The Number AI Regulators Keep Citing: 0.7 Percent of Models Get Audited

Archi4 min read

aipolicyauditingregulation

Contents
  1. The Number That Keeps Showing Up in Hearings
  2. What the 0.7 Percent Actually Counts
  3. Why the Gap Persists
  4. The Policy Debate Is Really About Sampling
  5. What the Number Means for Companies
  6. The Number to Watch Next

The Number That Keeps Showing Up in Hearings#

Somewhere between 0.5 and 1 percent of deployed AI systems get any form of independent audit. That figure is the one regulators and standards bodies keep returning to in testimony and draft guidance, and it is the reason so much recent policy fights over the mechanics of inspection rather than the principles. Call it the audit gap: almost every model that touches a loan application, a hiring decision or a medical triage queue has never been examined by anyone outside the company that built it.

The number is soft. It comes from surveys of compliance teams, from counts of certified auditors, and from the small share of systems that fall under existing financial or medical device review. But the direction is clear, and the size of the gap explains a specific pattern in AI policy: proposals that sound like they are about fairness or safety usually turn into arguments about who does the checking and how often.

What the 0.7 Percent Actually Counts#

Most of the coverage of that figure treats it as a compliance score. It is not. It is a measure of inspection capacity.

Think about how many AI systems exist in production. Large firms run hundreds. Banks have models for credit scoring, fraud, customer service routing and document review, plus older statistical models that were never labeled AI. Now count the people qualified to review any of them independently: auditors with machine learning backgrounds, regulators with compute budgets, academic groups with data access agreements. That pool is small, and it is smaller than the number of systems that would need review under almost any serious rule.

The gap is not caused by industry refusing inspection. It is caused by arithmetic. You cannot audit tens of thousands of systems with a few hundred reviewers, no matter how good the rules are.

Why the Gap Persists#

Three forces keep the inspection rate low.

Access. An audit needs data, training logs and sometimes source code. Companies guard these for trade secret and liability reasons, and the legal agreements that grant access take months to negotiate.

Standards. There is no common checklist. Two auditors reviewing the same hiring model can reach different conclusions because they test for different failure modes. Until test methods converge, audit results are hard to compare and easy to dispute.

Cost. A serious review of a mid-sized model can run into six figures and several weeks. For a company with a dozen models in production, that adds up fast, and the cost lands before any finding exists.

The Policy Debate Is Really About Sampling#

Once you accept the 0.7 percent figure, the interesting question becomes how to raise it without inspecting everything. Regulators are quietly converging on sampling logic borrowed from food safety and financial accounting: inspect a statistically chosen subset, require self-assessment for the rest, and escalate when something in the sample looks wrong.

That approach has a known weakness. Sampling works when items are similar. AI systems are not. A credit model and a chatbot share almost no risk profile, so a sample drawn across a company tells you little about any single system. Sector-specific sampling, where the review pool is narrower and the failure modes are better understood, is the version most likely to survive contact with reality.

The other lever is documentation. If companies publish structured model cards covering training data, evaluation results and known limits, reviewers can triage faster and spend expensive audit hours on the systems that need them. This shifts work from inspection to disclosure, which is cheaper but also easier to game.

What the Number Means for Companies#

The practical lesson for anyone shipping AI is that external review is unlikely to arrive on schedule. Do not build a compliance plan that assumes a regulator will audit your system before you deploy it. Build one that assumes nobody will look until something breaks.

That means internal testing with the same discipline an auditor would apply: hold out data, test for subgroup performance, log decisions, and keep records a stranger could follow. Companies that treat the audit gap as a reason to skip testing are the ones most exposed when the first serious review finally happens, often after an incident.

The Number to Watch Next#

The audit rate is the number to track because it moves slowly and tells you more than any single law. A jump from 0.7 to even 3 percent would mean thousands of new reviewers, new standards and new costs baked into product timelines. A flat year means the rules on paper are still ahead of the capacity to enforce them.

Watch two signals: the number of accredited AI auditors and the share of high-risk systems that get any external review before deployment. Neither is glamorous. Both determine whether the next round of AI rules changes anything on the ground.

More articles