Open your P&L. Find the AI line item you added six months ago. It’s small, easy to miss between the other software subscriptions. Now look two lines down, at the new salary about to join it, for a role that does roughly the same work the memo was supposed to touch. You didn’t cancel the subscription. You’re not about to cancel the hire either. You’re about to pay for both.
Run the test yourself. Name the tool you rolled out. Now name the seat it replaced.
If the second answer comes slower than the first, or doesn’t come at all, you didn’t roll out AI. You bought a subscription and wrote a memo about it.
Is buying AI seats for your whole team a mistake?
No. Tool-level adoption, ChatGPT or Copilot for the team, is a reasonable first move for a company your size. The mistake is stopping there and calling it a strategy.
You don’t redesign a role around a piece of technology before you know what that technology can do to your numbers. That’s not cowardice. That’s how a company with a real P&L is supposed to behave. Restructure first and you’re the founder who tore up a job description because a demo looked impressive, then spent six weeks rebuilding it by hand while a client waited.
There’s also a real economic effect working against you here, and it has a name. When AI makes a task faster, most companies don’t shrink the team that does the task. They ask the same team to do more of it. Economists call this a version of Jevons paradox: efficiency gains get absorbed as higher output, not lower headcount. Your account managers start using Copilot to summarize calls and follow up faster. Instead of needing four people to carry the book, three can carry it. Except nobody told the three that’s what happened, so you still post the fourth req when the next batch of accounts lands. The org chart looks exactly the same shape it always did. Something real still happened underneath it, and you missed the chance to bank it.
So the excuses that feel true in the moment are, in fact, half true:
- “We’re not big enough to build something custom yet.” True if you mean building an agent from scratch. False the moment the seats already exist for you to build on top of.
- “The team’s already using it every day, isn’t that the point?” That’s individual speed. It’s not the same claim as “we need one fewer person on this team.”
- “I don’t have time to redesign how we work right now, I’ve got a business to run.” This is the real one. Redesigning a role needs the one person who can authorize it, and that person’s calendar is already full of fires that didn’t wait for AI to show up.
None of that makes you wrong to buy the seats. It makes the seats step one of two steps, and most companies stop after step one.
What does “did your org chart move” mean?
It doesn’t mean redrawing boxes on an org chart diagram. It means a hire you were about to make didn’t happen, because a seat now covers what that hire would have covered.
Go back to your account managers. If Copilot genuinely lets three people carry the load of four, ask the only question that matters: did that let you skip the fourth hire, the one already sitting in your headcount plan for next quarter? If yes, your org chart moved, even though nobody’s title changed and nobody got a new box on the diagram. You avoided a hire. That’s the whole test.
If the answer is no, if you still posted that req because nobody reassigned who owns what, then the tool made three people faster and the business still needed a fourth person anyway. You got a productivity bump and kept the growth ceiling. Both of those things can be true at once, and a lot of “we rolled out AI” memos are true in exactly that limited way.
So here’s the sharper version of the test, the one worth running instead of eyeballing an org chart: can you point to a specific hire that didn’t happen, and say why? Not “we’re piloting things.” Not “the team seems more efficient.” A name, a role, a job req that’s still sitting in a drawer instead of on the careers page.
The three-layer way to think about this
A job is one unit of work a tool can finish today. A seat is a stack of jobs large enough to cover what your next hire would have done. Buying tools gets you jobs. Almost nobody stacks them into a seat.
This is how I build for clients, and it’s the reason most rollout memos stall exactly where they do. A job is something concrete: summarize this transcript, draft this follow-up, chase this overdue invoice. A seat is a bundle of jobs big enough that you can point at it and say “that used to require a hire.” An org is what you get when enough seats are running that growth stops meaning headcount.
Almost every “we rolled out AI” memo stops at job one and never stacks toward a seat. That’s the tell. Ask what job the tool is doing, then ask if that job, plus the next one, plus the one after that, adds up to a role you no longer need to fill. If the honest answer is “it just makes people faster at the job they already have,” the org chart was never going to move, because nothing was ever built to replace the hire. Only to assist it.
| The move | What it looks like | What changed |
|---|---|---|
| Buy the seat (the memo) | Company-wide ChatGPT or Copilot licenses, a prompts channel, maybe one workshop | People get faster at the job they already have. No job gets fully covered. |
| Build the seat (the roadmap) | Specific jobs stacked one at a time, drafting the reply, summarizing the call, chasing the invoice, until they cover a full role | The next hire for that role doesn’t happen. The req stays in the drawer. |
The seats that hold up under this test aren’t theoretical. Knowledge management, so the company doesn’t lose its institutional memory every time someone leaves. Email and customer support, the seat that used to be a support hire. Project coordination, the seat that used to be an ops hire. Research, the analyst role most companies could never justify hiring full time. Each of those is a stack of jobs big enough to answer “which hire did this replace” with an actual name, not a shrug.
What does skipping the seat-building step cost?
More than the wasted subscription. It costs the hire you make anyway, the gains that never show up on the P&L, and, per Gartner, roughly a one in three chance the whole project gets abandoned before it produces anything.
Start with the abandonment rate. Gartner predicted in mid-2024 that at least 30 percent of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025, and named the reasons: poor data quality, inadequate risk controls, escalating costs, unclear business value. None of those reasons are “the model wasn’t smart enough.” They’re all versions of nobody deciding what job the tool was there to finish.
Then there’s the workload paradox. In a 2024 Upwork Research Institute survey of 2,500 professionals, 96 percent of C-suite leaders expected AI to boost productivity. 77 percent of the employees using the tools said it added to their workload instead, and 71 percent reported feeling burned out. Handing someone a faster way to do their job, without removing anything from that job, doesn’t buy you anything. It just builds a faster treadmill.
Boston Consulting Group studied more than 1,500 companies going through this and landed on what they call the 10-20-70 rule: the companies that get results put roughly 10 percent of the effort into the algorithms, 20 percent into the technology and data, and 70 percent into people and process. Most small companies buy the 30 percent, the tech, and skip the 70 percent, the part where somebody redesigns who does what. That’s not a rounding error. That’s most of the work, skipped.
There’s a reason skipping it doesn’t feel neutral, it actively costs you for a while. Economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson studied this pattern across general-purpose technologies going back to electricity and named it the productivity J-curve: measured output often dips before it rises, because the real payoff sits in investments nobody puts on a balance sheet, retraining, workflow redesign, deciding who’s responsible for what now. Skip that investment and you don’t get a smaller gain. You sit in the dip.
The dip has a price tag you can run yourself. A ChatGPT or Copilot seat costs a few hundred dollars a year per person. A fully loaded hire, once you add benefits, payroll tax, and ramp time, typically runs a quarter to a third more than the number on the offer letter. On an $80,000 role, that’s another $20,000 to $30,000 nobody put in the budget conversation. You can buy a lot of seats for the cost of one hire you didn’t need to make.
What should you do this week?
Pick the role you’re about to post a req for. Before you post it, list every job that role would do in a normal week, and check which of those jobs a tool could already do today. Stack the ones that qualify. If they cover most of the role, you just found a seat instead of a hire.
This isn’t a call to freeze hiring or distrust every tool your team already uses. It’s a call to run the audit before the req goes up, not after. Most founders run it backwards, if they run it at all: hire first, notice the AI tool sitting half-used in the corner second, wonder later why the subscription and the salary are both on the books.
If your AI strategy didn’t change your org chart, you don’t have one. You have a software bill. Not because the tools failed. Because nobody ever asked them to replace anything. They were only ever asked to help.
Frequently Asked Questions
Does rolling out ChatGPT or Copilot company-wide count as an AI strategy?
No, by itself it’s a purchase, not a strategy. A strategy names which specific jobs the tool is meant to fully cover and tracks whether a planned hire got avoided because of it. Company-wide seats without that tracking are adoption, not restructuring, and adoption alone rarely shows up on the P&L.
How do I know if AI replaced a role in my company or just made someone faster?
Ask whether a specific, planned hire didn’t happen because of it. If your team is faster but you still posted the req you were going to post anyway, AI assisted the job. If the req never got posted because the work is now covered, AI replaced it. Speed and replacement are different claims.
Why do most small business AI rollouts fail to change headcount plans?
Because they stop at buying the tool and skip redesigning who does what. Gartner found at least 30 percent of generative AI projects get abandoned after proof of concept, largely from unclear business value, and BCG’s research puts 70 percent of what drives results in the people and process work most companies never get to.
What’s the difference between a “job” and a “seat” in an AI org?
A job is one unit of work a tool can finish today, like drafting a follow-up email or summarizing a call. A seat is a stack of jobs large enough to cover what your next hire would have done. Buying tools gets you jobs. Only stacking them deliberately gets you a seat.
What should a founder do before posting their next job req?
List every task the new hire would do in a normal week, then check which of those tasks an AI tool could already handle today. If enough of them qualify, you may have found a seat you can build instead of a role you need to fill. That audit takes an afternoon, not a consultant.
If you want help running that audit before your next req goes up, that’s what the AI Org Roadmap is for: map your org, name the seats an AI team can fill, and get the build order. Start at davidchung.ai.