AI won’t take your job. It will take your queue.

Last month I built an IT department. Eighty people across twenty-four roles — architects, analysts, engineers, testers, a service desk, a senior management layer, a change advisory board with strong opinions about Thursdays — carrying roughly £7.4 million a year in payroll. Then I let an AI into it, one role at a time, and watched what happened to the machine.

Almost every conversation about AI adoption is a conversation about individuals. Which tasks, whose job, what percentage of a developer’s day. The vendors reassure people, the commentators coach people, the surveys poll people. It is all very human-scale, and all perhaps even a little comforting because it implies the disruption will arrive one desk at a time and a human pace.

I think that framing is badly wrong.

An organisation is not just a collection of assorted jobs. It is a network, a literal machine made of queues. Work arrives, gets shaped, waits, gets built, waits, gets reviewed, waits, and, eventually, if it’s lucky, limps out the far end as value. AI does not subtract workers from that machine one at a time. It re-plumbs the machine. The unit of disruption here is not the job description; it is the flow. And flows do not fail politely, one desk at a time. Any changes to them impact all points quickly as bottlenecks are removed and new chokepoints emerge.

So rather than write another opinion about this, I built the machine and put it where you can pull its levers.

A machine made of diaries

The tool is a live simulation of an IT department as a flow network. Eleven functions, twenty-four roles — senior management included, because it turns out they matter — each with a headcount and a UK salary range (editable, right down to a CSV you can maintain in a spreadsheet, because you will not necessarily agree with my numbers and nor should you). Work moves between specific roles as artefacts: an initiative brief becomes a requirements pack, becomes a solution design, becomes a security review, a work breakdown, a build, a code review, a test cycle, a change approval, a deployment. Incidents pour in at the service desk and route out to whoever fixes things. Status reports rain gently and perpetually on the PMO. Perhaps that last fact might diminish my model’s credibility, but I’ll stick with it!

Two numbers travel with every artefact. The first is effort — honest person-days of work. The second is coordination — the days that exist purely to get the right people in the same meeting: the design authority slot, the CAB calendar, the workshop that couldn’t land until the week after next.

Underneath, the whole thing runs on queueing mathematics; my favourite course from my degree at Oxford, and one I should therefore make much more effort to use more! Every utilisation, every bottleneck, all governed from the settings. The animation is just that equation, wearing its best clothes.

Here is what the default, perfectly ordinary, medium-sized department looks like before AI touches anything: a project takes about thirty weeks door to door, of which barely ten are hands-on work. The rest is waiting: thirteen weeks in queues behind an overloaded engineering team; seven more in diaries, waiting for the meeting about the work to be scheduled so the work can be allowed to start.

Two-thirds of your lead time is nobody doing anything.

The Constraint Migration

Then you press the button, and this is where it stops being about individuals.

The AI absorbs the most overloaded role — at default settings, the software engineers, running at 109% of capacity with a backlog quietly strangling everything downstream. Two things happen immediately. First, the backlog stops being a tragedy: an AI doing tasks for you doesn’t just work on them quicker, it works on them in parallel, destroying backlogs in a way unfeasible with mere staff augmentation, so ninety-odd stalled builds clear in an afternoon for about £1,100 — work the humans needed £171,000 and another seven weeks to grind through. A backlog stops being a management failure requiring intervention and oversight and just becomes a purchase decision to buy the problem away.

Second — and this is the bit I want you to sit with — the pressure does not disappear. It moves. The hot spot migrates to the solution architects; a wave later it lands on the PMO, whose analysts have been quietly running at 102% all along — because in this organisation governance genuinely gates delivery: every initiative needs their funding sign-off and release assurance, and every doomed project needs their recommendation before senior management can kill it. Absorb them and the whole machine’s diaries loosen at once. Every wave, the constraint packs its bags and turns up somewhere that felt safe the week before. I’ve started calling this the Constraint Migration, because “transformation roadmap” doesn’t capture the feeling of watching your queue arrive at your desk.

There is a second story running up the governance side of the chart while this happens. Twelve per cent of everything the business asks for is doomed — it just doesn’t know which twelve per cent. With the PMO drowning, management sight runs at 57%, so the kill chain — flagged at design review, recommended by the PMO, cancelled by senior management — catches barely half the doomed work early; the rest sails through build, test and deployment and ships as waste, indistinguishable from value in a throughput report. And when the machine falls behind, the tool offers you the old playbook before it offers you the button: hire the gap it computes for you — three and a half engineers and a PMO analyst, recruitment fees booked, payroll recurring annually — or raise the standing kill rate and pay in cancelled ambition. AI is the third option. Arranging for you to feel that, rather than be told it, is most of what the tool is for.

Run it at different scales and the first casualty changes, which is exactly what a network effect looks like. In the fifteen-person shop, the constraint is the one infrastructure engineer everything eventually lands on. In the eighty-person department, it’s engineering. In the three-hundred-person enterprise, the first thing the AI should eat is not code at all — it’s the change-approval board and the design authority, drowning at 140%-plus while everyone argues about developer productivity.

And the migration runs further than I first let it. Product managers go at wave seven — drafting an initiative is knowledge work like any other. Senior managers go at wave nineteen, as a pure cost play, because by then the machine’s decisions mostly make themselves. What the model refuses to automate is smaller and harder than “management”: on the default numbers, the end state runs on six people out of eighty — two directors, one owning demand and one owning accountability, kept human because the law requires someone to carry the outcome and the machine cannot want — plus four stewards, one for each collapsed cluster of functions. Throughput holds. Lead time falls from thirty weeks to about one. And the tool dates every step: the full absorption takes a little over four years of simulated weeks, printed as a timeline from the day you let AI in — which team dissolved in which week, in what order.

The going rate for a build is £28

Let me give you the numbers the deck your boss is being shown will contain, because the tool prices everything.

A build, done by a mid-level engineer on a UK salary with employer on-costs, runs to roughly £4,500 and fourteen elapsed days once you count the queue and the diary alignment. The same artefact, executed by an AI agent with a human review step, costs about £28 and four hours. A solution design: £2,300 against £15. An incident fix: about £400 against £2.50. The whole department’s delivery engine, fully absorbed, costs around £370,000 a year in compute and platform — against £6.5 million of released payroll.

Then try the demand shocks, because this is where the asymmetry turns brutal. Double the demand on a human organisation and the model prices it honestly: dozens of hires, twenty per cent recruitment fees, twenty-two weeks to productivity, and a mountain of backlog accumulating while you interview. Double the demand on the AI-run version and the compute bill changes that afternoon. Halve demand and the comparison is worse: redundancy costs and consultation periods against simply turning the dial down.

Elastic beats loyal, on price and on speed. That sentence should make you deeply uncomfortable. It is also just arithmetic.

Three people this should bother

If you work inside one of these boxes. Somewhere right now, a supplier is presenting a version of this model to your boss — less honest than mine, prettier, with your function expressed as a row: a utilisation figure, a £-per-artefact comparison, a “transition timeline”. If you are not already using AI at this level — modelling your own work, automating your own queue, learning where your role sits in the flow — then you are not in that conversation. You are the subject of it. Be the person running the model, not a row in somebody else’s.

If you run one of the boxes. The head-count saving is a poisoned chalice, and I say that as someone showing you a tool that computes it to the pound. Every person you release from an AI-absorbed function walks out knowing three things: exactly how the work really gets done, exactly how much of it an AI can now carry, and exactly how little capital it takes to try. I have written before about the ex-colleague as competitor; this is that argument with a progress bar. They leave with your playbook, without your payroll, and — the part that should genuinely frighten you — without your change-approval board.

If you run the whole thing. Then understand that people like me — and there are versions of me less honest and considerably more mercenary — are no longer modelling departments. We are modelling entire businesses as machines: your onboarding, your underwriting, your claims handling, your compliance reviews, every artefact and every queue — and pricing how few people it takes to run a competitor that does only the parts your customers actually pay for. The boutique insurgency does not need your scale. Scale is the thing it is betting against.

None of the people in this story are villains, by the way. The vendor selling to your boss is doing their job. The boss eyeing the payroll line is answering to a margin that answers to a market. The mercenary building your replacement read the same arithmetic you’re reading now — just earlier, and without the comfort of assuming it applied to someone else. The machine doesn’t hate anyone. The machine doesn’t know you’re in it.

The honest part

The model is a toy. I want to be plain about that, because its honesty is the only thing separating me from the deck-wavers. Every service time, salary, defect rate and AI cost in it is an illustrative default, editable live and exportable to CSV, precisely because every one of them is arguable. It compresses years of organisational change into minutes. It assumes AI output quality holds with a human steward on review — an assumption I’ve made visible rather than buried, which is not the same as having proven it. It ignores politics, regulators, contracts, morale and all the other glorious friction that makes real transformation slower than mathematics.

And if your reaction to any number in this piece is “that’s nonsense, a build doesn’t cost £28” — good. Open the tool. Change it. Watch what your number does to the network. Being the person in the room who has argued with the parameters is rapidly becoming the whole job.

I could stop there and feel thoroughly pleased with myself. I won’t, because the boast needs deflating: I am not the impressive part of this story. An experienced practitioner with an AI built a working, priced, animated simulation of an entire department in a handful of evenings. Five years ago that was a consultancy engagement — a team of six and a quarter’s fees. Whatever this tells you about IT departments, it tells you rather more about what one person can now point at yours.

The queue is already moving

Play with it. Load your organisation’s real shape into the CSV, set the demand to your portfolio, and press the button — then watch where the pressure goes, because that is the conversation your leadership will be having within the year, with or without you in the room.

The machine is humming either way. The constraint is migrating either way. The only variable still open is which of the three people above you decide to be — the one who runs the model, the one who becomes a row in it, or the one who reads about it afterwards in a competitor’s case study.

Act. Live. Move.