Chapter 19
Hiring the Fox
The Utopia of the Agentic Enterprise, Part IV. Doing Less
In December 2024 Australia’s Department of Employment and Workplace Relations commissioned Deloitte, for about A$440,000, to review the Targeted Compliance Framework, the IT system that docks welfare payments automatically when jobseekers miss their obligations. The report found that the system’s rules couldn’t be traced back to the legislation, and that the system was “driven by punitive assumptions of participant non-compliance”. It was published in July 2025. Christopher Rudge, an academic at the University of Sydney, then noticed that some of its references didn’t exist, among them reports by professors at his own university and at Lund, and that a reference to a court decision in a robodebt case had been made up. The revised report disclosed in an appendix that part of the work had used a GPT-4o tool chain, licensed by the department and hosted on its own cloud. Deloitte agreed to repay the final instalment of the contract and said the corrections “in no way impact or affect the substantive content, findings and recommendations”.1
Rudge hesitated to call the report illegitimate, because its conclusions agreed with other evidence. But in the revised version each invented reference had been replaced by five or more real ones, which suggested to him that the claims in the body of the report hadn’t rested on any particular source. The footnotes were there to show that somebody independent had checked. Graeber would have called them a box ticker’s work, a record that lets an organisation say it has done something, and a model made them cheap to produce. The person who did check them wasn’t part of the contract.
The department wasn’t unusual in buying its expertise from outside. In the year to June 2022, 112 Australian government agencies paid about A$21 billion for almost 54,000 full-time staff who weren’t public servants.2 Senator Deborah O’Neill’s advice after the Deloitte refund was that “anyone looking to contract these firms should be asking exactly who is doing the work they are paying for”.3
In IT services we tell a story about why clients hire us. Any new technology takes a while to produce enough experts for a company to staff it in-house, so the company hires a consultancy or a systems integrator to do the work. And once the system is built, there isn’t enough interesting work left to keep the experts entertained, so if the company had hired them they would have left it with a system it doesn’t understand. That’s the story, anyway.
As an industry we’ve put a lot of effort into keeping up the mystique of what we do. Most of it isn’t that complicated, and when it is, we don’t know either and look it up. That doesn’t bother the salespeople, who sell the scarce expert regardless. Graeber found that people in service jobs seldom thought their whole industry pointless, and IT providers were one of the exceptions. Many of them, he wrote, “were convinced they were basically engaged in scams”.4
Economists have a name for what the client buys. Uwe Dulleck and Rudolf Kerschbamer’s survey of credence goods is titled “On Doctors, Mechanics, and Computer Specialists”. In such a market the seller diagnoses what you need and then sells it to you, and you can’t tell afterwards whether you needed it.5 Hiring the fox means asking the seller what you need.
Then the contract changes the market. Oliver Williamson called it the fundamental transformation: before the contract many suppliers bid, and after it the winner knows things about the client that no rival does, so a contest among many becomes a relationship between two.6
If you work in this industry, you find quickly that the support is skewed towards a long-term dependency on a single person, even when the integrator says there’s a big team behind them and any individual could be replaced at the snap of a finger. The contracts are written that way, and it gives everyone peace of mind. The client can see the arrangement from its side. When the system misbehaves, everybody calls the same person, and that person works for the vendor. Changing a threshold or a line in a prompt takes a change request, with a price and a lead time.
The client pays twice, in the rate and in what its own people never learn. Mariana Mazzucato and Rosie Collington built The Big Con, their book on the consulting industry, around the question “What happens to the brain of an organisation when it is not learning by doing because someone else is doing the doing?” Their answer was that “the more governments and businesses outsource, the less they know how to do”.7
The experts play along. If you run a manufacturing company on tight margins, you compete for them with consultancies that offer six weeks of paid holiday and a retreat by the beach to learn about technology. You could pay the salaries that buy the beach, and easily, since you’re already paying the consultancy and it has to take the money from somewhere. What you can’t do is explain to a two-shift assembly line why the people in IT get more holiday. So you don’t compete, and the day rate pays for the beach.
It has distorted the market. Consultancies and integrators have held on to IT people with no projects in sight, which dried the market up and made the scarcity something to sell. In June 2024 an Indian IT employees’ union counted more than 10,000 graduates who held offers from five IT services companies and were still waiting for a start date. Over 2,000 of them had been recruited by Infosys and had waited more than two years.8 Those offers were made before there was work for the graduates to start on. None of this needs a conspiracy. Each firm keeping its bench full and its contracts tight is enough, and the client can’t hire what the market no longer has.
Firms do sell something a client can’t get elsewhere at the start. They have people who’ve done the work before the labour market has them, and they’ve seen what went wrong at other clients. A fixed price moves the risk of a late delivery from the client to the firm, which is worth paying for. The real work inside an engagement is bringing a capability in and leaving it behind.
Robert Townsend, who ran Avis in the 1960s, wrote of consultants: “They are people who borrow your watch to tell you what time it is and then walk off with it.”9 The integrators kept the watch and had a few juniors tell you the time, by the hour. An agent can now tell the time for next to nothing, and an industry paid by the hour can’t credibly sell that, because if it did, collectively, its revenue would go and its business model with it.
Suppliers who sell services by the head face the same squeeze as their clients, because AI replaces the labour they sell. So the supplier is supposed to enable the client’s AI transformation while defending revenue that depends on the inefficiency the transformation is meant to remove. Willingly or not, the industry’s answers each fit one of Graeber’s types.
The first is a new label. An organisation that can’t staff its AI work internally may announce a centre of competence and fill it through a consultancy, and the consultancy may not have changed how it delivers anything, only relabelled the people it already had, which is a lot faster than retraining them. The by-the-hour juniors become AI experts. In late 2025 Accenture began calling its nearly 800,000 employees “reinventors”, and tested a version of its HR portal that listed them as reinventors rather than workers.10 Early the next year it made regular use of its AI tools a condition for senior promotions.11 For the sponsor who bought the centre of competence, the new label does a flunky’s work in Graeber’s sense, making the sponsor feel the transformation is staffed.
The second is governance. An integrator can set up an AI control board, staffed with its own people, and the client’s staff can’t change a prompt or a threshold without its sign-off. Advisory firms profit from complexity, and AI governance is complexity that can be sold before the client has any value to govern. It becomes a service line that produces more governance, a box ticker’s file that lets the client say its AI is under control. The template is older than AI, as old as the ISO 9001 manuals consultancies wrote in the 1980s. Deloitte’s footnotes were the same kind of box, and a model filled it.
The third is security. A security consultant can find reasons why staff shouldn’t use a chatbot, and some of the reasons hold, since client data pasted into a consumer tool leaves the company. But a ban on the cheap tool also protects the integrator from its cheapest competitor, and the same market will sell the client a secured version, delivered by the hour. Graeber’s goons further the interests of whoever employs them. This one is paid by the client and furthers the integrator’s.
A proposal can be read the way an agent’s log can, by sorting what it delivers by who consumes it.
| The deliverable | goes to | Graeber’s type |
|---|---|---|
| the monthly steering pack | the sponsor, as reassurance | flunky |
| the vendor’s governance board | your vendor-management office, which it faces | goon |
| hypercare, or a managed service | a fault the project left behind | duct taper |
| the AI framework or control matrix | a file, as evidence | box ticker |
| the workshops and status cadence | your staff, as tasks | taskmaster |
| the change the project was for | someone it changes something for | real work |
Put the day rates beside each row, and add up the last row.
Then, for each place the proposal uses AI, ask whose hours it cuts. If the vendor’s agent drafts the status reports and the client’s people still review them, the vendor’s margin went up and the client’s work didn’t go down. Where the proposal says AI makes something faster, ask for a fixed price for it. A vendor that believes its claim can afford one.
An RFP has the same problem as the cover letter. A long answer used to take a bid team weeks, which told the buyer something about the bidder. Now both sides’ agents write, and the proposals read alike. So the RFP has to ask for what’s still expensive to fake: a pilot on the client’s own data with the client’s people at the keyboard, and the names of the people who’ll do the work, with a clause that keeps them on it.
Vendors also offer to optimise what they already run, with their own agents reading the client’s systems. Let the agents read, and ask that each recommendation show which of the vendor’s own revenue lines it touches. Someone at the client decides. The test of stopping a piece of work for a month to see who notices works on a vendor’s reports and services too.
Mazzucato’s own proposal was to write “learning by doing” clauses into consulting contracts, so that the client’s staff pick up the skills their departments have unlearned.12 A contract can go further and pay the vendor for leaving. Knowledge transfer is signed off by the people who’ll run the system, not by the vendor’s project manager. The no-hire clause goes, or the vendor puts a price on each person the client wants to keep. Build–operate–transfer contracts, in which a provider sets up an operation, runs it and then hands it over, already exist for offshore centres, so the lawyers on both sides know the form. And before the final payment, the vendor steps away for a month and the client’s people run the system on their own. Whoever they still have to call by the end of the month is the knowledge transfer that hasn’t happened yet, and the final payment waits for it.
Checklist
- Sort each deliverable in the proposal by who consumes it, with the day rate beside it. What share of the fee pays for the change the project is for?
- For each place the vendor uses AI, whose hours does it cut, yours or the vendor’s?
- Where the vendor says AI makes something faster, will it fix the price for that part?
- Which person at the vendor does everyone call when the system misbehaves, and what happens if they leave?
- Can your own staff change a prompt or a threshold without a change request?
- Does the contract stop you hiring the vendor’s people, and at what price would it let you?
- Before the final payment, can your people run the system for a month without the vendor?
- In a vendor-led optimisation, who at your company decides, and does each recommendation show which of the vendor’s revenue lines it touches?
- What does your RFP ask for that’s still expensive to fake?