Why doesn't giving everyone an AI license improve company productivity?

Because individual speed is rarely the binding constraint. Most work inside a company waits on someone else: a call to be scheduled, numbers to be collected, an approval to be granted. An AI tool makes a person reach the next dependency faster, then they wait in the same queue as before. Output at the company level only changes when the waiting gets shorter.

What did Meta's internal AI data actually show?

According to Reuters reporting on Meta's Project OT restructuring, code changes to internal platforms rose 220% year on year while new or upgraded features reaching users rose only 36%. Major technical and security incidents climbed 40%, and time spent responding to them rose 70%. Individual output rose sharply; the throughput of the organization did not follow.

Who captures the productivity gain from an AI rollout?

The individual does, first and by default. If an employee reveals an efficiency and the reward is more work, keeping part of the gain is a rational response. Unless a company explicitly renegotiates who benefits when a job gets easier, a meaningful share of the saved time stays with the person who saved it.

How do you find the real bottleneck in a business?

By asking people at every level of the organization what slows them down, then testing which answer carries the most cost. Executives typically name visibility, because that is what they lack. Frontline staff name specific mechanical failures: a form that will not submit, an approval that waits on one person, a report that has to be rekeyed by hand. The answers differ by role, and the expensive one is rarely visible from the top.

Should a value creation plan assume an efficiency gain from AI licenses?

Not on licenses alone. A plan assuming a 20% efficiency gain because everyone now has access has underwritten the gain at the employee level, which the company does not automatically capture. Defensible assumptions attach to specific changes: an approval step removed, a handoff collapsed, a reporting cycle shortened from weeks to days.

Why AI Rollouts Don't Raise Company Productivity

Give everyone in a company an AI license and individual output rises within days. Company output usually does not move at all. The reason is not adoption, training, or model quality. It is that the speed of any single person was almost never the thing holding the work up.

This matters most for anyone underwriting an operating improvement. A value creation plan that assumes a 20% efficiency gain because a business now has licenses has underwritten a gain at the employee level, and the company does not automatically capture it.

The gain lands on the person first

A senior lawyer at a firm that had rolled out AI to everyone described the arithmetic plainly: roughly half the productivity gain he was getting, he kept for himself. Maybe a little more.

He was not being obstructive. He had spent years asking for more support, and what arrived was software rather than a person. Nobody at the firm ever discussed who the saved time belonged to, so he decided.

This is a predictable response rather than a cultural failure. If every efficiency an employee reveals is converted into additional work, retaining part of the gain is rational. Some of the saved time improves the quality of the output. Some becomes an earlier evening.

The same logic explains why saved minutes do not aggregate into capacity. Asked whether the tool meant he could take on another client, he said no. Another client means another set of facts to hold, another relationship to manage, and one more thing that can become urgent on a Friday. Judgment, responsibility and context switching set his ceiling. The drafting never did.

Meta measured the gap

Meta ran the largest version of this experiment and documented the result internally. Reuters reported the numbers in August 2026, drawn from a restructuring program called Project OT.

Code changes to Meta’s internal software platforms rose 220% year on year. New or upgraded features that actually reached a user rose 36%. Major technical and security incidents climbed 40%, and the time employees spent responding to them rose 70%.

This is the most aggressive adopter among the large caps, with its own frontier lab and no budget constraint. The individual effect worked exactly as promised. The output then arrived at review, integration and on-call, none of which got faster, and a significant portion returned as rework.

The constraint had moved. It was no longer writing the code. It was deciding what should ship, integrating it safely, and handling what broke.

Companies run at the speed of the queue

Most work inside an organization involves waiting on someone else. Getting four people onto a call. Collecting numbers from a team that owns a different system. Waiting for a decision that needs a person who is traveling this week.

An AI tool can prepare the meeting and draft the follow-up. It cannot make Thursday’s call happen on Tuesday.

A seat-based rollout treats a company as a collection of individual jobs. A company is a system of dependencies and decisions. Everyone reaches the next dependency faster, and then waits exactly as long as before.

This is why small teams report results that large organizations cannot reproduce. In a team of three, the individual genuinely is the bottleneck, so removing the constraint on the person removes it on the business. Nothing sits between deciding and doing.

Five people, five different bottlenecks

The practical work is identifying where the waiting actually costs money, and the answers vary sharply by role.

Executives usually name visibility, because visibility is what executives lack. Ask a field engineer and the answer is mechanical: the job form takes twenty minutes and will not submit from a phone in a basement, so nothing gets invoiced until he returns to the depot on Friday. Sales leadership names lead volume; the representative says every quote above a certain size waits on one person who is in meetings all day. A CFO names month-end close; the controller says two subsidiaries send numbers in different formats and someone rekeys them by hand.

None of those are fixed by a license. Only one of them was visible from the top.

Across our implementations at the operating company level, this diagnostic pass is most of the early work, and it is far less technical than people expect. It is largely a matter of sitting with people at every level and testing which of their answers carries real cost.

In one PE-backed services business, problems surfaced in the monthly review rather than while there was still time to act. A margin slip or a billing error appeared weeks after close, by which point it was a customer dispute. Staffing was set by last period’s volume rather than next period’s demand. Reconciling the operating, billing and accounting systems into one foundation, then putting predictive models on top, moved the detection point forward rather than making anyone type faster. Incidents fell around 30%.

Shorten the chains of translation

What to look for is the chain of translation: every point where information changes hands, format, or system, and loses a day and some of its meaning along the way.

A field engineer’s notes become a form. The form becomes a spreadsheet. The spreadsheet becomes a line in a board pack. By the time anyone can act, the month has closed and the number is history.

Portfolio reporting is the clearest case. In a mid-market PE firm’s monthly cycle, P&Ls arrived from portfolio companies in every conceivable format and were reconciled by hand into one standard. Automating that reconciliation returned roughly 40 hours a month, not because anyone worked faster, but because a translation step stopped existing.

Shorten those chains and the business speeds up, because the time comes out of the gaps rather than out of anyone’s typing.

What actually reaches the P&L

Giving a workforce access to AI remains sensible. It makes tedious work less tedious, gives unsupported people something resembling junior help, and raises the baseline quality of everyday output. Those are real, and they are worth having.

They are not a company-level productivity gain, and a plan should not be built as though they are.

The changes that reach the P&L are structural: an approval removed, a handoff between two teams collapsed, one team’s data made reachable by another without a request and a two-week wait. Each is organizational work rather than procurement, which is precisely why it gets deferred. Buying licenses is a purchase order and one meeting. Establishing that sourcing data and portfolio operating data have never lived in the same place, and then putting them there, is a project with a sponsor and an argument attached.

That is also the difference between a rollout everyone describes as a success and one that shows up in a number somebody underwrote. For more on why the early progress feels real while the ceiling stays where it was, see why becoming AI native is harder than it looks.