MSc Thesis · Oxford Internet Institute · In progress · Submission Jun 2026

Procurement and Process:
what AI rules did — and what scrapping them didn't.

How a pair of 2024 White House memos — and their 2025 repeal — changed how federal agencies buy artificial intelligence, traced through the public contract record.

Even though the Biden-era AI rules set strict new requirements for how the public sector could use AI, federal agencies began signing AI contracts markedly faster once they took effect. When the Trump administration repealed those rules months later, AI contracting didn't suddenly expand — it showed no difference from federal tech contracting overall.

Putting rules in place moved the needle — scrapping them didn't move it back.

Method
Interrupted time series · OLS, Newey-West HAC & GLS AR(1) at equal standing
Source
USAspending.gov · Oct 2021 – Sept 2025 · 48 monthly obs (extended to Jan 2026 for the repeal test)
Treatments
M-24-10 (Mar 2024) · M-24-18 (Sept 2024) · 1 Dec 2024 deadline · M-25-21/22 repeal (Apr 2025)
n
4,616 civilian AI contract actions · 41 agencies · 788 recipients
Tooling
R · data.table · sandwich · nlme · ggplot2
01 — The question

Did AI guidance actually change what government buys?

In 2024, the White House told federal agencies how to buy and govern AI in two OMB memos — M-24-10 and M-24-18 — that set rules for acquiring it responsibly. In 2025, a new administration scrapped them and replaced them with lighter-touch guidance. This project asks a blunt question: did any of it change what agencies actually buy?

Three things, in plain terms. One — when the rules took effect, did agencies buy more AI, and was that rise really about AI rather than a general bump in tech spending? Two — when the rules were repealed, did AI buying rise relative to the rest of federal tech? Three — does the government's new near-free AI access program, OneGov, show up in the spending data at all? The answers, in order: yes, no, and barely.

AI contracts per month, after the rules
+5.9 /mo
the monthly pace kept climbing once the rules took effect p = .0004
Total civilian AI contracting, 2021–25
$2.9B
across 4,616 contracts and 41 federal agencies
Won by Palantir alone
$0.9B
the top AI vendor by far — ahead of Anduril and Booz Allen
AI contracting after the 2025 repeal
−7%/mo
the same rate as federal tech overall — no AI-specific drop p = .49
PRELIMINARY Federal civilian AI contracts and dollars, by month · Oct 2021 – Jan 2026
Period
Hover the chart
AI contract actions
AI obligations ($M)
AI contract actions (count)
AI obligations ($M)
Rules take effect (solid)
Other policy dates (dashed)
FINDING

The rules left a mark; the repeal didn't. Once the Biden-era requirements took effect, agencies signed AI contracts faster — gaining roughly six more a month, a result that holds across all three statistical models. And the increase looks specific to AI, not just part of the wider rise in government tech buying: AI's share of federal tech contracting climbed at the same moment (significant under the most conservative test). The memos' earlier announcement dates show weaker, less consistent effects. After the 2025 repeal, AI contracting fell about 7% a month — but federal tech contracting overall fell just as fast, so the drop was a government-wide pullback, not a reversal aimed at AI.

02 — Where the money goes

Who gets the money, and who's buying.

Top 10 vendors by total AI dollars

The biggest civilian AI contracts go less to traditional IT integrators than to AI-focused firms. Palantir leads by far (about $0.9B across two business units), with Anduril close behind — ahead of integrators like Booz Allen and Accenture, and a set of high-volume resellers (Four Points, ThunderCat) that handle many smaller buys.

Which agencies are buying (since the rules took effect)

By number of contracts. The General Services Administration is the single biggest buyer — but only about a fifth of contracts run through it. AI purchasing is spread widely across Health & Human Services, Homeland Security, Treasury, NASA and dozens more.
03 — The repeal

Repeal without reversal.

In April 2025 the Trump administration rescinded the Biden memos and replaced them with lighter-touch guidance, trading risk-management requirements for faster adoption and streamlined buying. That second turning point can be tested the same way — and it comes back empty.

AI contracting fell after the repeal by about 7% a month, but that was not an AI reversal. Federal tech contracting as a whole fell at the same rate — and AI's share of those contracts didn't budge. The buying didn't move elsewhere, either: AI bought through resellers fell too, rather than rising to soak up the lost contracts.

That asymmetry is the real finding. Rules that made AI easy to buy left a clear, AI-specific mark on the record; taking those rules away did not erase it. The most natural reading is that the rules built procurement capacity — agencies learned how to buy AI, set up the vendors and the paperwork — and that capacity didn't vanish when the rules did. Once AI became something agencies knew how to contract for, it stayed that way, whether or not the rules requiring it were still on the books. Regulation and deregulation aren't mirror images: it is easier to set federal buying in motion than to wind it back.

04 — The blind spot

The biggest move barely shows up.

One big piece of the story never reaches the data at all. GSA's OneGov arrangement, live from spring 2025, negotiates near-free, government-wide access to commercial AI tools like ChatGPT, Claude and Gemini by treating Washington as a single customer. By GSA's own count it has reached around 3.4 million users and avoided roughly $1.15 billion in cost. Yet in the contract record it is almost invisible: it shows up only as a handful of $1 subscription line items. The single largest expansion of federal AI access leaves almost no trace in the data meant to track federal buying.

It isn't the only blind spot. Thousands of contracts that agencies themselves flag as AI never say so in their text, so keyword searches miss them. And real AI work — from the likes of Palantir, Databricks and Clearview — often sits hidden as subcontracts beneath prime contracts that don't look like AI at all: at least $112 million of it.

The point is simple: what shows up in the contract record isn't the same as what the government actually buys. AI can enter an agency through a $1 subscription, a buried subcontract, or a contract description no keyword catches — and none of it looks like a big AI purchase. The record is a useful lens on federal AI, but a partial one.

05 — Method

Interrupted time series, civilian universe.

Monthly aggregation of civilian AI-related contract actions, identified through a layered filter combining NAICS coding and technical-vocabulary keywords, with supplementary capture for named AI vendors and products and a cross-walk to the OMB AI use-case inventory. Defense and intelligence agencies are excluded throughout — the memoranda govern civilian agencies, and defense contracting is a structurally distinct procurement regime, so it is dropped rather than used as a placebo.

Specification
ITS · level & slope shifts at each intervention
Outcome 1
AI contract action count — primary
Outcome 2
AI obligation $M — secondary corroboration (counts ↔ dollars r = 0.86)
Treatments
M-24-10 (Mar 2024); M-24-18 (Sep 2024); 1 Dec 2024 compliance deadline (headline break); M-25-21/22 repeal (Apr 2025)
Inference
OLS, Newey-West HAC & GLS AR(1) at equal standing; one-method-only significance reported as suggestive
Internal control
AI share of civilian IT (within-civilian normalisation)
Seasonality
Sept fiscal-year-end & Oct fiscal-year-start indicators; full month fixed effects as robustness

The chart above visualises the descriptive series with intervention markers; the thesis develops the formal estimation, the announcement-versus-enforcement timing tests, and the measurement-visibility analysis in detail.

06 — Implication

What this means for how we read AI policy.

The pattern is telling. Once the rules took effect, agencies bought AI more often without spending much more per contract — which suggests the memos mainly lowered the paperwork cost of buying AI rather than unlocking bigger purchases. Agencies started experimenting at a higher tempo, not betting larger budgets on flagship systems.

That matters for anyone trying to read federal AI activity from the outside. A jump in AI contract counts after a policy memo is not, by itself, evidence of a spending surge. It is just as plausibly evidence of permission — agencies that could already run AI procurements finding the process less of a slog once the rules arrived.

The repeal makes the same point in reverse. Taking the rules away did not produce an equal and opposite effect — the rules left a deeper institutional mark than their removal did. And OneGov is the reminder that contract data has limits as a window onto AI policy: the biggest access push of all moves enormous adoption while barely showing up in the record.

It also points to where to look next. Where the number of contracts rises faster than the dollars, the interesting question is what kind of AI government is actually buying — cheap per-seat licences for off-the-shelf tools like ChatGPT or Copilot, rather than big custom builds. That's the thread the thesis picks up from here.