If your company trains AI models and your corporate family clears $500 million in annual revenue, the state of Illinois now has a to-do list for you: publish a formal catastrophic-risk safety framework, submit to an independent third-party audit every single year, and report serious safety incidents within 72 hours — 24 if someone could get hurt. Miss any of it, and the state attorney general can fine you up to $1 million the first time and $3 million after that.
Even if you are nowhere near that revenue line today, keep reading. Illinois is the third state — after California and New York — to enact frontier-AI safety rules, and together the three states represent roughly 40% of the U.S. AI market. Lawmakers are openly counting on that weight to set a de facto national standard while Congress stays on the sidelines. What starts as a big-lab obligation has a way of becoming your enterprise customer's vendor questionnaire, your investor's diligence checklist, and eventually your problem. The law's heaviest requirements bite on January 1, 2028, which gives you about sixteen months to get ready. Here is how to use them.
What Illinois actually passed
On July 6, 2026, Illinois enacted the Artificial Intelligence Safety Measures Act, aimed at a narrow but severe category of harm the statute calls "catastrophic risk": a foreseeable, material risk that developing, storing, using, or deploying a frontier model will materially contribute to death or serious injury of more than 50 people, or more than $1 billion in property damage from a single incident. The covered scenarios center on models assisting with chemical, biological, radiological, or nuclear weapons development, unsupervised cyberattacks, and similarly grave misuse.
The law applies to "large frontier developers" — companies that, together with their affiliates, brought in more than $500 million in gross revenue in the prior calendar year. That is the same revenue line California and New York use, which is deliberate: the three states want one coherent compliance target, not three.
If you are covered, five obligations matter.
1. Publish a frontier AI framework
Beginning January 1, 2028 (or 90 days after you first cross the qualifying threshold, whichever is later), you must write, implement, follow, and conspicuously publish on your website a framework describing how you manage catastrophic risk. The framework has to spell out the thresholds you use to decide whether a model has dangerous capabilities and the mitigations you apply when it does. This is not a marketing trust-and-safety page. It is a public, operative document your auditors — and the attorney general — will test you against.
2. Pass an annual independent third-party audit
This is the provision that makes Illinois the toughest regime in the country. New York requires a single independent audit when a developer first becomes large enough to qualify. Illinois requires one every year. The auditor must work to generally accepted auditing standards, demonstrate real technical competence in frontier-model safety, and opine on whether you have substantially complied with the law — deviations and all, with recommendations for fixing them.
Within 30 days of receiving the report, you must publish a redacted copy with a summary and send it to both the Illinois Emergency Management Agency and Office of Homeland Security and the state attorney general. And you must keep each audit report for the life of the model plus five years. Start thinking now about what "the life of the model" means for your versioning and retention schedules, because your lawyers and your accountants will define it differently.
3. Report critical safety incidents fast
You must report any critical safety incident to the state agency and the attorney general within 72 hours of learning facts strong enough to reasonably establish that it happened. If an incident poses an imminent risk of death or serious physical injury, the clock shrinks to 24 hours, with disclosure to appropriate law enforcement or public-safety authorities. Note the spread: New York also uses 72 hours, but California allows 15 days. If you operate nationally, build your incident-response runbook to the shortest clock — Illinois's — or you will miss it.
4. Register, designate contacts, and pay your share
Starting January 1, 2027 — a full year before the framework and audit duties kick in — covered developers must file a disclosure statement with the state agency identifying the company and its designated points of contact, renewed annually and on any model-ownership transfer or material change. The agency can also assess administrative fees, split pro rata among the covered developers, to fund its oversight. Expect this to function like any other regulated-industry assessment: budget for it, and do not be surprised when the number moves.
5. Protect internal whistleblowers
The law bars developers from adopting or enforcing any rule, policy, or contract term that prevents employees from disclosing violations — or retaliating against them for doing so. Covered developers must also maintain a reasonable, anonymous internal reporting channel. If your employee handbook, severance templates, or contractor NDAs contain broad non-disparagement or confidentiality language with no carve-out for legal disclosures, that language needs surgery before 2027.
Who this covers — and the gap you should not fall into
Read strictly, the law binds only the largest labs: more than $500 million in affiliated revenue, training frontier-scale models. California adds a compute test on top — models trained with more than 10^26 FLOPs (floating-point operations, the standard yardstick for training compute), including fine-tuning — and only a handful of companies publicly acknowledge crossing it today. If you are a twelve-person team fine-tuning an open-weights model for a vertical SaaS product, none of these statutes name you as a regulated party.
Do not mistake "not named" for "unaffected." Three transmission belts will carry these obligations downhill to you:
- Enterprise procurement. The moment large developers standardize on published safety frameworks and annual audits, Fortune 500 buyers will paste those expectations into vendor security reviews. A startup selling AI features into healthcare, finance, or government should expect to be asked for its own safety documentation long before any statute requires it.
- Investor and acquirer diligence. Frameworks, incident logs, and audit trails are becoming diligence artifacts. A company that kept them from early on is simply worth more, and cheaper to acquire, than one that must reconstruct two years of safety history under deal pressure.
- The ratchet. California's compute threshold is subject to annual review by its Department of Technology and can be lowered. Training costs fall every year, which means today's frontier is tomorrow's mid-market. Building compliance habits at small scale is dramatically cheaper than bolting them on after you cross a threshold.
What compliance actually costs — and how to budget for it
Nobody will hand you a fixed price, because the audit market for frontier-model safety is still forming. But adjacent benchmarks exist: industry surveys put a typical AI compliance audit cycle at roughly $25,000 to $150,000 depending on system complexity, with mandatory third-party assessments running up to 40% above a purely internal review. A frontier-model audit — red-team evaluations, framework testing, incident-log review, a formal opinion letter — will land at the top of that range or above it, and Illinois demands one annually, not once.
For a startup, the smart move is to treat compliance as a capital project with an annual operating tail:
- This year: the trial audit. Legal analyses of the Illinois law explicitly suggest running a voluntary audit before one is legally required, to surface gaps while findings are still private. Even if you will never be a "large frontier developer," a lightweight gap assessment against the Illinois framework elements — capability thresholds, mitigations, incident definitions, reporting lines — is the highest-value money you can spend. Price it, capitalize the setup work properly, and expense the recurring portion.
- Ongoing: the compliance ledger. Track every dollar of safety and compliance spend in its own cost center: red-teaming, evaluation tooling, auditor fees, outside counsel, training time, the engineering hours spent writing framework documentation. When an enterprise buyer asks for your security posture, when an investor asks about regulatory risk, or when a future law finally names companies your size, that ledger is your receipts. It also keeps audit fees — which are generally deductible business expenses — cleanly separated from R&D spending that may face entirely different tax treatment.
- Always: the retention discipline. Illinois requires audit reports kept for the life of the model plus five years. Apply that instinct to everything: versioned safety documentation, incident logs with timestamps, training records. Storage is cheap; reconstructing evidence under a 72-hour reporting clock is not.
Your 16-month countdown
Now through end of 2026: read and map. Read the framework elements the law requires and map them against whatever safety practices you already have — evals, red-teaming, deployment gates, incident response. Most serious AI teams have 60% of this informally; the work is writing it down, assigning owners, and versioning it. Put your employee reporting channel and NDA carve-outs on the employment-lawyer list now, while there is no deadline pressure.
January 1, 2027: registration opens. If you are anywhere near the revenue threshold (remember: affiliates count), calendar the disclosure filing and designate your points of contact. If you are far below it, use the date as an internal milestone anyway: framework first draft done, incident-response runbook rewritten to a 72-hour/24-hour clock, compliance cost center live in your books.
2027: the trial audit year. Commission the voluntary audit, remediate the findings, and publish what you are comfortable publishing. A startup that can show an enterprise prospect a real audit summary — even an auctioned-down, voluntary one — stands apart from every competitor waving a trust-center page with stock photos.
January 1, 2028: the law bites. Frameworks published, auditors retained, reporting clocks live. Companies that spent 2027 preparing will experience this as a filing deadline. Companies that did not will experience it as an emergency.
The bigger picture worth watching
Supporters, including two of the largest AI labs, backed the Illinois bill while continuing to ask for a single federal framework instead of a state patchwork — and openly conceding that coordinated state action is the realistic path. Critics in the industry argued the law forces private companies to make subjective safety judgments without national standards or certifications to judge against. Both things are true, and the tension is the point: the states are deliberately creating facts on the ground that a future Congress will have to reckon with. Didech, the bill's House sponsor, has already flagged medical care and education as the next frontiers for AI safety scrutiny. If your product touches either, assume your sector's turn is coming and prepare on Illinois's template.
There is also a legitimate concern worth naming: annual third-party audits cost six figures and then some, the law offers no small-developer on-ramp, and fixed compliance costs inherently favor incumbents. That critique does not change your obligations, but it should change your strategy — shared evaluation tooling, industry-standard framework templates, and audit readiness built into normal engineering hygiene are how smaller teams keep the fixed costs from becoming a moat they cannot cross.
Keep Your Compliance Spending Audit-Ready From Day One
Preparing for a regime like Illinois's is, at its core, a record-keeping discipline: versioned safety documents, timestamped incident logs, categorized compliance spend, and reports you can produce on a deadline. That is exactly the muscle that good accounting builds. Beancount.io provides plain-text accounting that gives you complete transparency and control over your financial data — every compliance dollar tagged, every record version-controlled, no black boxes. Get started for free and build the financial paper trail your future auditor will thank you for.