How this assessment works
The assessment marks 24 controls from 0 to 3 on the evidence behind each one, weights them for the framework you picked, and renormalizes over what is in scope. The result is a score out of 100 and one of three verdicts. Two gates can hold a verdict down regardless of the score, and neither can lift one.
It is deterministic. The same answers always produce the same score, the same band and the same first thing to fix. There is no model in the scoring path, which is why the whole method fits on one page.
The arithmetic
weighted(control) = inScope ? (mark / 3) * weight : 0
score = round( sum(weighted) / sum(inScope weights) * 100 )
verdict = score >= 80 -> AUDIT-READY AS DESCRIBED
score >= 55 -> PATCHABLE
otherwise -> WOULD NOT PASS
gate 1 (no owner) any in-scope control unowned -> WOULD NOT PASS
gate 2 (dark surface) any in-scope control at mark 0 -> AUDIT-READY AS DESCRIBED held at PATCHABLEThe first three lines are the shipped 90-day audit readiness triage (opens redhub.ai in a new tab) workbook, formula for formula, and a check in the RedHub repository reads them back out of that file and compares them against this site over more than fifty thousand generated cases. Gate 1 is that product’s own gate, unchanged. Gate 2 is the one addition, and it is the only thing the two engines disagree about.
Why two gates, and why they only ever hurt
A weighted average is a machine for hiding a hole. Six controls done well and one done not at all still averages to something respectable, and the number that comes out the other end says nothing about the one that is missing. Both gates exist to stop that.
The no-owner gate fails an assessment when any in-scope control has no named person on it. Not a team, not a job title with nobody in it. Ownership is one of the first things an audit tests, and a control nobody owns has no one who can attest to it, explain the exceptions, or say what changed since last quarter.
The dark-surface gate holds a top verdict at PATCHABLE when any in-scope control reads zero. It exists because we found the failure in our own work: across 132 top-band verdicts our Agent-Readiness Scorecard produced, 64 of them, 48%, sat on a control that was completely absent. The average was filling the hole in.
Neither gate can improve a verdict. A gate that could lift one would be a scoring rule wearing a costume.
Every control and its weight
Weights total 100 for each framework. A dash means that framework does not ask for the control, in which case it never enters the calculation at all, and is not scored and forgiven.
| Control | NIST AI RMF | ISO/IEC 42001 | EU AI Act |
|---|---|---|---|
Inventory and classification Is there a written list of every AI system and AI use case in the business? | 6 | 6 | 5 |
| Is each system classified by risk or impact tier, using the framework you are measuring against? | 5 | 5 | 8 |
| Is there a route by which AI tools nobody approved get discovered and added to the list? | 3 | 3 | 2 |
Ownership and accountability Is one named individual accountable for AI across the business? | 6 | 7 | 5 |
| Does each in-scope system have a named owner, meaning a person and not a team or a job title with nobody in it? | 6 | 6 | 5 |
| Can that owner stop or roll back the system without asking permission? | 3 | 3 | 3 |
Policy and approval Is there a written acceptable-use policy for AI that staff can find? | 5 | 6 | 4 |
| Is there a defined approval step before a new AI use goes live? | 5 | 5 | 3 |
| Are there AI uses your organization has written down as off-limits? | 3 | 3 | 6 |
Risk assessment and testing Is there a documented risk or impact assessment for each in-scope system? | 7 | 7 | 7 |
| Is anything monitored after launch, or was it assessed once and left? | 6 | 5 | 5 |
| Is a system tested against the failure modes you named, before it goes live? | 5 | 3 | 3 |
Data governance Can you say where the data behind each system came from, and on what terms you may use it? | 5 | 5 | 5 |
| Is personal data in AI inputs and outputs handled under a stated basis and a retention rule? | 3 | 4 | 4 |
Record-keeping and traceability Are AI decisions or outputs that matter retained automatically, or only when someone remembers to save them? | 5 | 5 | 5 |
| Can you reconstruct which model version and which prompt or configuration produced a past output? | 3 | 3 | 3 |
Transparency to affected people Are people told when they are dealing with an AI system, or looking at AI-generated content? | 3 | 3 | 6 |
| Could someone affected by an AI-assisted decision get an explanation of how it was reached? | 3 | 2 | 2 |
Human oversight Can a human review and override an AI-assisted decision before it takes effect? | 4 | 3 | 4 |
Vendors and sub-processors Is there a register of AI vendors and the sub-processors behind them? | 3 | 4 | 3 |
| Do your AI vendor contracts cover data handling, model change and notice? | 3 | 4 | 3 |
Incidents Is there a defined route for reporting an AI failure, with someone who receives it? | 4 | 4 | 4 |
| Has that route been exercised, or only written? | 1 | 1 | 1 |
AI literacy Have the staff who use AI had training appropriate to what they use it for? | 3 | 3 | 4 |
| Total | 100 | 100 | 100 |
What each framework is, and what it is not
NIST AI Risk Management Framework 1.0
A voluntary US framework organized into four functions (GOVERN, MAP, MEASURE and MANAGE) for identifying and managing risk across an AI system's life cycle.
The NIST AI RMF is voluntary and confers no certification. A score here describes what you told this tool, not conformance with the framework.
ISO/IEC 42001:2023
An international management-system standard for artificial intelligence: a documented AI management system, with an Annex A set of controls, that an accredited body can certify.
Accredited certification against ISO/IEC 42001 is issued by an accredited certification body after an audit, and a certificate from anyone else is an attestation from whoever issued it. This assessment is neither, and a score here is not a gap analysis performed by a certification body.
EU AI Act (Regulation (EU) 2024/1689)
An EU regulation that assigns obligations by the role you play (provider, deployer, importer, distributor) and by the risk category a system falls into, including prohibited practices, high-risk systems, transparency duties for certain systems and content, and a separate set of duties for general-purpose AI models.
Obligations under the EU AI Act depend on your role and on how a system classifies, and the calendar has moved. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026) deferred both high-risk routes: the Annex III category to 2 December 2027 and the Annex I product category to 2 August 2028. The Article 50 transparency obligations were not amended and apply from 2 August 2026, with a grace period to 2 December 2026 for the machine-readable marking in Article 50(2) on generative systems already on the market. This assessment is not a conformity assessment and does not determine your classification.
What this assessment cannot do
- It cannot verify a single answer. Everything on the result page describes what you told it.
- It cannot classify your systems or determine which role you play under a regulation. Those turn on facts a questionnaire cannot see, and they are what decide which obligations attach.
- It is not an audit, not a certification, not a conformity assessment, and not legal advice. A good score is not evidence of compliance.
- It does not reach into your systems. Nothing is fetched from a hostname you give it, because a tool that fetched a stranger-supplied target would be a security problem wearing a marketing page.
Common questions
- How is the score calculated?
- Each control is marked 0 to 3 on the evidence that exists. That mark is divided by 3, multiplied by the control's weight for the chosen framework, and summed. The total is divided by the weight of everything in scope and multiplied by 100, then rounded to the nearest whole number. Marking a control not applicable removes its weight from the denominator instead of scoring it zero, so excluding something you do not do never counts against you.
- What are the bands?
- 80 and above is AUDIT-READY AS DESCRIBED, 55 to 79 is PATCHABLE, and below 55 is WOULD NOT PASS. Two gates can push a verdict down from there and neither can lift one.
- Why does the no-owner gate exist?
- Because ownership is one of the first things an audit tests, and a control with strong evidence but nobody's name on it has no one who can attest to it. The gate is taken unchanged from RedHub's shipped 90-day audit readiness triage, whose worked example scores 92 out of 100 and still fails on a single unowned row.
- Can I recompute this by hand?
- Yes, and that is the point of publishing the weights. Every control, its weight for each framework, both floors and both gates are on this page. A score you cannot recompute is an opinion with a number attached.
- Do the three frameworks share one set of weights?
- No. Each control carries a separate weight per framework, and a weight of zero means that framework does not ask for it, in which case the control never enters the calculation. That is why the same answers produce different scores against different frameworks.