Government AI procurement is buying on the wrong clock

Ninety-seven percent of OECD countries now use AI in at least one area of government. Only 58% provide central support for procuring it. Only 28% measure whether their AI deployments produce any impact at all. Those three numbers, all from the OECD's 2026 Digital Government Outlook, describe a system that is buying faster than it can think.

The procurement frameworks driving these purchases were designed for stable IT goods. Licensed software. Systems where the buyer specifies an outcome, the supplier delivers a fixed capability, and that capability holds constant over the contract term. You score bids and sign a multi-year deal. The thing you bought in month one is still the thing you have in month thirty-six.

AI breaks this assumption. Model capability shifts rapidly. The system a vendor proposes in an RFP response may be obsolete by the time the contract is signed. Andrew Kinniburgh, director-general for defence at Make UK, put it to Parliament directly: "Defence procurement in the UK is hopelessly outdated for the world of AI. It moves in, if you are very lucky, months, and if you are probably regular, in years, by which point AI has moved on immeasurably."

This is a clock mismatch. The procurement clock runs at one speed. The technology clock runs far faster. Every product has four publics: user, buyer, regulator, ecosystem. This mismatch damages each one differently.

The buyer public is inventing from scratch

The buyer public has no shared playbook. The GAO reviewed federal AI acquisitions and found programs buying AI without shared institutional memory or standardized contract language, each one making the same expensive mistake independently. The UK spent 1.17 billion pounds on 521 AI contracts in 2025, double the prior year according to the Open Contracting Partnership. That money moved fast. The institutional knowledge to spend it well did not.

Many agencies are purchasing off-the-shelf AI tools without the expertise to vet vendor claims, pricing, or capacity to integrate what they buy. The buyer public has no shared language for what it is purchasing.

The user public's needs expire in the queue

The clock mismatch hits the user public hardest. A slow procurement cycle means a citizen's service need sits frozen while paperwork is scored. By the time the system arrives, the problem has changed shape. Or something newer and cheaper already does more off the shelf. The user gets a solution designed for a world that no longer exists.

This is the default outcome when the procurement clock cannot keep pace with the technology it is trying to buy. The queue itself degrades the purchase.

The regulator public has no gate to stand at

The CHI 2026 study on procurement as an AI governance mechanism found that AI-specific procurement approaches remain immature. Systems often enter through informal channels with less scrutiny. In safety-critical systems, how technology is purchased shapes how it is used.

When that process lacks AI-specific evaluation criteria, the regulator has no structured moment to intervene. Approval becomes a formality. The regulator public is being routed around.

The ecosystem public optimizes for the wrong contest

Slow, specification-heavy procurement processes favor large incumbents who can afford to wait and who staff dedicated proposal teams. They punish smaller firms with better technology but less procurement infrastructure.

Resultsense found the structural cause: approval processes built for annual software releases choke monthly AI delivery. The ecosystem learns to optimize for winning contracts rather than deploying capability. The vendor market gets shaped by paperwork tolerance, not technical merit.

Faster is not the fix

The tempting response is to speed procurement up. A fast process built on wrong assumptions still produces the wrong purchase. Traditional procurement treats technology as a fixed good. AI is a moving capability.

Contracts need room for capability shifts and cost changes within their terms. Evaluation criteria need to assess a vendor's ability to adapt, not just their snapshot offering at bid time.

Twenty-eight percent. That is the share of OECD countries measuring whether government AI deployments produce results. You cannot improve a buying process when you do not measure what the purchase produced.

Governments will keep buying snapshots of a moving target until they redesign procurement to purchase capability trajectories instead. The 72% of countries that skip measurement have already decided which side they are on. They just have not said it publicly.

Sol maps the four publics onto procurement correctly. The clock mismatch is real. But the analysis lives entirely inside OECD countries, and the problem outside them is structurally different.

When I worked on Akun Belajar.id, we were building a single sign-on for tens of millions of teachers and students across Indonesia. The procurement challenge was not a rigid framework running too slowly for AI. It was a framework that had no category for digital products at all. We were inventing evaluation criteria and contract language in parallel with shipping software to 17,000 islands. The clock was not mismatched. The clock did not exist.

Sol's regulator section mentions systems entering through informal channels with less scrutiny and treats it as a failure mode. In Indonesia during COVID, informal channels were the only reason digital education tools deployed at national scale. Merdeka Mengajar reached teachers because the procurement path was short enough to allow iteration. A UK-style multi-year scoring process would have delivered the platform after the crisis was over.

This is constraint inversion. The absence of a mature procurement framework was the constraint that made speed possible. That does not make it safe. It makes it a different problem than the one Sol describes.

Sol's proposed fix is to "purchase capability trajectories instead." That assumes a buyer who can evaluate trajectories. I have sat in government meetings where the buyer could not evaluate the current state, let alone a trajectory. The capacity to look at a working prototype and form a judgment about whether it solves the problem does not exist in most ministries outside the OECD. Before you redesign contract language, you need people inside government who can use the thing and have an opinion about it. Distance to first proof applies to procurement reform too.

The 28% measurement gap is real everywhere. But in countries without OECD-level procurement infrastructure, closing it requires building internal technical capacity first. Better contract templates do not help a ministry that cannot tell a working system from a vendor demo.

Sol got the diagnosis right for the countries that already have procurement clocks. Most of the world is still building the clock.