Fiscal note
What this Act costs the adopting state, stated the way a legislative budget office reads a cost estimate: categories before numbers, startup apart from steady state, and no revenue anywhere in the arithmetic.
The standing rule this document is held to. Enforcement is never sold as self-funding. Penalties are never booked as revenue. Startup costs are stated apart from steady state. Every figure below is a bracketed default for the adopting state to replace. No fiscal reviewer has been found, and the brackets exist so that whoever takes it has arithmetic on a stated basis rather than invention from a blank page.
Status. Drafted 20 August 2026 against the v3.4 tag. Cost drivers are traced to the provisions that create them; bracketed figures are unpinned defaults and are flagged as such at § 6. Nothing here has been reviewed by a budget office.
1. What the Agency is, and what it is not
The single largest determinant of this Act’s administrative cost is a provision that looks like a limitation on power rather than a budget line. SEC. 3(b) forbids any standard, rule, or mode of validation from conditioning deployment, expansion, or release on the prior affirmative approval of the Agency or of any officer of the State.
The Agency does not operate a premarket licensing or approval program. It therefore has no application queue, approval clock, user-fee relationship, or appeals from license denials. This removes several major cost drivers found in food, drug, and financial preclearance regimes. It does not eliminate the possibility of workload backlogs in rulemaking, standards review, report handling, whistleblower matters, or referrals.
The Agency’s principal functions are receiving, publishing, maintaining, reviewing, and referring:
| Function | Source | Character |
|---|---|---|
| Adopt standards by rule, after notice and technical submissions | SEC. 3(a) | Startup-heavy, then periodic |
| Independently review any standard it incorporates, with power to modify or reject; re-adopt on every amendment | SEC. 3(a) | Recurring, and the most underestimated line below |
| Specify the mode of validation for each standard | SEC. 3(b) | Rides with rulemaking |
| Receive conformity assessments and nonconformity reports | SEC. 3(c)(2)(C)–(D) | Continuous; volume not yet quantified |
| Receive personal certifications | SEC. 8 | Event-based and periodic; volume not yet quantified |
| Receive critical-safety-incident reports | SEC. 9 | Episodic, unpredictable, clock-bound |
| Receive whistleblower reports; act or publish a reasoned declination within [180] days | SEC. 11(a), (e) | Continuous, clock-bound, the principal steady-state driver |
| Protect reporting persons’ identity, including in awards | SEC. 11(b) | Requires secure handling capability |
| Provide by rule for security-sensitive information in reports and awards | SEC. 11(f) | Startup, then maintenance |
| Administer the fund and pay awards | SEC. 10(f), SEC. 11(a) | Contingent on collections |
Prosecution is not an Agency function. The Attorney General enforces (SEC. 10(e)), and those costs are estimated separately at § 4.
2. Volume: bounded in scope, not yet quantified
SEC. 1 narrows the universe of covered models through a compute threshold and prospective frontier-equivalent designation. That does not, by itself, establish the number of regulated entities or annual filings. Duties can arise from developer, provider, and deployer conduct, and more than one entity or controlling person may be connected to a covered system.
Filing volume is driven by deployments, material expansions, releases, periodic certifications for sub-material changes, nonconformity reports, critical-safety incidents, and whistleblower submissions. It therefore cannot be inferred solely from the number of frontier-model developers. State population and AI adoption may also affect how many covered systems are made available to residents, how often reportable events occur, and how much enforcement work follows.
The records duty beginning at [10^22] operations reaches below full frontier-model coverage. Those records are ordinarily retained rather than routinely filed, limiting regular intake costs, but production requests and resulting investigations can still create workload.
Until a fiscal analyst inventories covered entities and models, estimates the relevant event frequency, and tests low-, central-, and high-volume scenarios, this note cannot responsibly claim a number of annual filings. The staffing figures below are bracketed structural placeholders, not volume-derived estimates.
3. Startup costs (year one)
| Category | Driver | Bracketed default |
|---|---|---|
| Standards development and rulemaking | SEC. 3(a) — technical submissions, drafting, notice-and-comment, independent review of any incorporated standard | [2.0–3.0 FTE] for the year, plus [$—] contracted technical expertise |
| Public availability of incorporated material | SEC. 3(a) requires all incorporated material to be publicly available without charge | [$—] — see § 5, the paywall line |
| Secure information handling | SEC. 11(b), (f) — identity protection and security-sensitive material | [$—] one-time, facilities and systems |
| Intake and publication systems | SEC. 3(c)(2)(C), SEC. 9, SEC. 11 | [$—] one-time; capacity to be set from low-, central-, and high-volume scenarios |
| Fund establishment | SEC. 10(f) | Administrative, nominal |
Year one carries workload that the steady state does not because the Agency’s own standards do not yet exist. The truth-telling, reporting, records, whistleblower, and administrative provisions operate from the effective date; the core SEC. 2 deployment duty begins at provisional commencement after [180] days using the interim standards in SEC. 3(c)(4). The Agency’s unfinished rulemaking therefore does not postpone that commencement, but the state still incurs startup costs while its permanent standards are developed.
4. Steady-state costs (annual, from year two)
| Category | Driver | Bracketed default |
|---|---|---|
| Report intake and review | SEC. 3(c)(2), SEC. 9 | [1.0–2.0 FTE] |
| Whistleblower processing | SEC. 11(e) — the [180]-day act-or-declare clock is the binding constraint | [1.0–2.0 FTE] |
| Standards maintenance and re-adoption | SEC. 3(a) — every amendment to an incorporated standard requires fresh adoption | [0.5–1.0 FTE] |
| Legal and rulemaking support | Across | [0.5–1.0 FTE] |
| Attorney General — enforcement | SEC. 10(e) | Stated separately below |
Attorney General. Enforcement cost is episodic and cannot yet be reduced to a reliable annual average. The Act creates document- and expert-intensive civil and criminal matters against potentially well-resourced defendants. Whether this requires standing staff, contingent appropriations, outside technical expertise, or some combination depends on expected case volume and the adopting office’s existing capacity. A fiscal analyst should model investigation, electronic discovery, expert evidence, prosecution, and appellate work under low-, central-, and high-enforcement scenarios. A first litigated case is likely to cost materially more than later cases, but no amount is asserted here without a state-specific basis.
5. Two line items that are easy to miss
Re-adoption on amendment. SEC. 3(a) provides that no amendment to an incorporated standard takes effect in this State until the Agency adopts it by the same procedure. Where a state incorporates a standard maintained by an outside body that revises frequently, this converts a one-time rulemaking into a recurring one. This is a deliberate anti-delegation feature, and it has a price; the price belongs in the estimate rather than in a footnote.
Paywalled standards. SEC. 3(a) requires all incorporated material to be publicly available without charge. If the Agency incorporates a standard that is ordinarily sold, the State must secure public availability — by license, by negotiation, or by declining to incorporate that standard. This is a real budget line that regimes of this kind routinely discover after enactment rather than before.
A cost claim from the other side of the argument — added 23 August. The accelerationist literature’s own futurist, writing at FAI in May 2026, on what autonomous systems do to oversight economics: systems “can record every action they take, with precise telemetry … ‘auditing’ and ‘inspection’ become dramatically lighter weight; the regulator’s agents can check in at essentially no cost, and compliance becomes verifiable in seconds” (Ball, Hyperdimensional, 28 May 2026, ⚠ P — press corpus § 5). If that is right, SEC. 12’s records duties and the Agency’s steady-state review costs in § 4 are overstated above, not understated — recorded here because a fiscal note should log the argument that cuts against its own caution, and this one comes from the opposition’s pen.
6. Revenue: none, stated deliberately
This Act creates no fee authority. SEC. 3(b) bars pre-approval, so there is no application to charge for, and nothing else in the text authorizes the Agency to charge anyone anything. Administration is funded by appropriation, and by nothing else.
The fund at SEC. 10(f) receives penalties, fines, disgorgement, and other monetary recoveries after restitution, and the adopting state elects whether the balance is appropriated to the Agency’s functions or reverts to the general fund. It is not a funding source and must not be scored as one, for three reasons stated plainly:
- It receives nothing until there is a successful action, and the Act is drafted to make successful actions rare by making the underlying conduct rare.
- Whistleblower awards under SEC. 11(a) are paid from the same fund at 10 to 30 per cent of sanctions collected. Where the fund is also the Agency’s appropriation source, the two draw on one pool, and a state making that election should say so on the face of its appropriation rather than discover it in year three.
- A regime whose budget depends on collections has a stated incentive to collect. The whole architecture of this Act is an argument that enforcement incentives should be legal rather than financial, and it would be incoherent to fund it in the manner it criticizes.
The correct scoring posture: cost is appropriated, recoveries are windfall, and the estimate is sound if not a dollar is ever collected.
6b. A state fiscal office has now priced an AI act — the first dollar donor
Added 24 August 2026. Colorado’s Legislative Council Staff issued a fiscal note for SB 26-189 — the successor to the repealed duty-of-care act — on 4 May 2026, and it is the first professionally-produced dollar figure this note can cite for AI-act administration: a $46,190 General Fund appropriation, 0.4 FTE (an Assistant Attorney General in the Department of Law for rulemaking and stakeholder engagement), $56,286 total expenditures with centrally appropriated costs stated separately ($20,193), startup isolated to FY 2026-27, and out-year expenditures honestly zeroed. Attribution: analysts John Armstrong and Dhivahari Vivek, Legislative Council Staff (primary PDF; retrieval logged in the verification record’s next update).
What it does and does not do for this note. It does not price this Act — Colorado’s bill is a civil disclosure regime enforced by an existing office, while this Act stands up an Agency and a criminal docket, so the figure is a floor for the narrowest administrative posture, not an estimate. What it supplies is the genre’s method, demonstrated on this subject matter: one named office, startup severed from steady state, central costs broken out, revenue at zero — the discipline §§ 3–6 of this note already promise and can now cite in practice. The fiscal seat’s question 1 — “is a note with no numbers reportable in your state?” — now has a comparative answer in hand: Colorado reported one, and this is what its arithmetic looked like.
Correction, 24 Aug evening (the primary arrived). The figures above are conformed to the final revised fiscal note of 6 May 2026, now in hand: a $46,190 General Fund appropriation and 0.4 FTE Assistant Attorney General “in FY 2026-27 only,” $56,286 total expenditures ($46,190 personal services + $10,096 centrally appropriated), out-years at zero. The larger figures this section first carried ($100,403 / 0.8 FTE / $120,596, dated 4 May) came in ⚠ from reporting of the initial note; whether that initial note carried them is unverified — the final revised note controls, and the register carries the entry (E36). The floor is therefore half what this file first reported — which sharpens, not weakens, the point the section makes. The same note records, in passing, that “the Attorney General was ordered by U.S. District Court to not initiate enforcement” in X.AI LLC v. Weiser — logged at the standing watch.
6c. The second note in the genre — pricing a delay at zero
The same office priced the delay itself. SB 25B-004 (2025 special session) — the bill that moved SB 24-205’s duties from 1 February to 30 June 2026 — carries a final fiscal note of 10 September 2025: $0 revenue, $0 expenditures, 0.0 FTE in every year, no appropriation required (primary in hand). The method survives even at zero: impacts stated by year and type, the enacted bill reflected, the zeros written down rather than implied. For § 7’s comparator work, that is the point — the genre’s discipline is the table, not the total.
7. Open items for the fiscal seat
Named so that the seat’s work is review rather than design.
- Every bracketed figure above. FTE ranges are structural defaults reasoning from the functions at § 1, not estimates drawn from a comparator agency’s actuals. They are the item most in need of replacement.
- Comparator selection. The nearest enacted siblings are the California, New York, and Illinois frontier regimes whose standards SEC. 3(c)(4) freezes. Whether any of them carried a published fiscal note, and what it assumed, is unpinned. Pinning one is worth more than refining the brackets.
- The Attorney General line — contingency figure or reasoned non-quantification (§ 4).
- The paywall exposure (§ 5) — whether it should be a stated line or a conditional note, and at what assumed figure.
- The fund election (§ 6) — whether the appropriated-to-Agency election should carry a drafting note warning of the SEC. 11 interaction.
Corrections to FrontierAIAccountabilityProject@proton.me; they enter the errata register like everything else. Nothing in this document is legal or fiscal advice, and it has not been reviewed by any budget office.