The capacity plan counted licenses

Meta description: The budget bought access. A quarter later, a hidden queue of human review work appeared that nobody estimated. The budget bought access. The work bought a queue.

The capacity plan was approved without much debate. It counted what AI capacity plans count: licenses, seats, compute, vendor tiers — capacity as something purchased from suppliers. A quarter later, I am looking at a different reality: a queue of AI-generated work awaiting human review that nobody estimated, staffed by people whose calendars were already full, growing at the rate the models produce. The plan was not wrong about the licenses. It was silent about the humans, and the silence is now a queue, an error rate, and a team too proud of the program to say they are drowning.

[AWAITING MIKLOS — scene: a budget that bought access and a quarter later a hidden queue of human checks]

What do capacity plans currently count?

The plans I read count supply: how many seats, what tier of model, how much compute, what the contract allows. This is procurement arithmetic, and it answers the CFO’s question — what does AI cost? — with a vendor-shaped answer. What the arithmetic cannot see is demand on the other side: every generated draft requires reading, every suggestion requires judgment, every automated output requires someone accountable for accepting it. The model’s capacity is purchased. The human capacity to absorb what the model produces is assumed, and assumptions do not appear in budgets.

Where does uncounted work appear?

Miklos Roth, high-velocity AI consultant

[AWAITING MIKLOS: the capacities he has seen omitted.]

The uncounted work appears in the seams of people’s days. I have watched it accumulate: a reviewer who now checks forty machine-drafted documents where she once wrote ten, a manager whose team produces three times the analysis and whose sign-off time tripled with it. None of this was in anyone’s objectives. It arrived as an implicit duty — the model produced it, someone must vouch for it — and implicit duties expand until they meet a limit. The limit is usually quality: first the review gets faster, then thinner, then ceremonial. The queue and the corner-cutting are the same phenomenon at different stages.

Why do seat models repeat the error?

The false explanation is seats: buy more licenses, extend access, democratize. I have watched seat counts double while the review queue doubled faster, because each new seat is a new source of generated work and the reviewing capacity did not grow at all. Seat models assume the scarce resource is access to the tool. In the organizations I visit, access is abundant and the scarce resource is judgment — the humans who can responsibly accept the tool’s output. Judgment cannot be bought in seat increments, so the seat model cannot see the scarcity it is worsening.

What pattern marks a missing estimate?

The marker is where the truth lives. In companies with a missing estimate, the real capacity picture exists in private alone: a team lead who knows her people are saturated, a reviewer with a personal triage system, a quiet agreement to let some outputs through unchecked. The official picture — headcount, licenses, budget — is complete and false. I have learned that the distance between the official picture and the private knowledge is itself a measure: the wider the distance, the more certain the coming quality failure, and the more surprised leadership will be by it.

[AWAITING MIKLOS: the hidden-queue scene.]

[AWAITING MIKLOS — scene: second scene]

What question forces the hidden queue into the plan?

Budget processes ask what the company will buy. The hidden queue exists because nobody asked what the company must absorb — what human capacity the bought capacity demands. The question that fixes the plan is not a sizing formula; it is a question about the other side of the ledger, asked before the purchase, with the people who will staff the queue in the room. In my experience they are never in the room. The queue is their introduction.

[AWAITING MIKLOS: the question. Not a sizing formula — and why it is the hinge]

PR excerpt: [PR — awaiting Miklos material; cut 80–120 words from his paragraphs once available.]

By Miklos Roth — AI marketing consultant, Budapest. More: his field notes.