How to Procure an AI Ghost

Essay 1, The Unread Contract

Sometime last year, a state government published a tender. The document ran to dozens of pages and asked the market for something that resists easy description: a fully staffed, operationally ready Artificial Intelligence Division, to be housed inside the Chief Minister’s Secretariat, designed, built, and run by whoever won the contract.

I read it twice. The second time more slowly.

There was no cabinet resolution authorizing the unit. No government order establishing its legal mandate. No specification of what problem it was meant to solve, what decisions it would support, or how its outputs would connect to anything the state government actually did. There was a detailed vendor requirement list. Experience in machine learning, natural language processing, computer vision. A requirement for senior AI researchers. A reporting line to the Principal Secretary. Fifty pages describing an institution that did not legally exist. The winning vendor would be asked to conjure it into being, complete with staff, systems, and a mandate that no elected body had approved. A ghost institution, procured at public expense. The title of this essay is not a metaphor.

The political logic is not hard to read. An AI Division inside the Chief Minister’s office signals technological seriousness, generates press coverage, and requires neither a legislative fight nor a cabinet consensus. If the division produces something useful, the government takes credit. If it produces something harmful or embarrassing, the vendor absorbs the first wave of blame. That asymmetry is a feature of the procurement design, not an accident of it.

The consultant fingerprints were visible throughout. The language was fluent and current, full of references to AI maturity models and transformation roadmaps. This is what strategy firms now sell to governments: the vocabulary of institutional seriousness, priced separately from its substance. The government acquires a document that looks like a plan. The consultant acquires a contract. The question of what the AI Division is actually for gets deferred, because answering it honestly takes months and there is a political calendar that doesn’t wait.

That last detail matters. In any line department, a procurement this structurally incoherent would have encountered resistance. Finance departments ask for project justifications. Line ministries have enough operational reality to sense when a proposal connects to nothing real. This tender originated at the political center of government, which meant the usual institutional friction was unavailable. The procurement moved forward because the political calendar rewarded the announcement. The checks that might have caught the problem were not available to the people who might have applied them.

This case is one entry point into a wider pattern. Public procurement records from 2024 to 2026 show senior government offices across multiple jurisdictions issuing tenders for AI divisions, roadmaps, and analytics units with a consistency that suggests standard practice rather than aberration. The procurement documents share three features. The vendor is required to design, build, and run the unit, including in several cases providing its executive leadership. The reporting line goes directly to a senior official or political delivery unit, bypassing the line departments that would actually use the outputs. And the unit being created carries no prior statutory or legally authorized framework at the time the tender is published. The vendor is being hired to constitute a government function, not support one. Each of these procurements produces a variant of the same thing: a ghost with a different name and a current vocabulary.

A contract can smuggle a function. It cannot smuggle a legal mandate, an accountability chain, or data governance. Those have to exist inside the government before the vendor arrives. When they don’t, the vendor builds something, it operates for a few years, it generates data, it informs decisions that affect citizens, and at the end of the contract period there is no legal framework for what happened, no institutional memory, and no government official who was accountable for any of it.

The state has been here before, and so have cities across the world. A decade ago, governments were spending billions on Smart City programs, building Integrated Command and Control Centers in city after city: large buildings, large screens, real-time data feeds aggregating traffic, utilities, and public services into a single dashboard. One national program alone covered 100 cities and spent approximately US$18 billion. The government’s own review declared all 100 centers operational. Independent auditors examined how the broader program actually landed. They found that city agencies had quietly dropped between 7 and 48 percent of their approved projects. In one city, only 45 of 645 contracted projects were completed on time. In another, 29 of 44 approved projects had not started four years after the program launched. The program closed after ten years. The command center buildings remain. The screens display data. In most cities, whether departments changed routine decisions because of what those screens showed remains unanswered in the official record.

The AI Division tender is the same architecture. A new technology, an identical instinct: announce the infrastructure, defer the question of what it is actually for, and leave that answer to whoever wins the contract. Another ghost, newer vocabulary. Governments that have not resolved basic service delivery are procuring AI to manage the complaints those failures generate. The technology arrives as a shortcut. The shortcut leads nowhere, because the underlying institutional problem remains.

This would be easier to dismiss if the failure were confined to governments with weak technical capacity. Australia’s Robodebt program makes that argument unavailable.

Robodebt automated welfare debt recovery by cross-matching annual tax records with fortnightly benefit payments, then issuing debt notices to anyone whose data showed a discrepancy. The logic appeared obvious to the officials who approved it. A Royal Commission later established it was legally indefensible. Welfare recipients with variable incomes, casual workers, seasonal employees, the majority of the people the system targeted, have earnings that don’t average cleanly from annual tax data. The program produced more than 500,000 unlawful debt notices before courts intervened. The total resolution cost, including refunds, compensation, and the Royal Commission itself, came to approximately A$1.8 billion. The Commission documented the suppression of internal legal advice, the absence of validation testing before deployment, and the systematic failure to ask whether the algorithm’s core method produced correct results for the population it targeted.

Australia has independent courts, a functioning civil service, and a Royal Commission process capable of establishing precisely what went wrong. What it lacked, when Robodebt was approved, was a procurement process requiring anyone to answer a foundational question: does this method produce correct results for the people it will be applied to?

The two cases are different in scale, consequence, and institutional context. They share the same upstream failure. The institution could not define the problem clearly enough to specify a solution, and nobody required it to try. In one case a consultant filled the gap with a roadmap. In the other, an algorithm filled the gap with income averaging. Both produced ghosts: systems that existed in procurement documents and press releases, consuming public money, without the institutional foundation to do what they claimed.

AI procurement is the procurement of public judgment under uncertainty. That is what makes these failures so consistent across income levels, governance systems, and technology cycles. Governments arrive at vendor meetings carrying the wrong question. The question on the table is which system to buy. The question that should precede it has three parts.

What decision changes? Name the specific decision, made by which official, about which citizen or transaction, that the AI system will alter. A vague answer produces a vague specification, the vendor fills the gap with their own definition of success, and the evaluation measures what the vendor built rather than what the government needed. The ghost institution tender had no answer. Robodebt’s answer turned out to be wrong for most of the people it targeted, and it was never tested before deployment. Brazil’s federal government answered this question before its MentorIA system went live in 2026: the tool advises procurement officers on regulatory compliance, draws only from named legal documents, and requires a human to review every output before a decision is recorded. One named decision. One named official. One specified source. The vendor had nowhere to go except where the specification pointed.

How will you know if it’s wrong? Name the test, the data, the frequency, the unit inside government responsible for reviewing it. Income averaging fails this test in thirty seconds if applied honestly to casual workers. The Robodebt architects never applied it. Allegheny County in Pennsylvania issued a second procurement contract alongside its AI-assisted child welfare screening tool: an independent impact evaluation, awarded to Stanford University before the tool went live, with monthly quality assurance reports tracking performance from day one. When the evaluation found disparities in outcomes, the county updated the model. The tool has gone through four versions, each revision driven by monitoring findings. That is a failure signal functioning as designed, not discovered in court.

Who is accountable when it is wrong? Assign institutional and legal responsibility for an adverse outcome affecting a citizen before the contract is signed. In the ghost institution tender the effective answer was the winning vendor, which is another way of saying nobody with actual public authority. In Robodebt the Royal Commission reconstructed the accountability chain from suppressed legal advice and ministerial testimony. Singapore’s Government Technology Agency, a statutory board under the Ministry of Digital Development and Information, was built on the opposite principle. GovTech owns and operates the systems it builds for government. The vendor supplies components. GovTech holds the system. The Ministry holds the policy. When something goes wrong, the citizen has a named public institution with legal standing. That accountability chain was established by Parliament before any procurement was signed, not assembled afterward from testimony.

These three questions take twenty minutes to answer honestly. They require no technical expertise. They do require a civil servant who has been told it is their job to ask them, and a procurement process that will not advance without written answers on record.

The vendor will not raise these questions unprompted. The consultant who helped draft the terms of reference has a financial interest in a procurement that moves forward. The political principal who wants an announcement has a timeline that runs faster than the one required to answer these questions well.

These tenders are remarkable for exactly the wrong reason. They assume that a vendor and a model can solve what years of governance reform could not. That assumption is where public money goes to disappear.

If the AI division produces a decision that harms a citizen, which government official is answerable? If that question has no clear answer before the contract is signed, the institution hasn't been built. It's been rented. That is the ghost in every AI tender: not the technology, but the accountability that was never assigned.

All opinions are personal