AI Governance That Works: Model Inventory, NIST AI RMF, ISO 42001, and the EU AI Act Timeline
An AI governance framework that works starts with a list, not a policy: one register of every AI system you build, buy or run, each with a named owner, your role under the law, a risk tier and a date. This lesson walks you through building that register in a half-day workshop, then shows where NIST AI RMF, ISO 42001, the EU AI Act timeline and the deployer duties plug into it, so the next hard question from your board takes an afternoon to answer instead of three weeks.
The Problem: A Board Question You Cannot Answer
"Which of our AI systems are high-risk under the EU AI Act, and are we compliant?" A non-executive director sent that one line, the board secretary forwarded it to legal, and legal promised an answer before the next board meeting.
Legal asks IT for the list of AI systems. IT has four projects, the ones the AI team built. Procurement, asked separately, knows of eleven vendor products with AI features switched on. Nobody has the agents. The spreadsheet that finally reaches the board took three weeks and was out of date on the day it was sent.
Nobody was careless, and the company probably has a perfectly decent rulebook, perhaps the one-page policy from your first AI acceptable use policy and readiness assessment. The trouble is that the director asked a counting question, and a policy cannot count. You cannot classify, control or report on systems you have never listed.
What does a missing inventory cost?
It costs you clear ownership, and it leaves you exposed to dates that already apply. Four facts for your next steering committee:
- Most breached firms had no AI policy in place. IBM's 2026 Cost of a Data Breach Report (July 2026) put the global average breach at USD 4.99 million, up 12% from USD 4.44 million in 2025. Coverage of the report by ComplexDiscovery on 1 August 2026 says 68% of breached organizations had no AI governance policy in place (35% had none, 33% had one still in development), and 92% of those hit by an AI-related incident lacked proper AI access controls. The figures are self-reported: a signal, not a census. The article's conclusion is this lesson's premise: "The practical response starts with discovery, not policy."
- The penalty ceiling is real money. For Article 50 breaches it is up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher, and for qualifying smaller firms the lower of the two, according to Jones Walker's analysis of 16 July 2026.
- Some dates have passed. Article 50 transparency duties have applied since 2 August 2026, including a deployer's duty to disclose deepfakes. The only grace period went to providers: those whose generative systems were placed on the market before 2 August 2026 were given until 2 December 2026 to add Article 50(2) machine-readable marking (Jones Walker, 16 July 2026; White & Case, 4 August 2026).
- The big date is closer than it sounds. Annex III high-risk obligations apply from 2 December 2027. From this lesson's publication on 12 November 2026, that is just under 13 months to put documentation, oversight and log retention in place for every qualifying system: a tight runway, on our reading.
The regulator that moved its own deadline
The EU AI Act entered into force on 1 August 2024. Its stand-alone high-risk obligations were due on 2 August 2026, and many programmes were built around that day.
Then the date moved. The Commission proposed the Digital Omnibus on AI on 19 November 2025. The result was published as Regulation (EU) 2026/1744 in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, days before the original deadline (European Commission AI Act policy page, updated 19 October 2026; White & Case, 4 August 2026). Annex III high-risk duties moved 16 months, to 2 December 2027. Annex I duties, for AI embedded in regulated products, moved to 2 August 2028.
The part people missed: what stayed put, and what the law added.
| Duty | Applies from | Moved by the Omnibus? |
|---|---|---|
| Original prohibitions and AI literacy | 2 February 2025 | No. Already in force. |
| Governance rules and general-purpose AI model obligations | 2 August 2025 | No. Already in force. |
| Article 50 transparency duties (chatbot disclosure, emotion-recognition notices, a deployer's deepfake disclosure under Article 50(4), public-interest text) | 2 August 2026 | No. No grace period for deployers. |
| Article 50(2) machine-readable marking, for providers of generative systems placed on the market before 2 August 2026 | 2 December 2026 | Grace period, for those providers only. |
| New prohibitions: AI that generates intimate or sexually explicit images of identifiable people without their explicit consent (Article 5(1)(ba)), and AI that generates child sexual abuse material (Article 5(1)(bb)) | 2 December 2026 | New. Added by the Omnibus. |
| Annex III high-risk duties, including the deployer duties in Articles 26 and 27 | 2 December 2027 | Yes, from 2 August 2026. |
| Annex I product-embedded high-risk duties | 2 August 2028 | Yes. |
The Omnibus also rewrote Article 4 on AI literacy: providers and deployers "shall take measures to support the development of AI literacy of their staff" and of others operating AI on their behalf, and a new sentence says the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual." The AI literacy program lesson turns those measures into a record you can show.
Two kinds of organization got hurt. One had built everything around 2 August 2026 and over-prepared for a duty that slipped. The other heard "postponed" and walked past the transparency duties that arrived on schedule. Jones Walker's advice is the line to remember: "This should be treated as a reprieve rather than a repeal." Conformity assessment, documentation, oversight and EU database registration still apply to high-risk systems, at the new dates.
A third kind had a quiet week. It had a current inventory: someone re-ran the classification column, changed the due dates and closed the file by the end of the afternoon.
Dates move. The inventory is the durable asset.
These dates are checked against the Commission's policy page, two law-firm readings and the EUR-Lex text of Regulation (EU) 2026/1744; confirm against the Official Journal before relying on them. This is not legal advice.
Where this lesson sits on the ladder
This is the fifth and final lesson of Level 3, AI Operational, on the L1 to L5 ladder. The previous four lessons each gave you one artifact per system: a release and rollback runbook (Lesson 11), a data foundation pack (Lesson 12), a threat map and control plan (Lesson 13) and, last time, a unit economics sheet from AI FinOps: token economics, model routing and controlling the AI bill. Today you build the one register across them that points to each as evidence. Next, the internal AI platform lesson opens Level 4 by turning what you repeat by hand into shared gateways and components.
How an AI Governance Framework Works: Think Like a Fleet Manager
Forget AI for a minute and think about a company that runs vehicles. The fleet manager's first tool is never the road traffic act. It is the register: every vehicle, its plate, its named driver, what it is used for, and whether the company owns it or leases it. Each vehicle has a class (bicycle, van, heavy truck), and the class decides the licence, the inspection rhythm and the insurance.
Road law sets what each class must have and from which date, some duties on the manufacturer and some on whoever drives. The fleet manager's routine (assign, record, inspect, repair) is the working method. An operator's certification is an outside party confirming the routine is real.
Two obvious things get forgotten with AI. First, nobody argues about road law before they know how many trucks they own. Second, a van that starts carrying chemicals changes class, though it is the same van. Class follows use, not hardware.
What is an AI inventory?
An AI inventory (also called a model inventory) is the single list of every AI system the organization builds, buys or runs, with its owner, purpose, role and risk tier. It is the fleet register. The vendor product with an AI feature switched on is a vehicle in your fleet, and so is the agent somebody set up last quarter.
Agent rows need extra fields: platform, the tools and datasets they reach, spending and iteration limits, and who can switch them off. Keep agents in the same register; a second list is how agents go missing.
What is risk tiering?
Risk tiering, or AI risk classification, means sorting systems into a small number of classes so that scrutiny matches potential harm. It is the vehicle class. Without tiers you either inspect the bicycle like a truck or the truck like a bicycle.
Use two tiers:
- The legal tier. The EU AI Act sorts systems into four tiers, per the European Commission: unacceptable risk (prohibited practices), high risk, transparency risk, and minimal or no risk.
- The internal materiality tier. Materiality is your own rating (high, medium, low) of who is affected and what happens if the system is wrong. An agent that moves money may sit in the law's lowest tier and your own highest.
How do NIST AI RMF, ISO 42001 and the EU AI Act fit together?
They do three different jobs: one is law, one is a way of working, and one is a management system an auditor can certify. You do not choose between them, and collecting all three is not operating any of them.
| Framework | What it is | In the fleet | What it gives your register |
|---|---|---|---|
| EU AI Act | Law. Four risk tiers, with duties and dates for each, split between providers and deployers. | Road law | The role column, the EU tier column and the date column |
| NIST AI RMF | A voluntary US framework built on four functions: Govern, Map, Measure, Manage. | The fleet manager's routine | A label for each governance activity, so you can see which ones you skip |
| ISO/IEC 42001:2023 | A certifiable AI management system standard. | The operator's certification | Scaffolding: policy, risk and impact assessment, internal audit, management review |
The NIST AI Risk Management Framework is voluntary. Version 1.0 was released on 26 January 2023, and its Generative AI Profile (NIST AI 600-1) followed on 26 July 2024, according to the NIST AI RMF page. NIST says version 1.0 is being revised and gives no date. Our working shorthand, not NIST's: Govern is policy and roles, Map is the inventory and its context, Measure is evaluation and monitoring, Manage is treatment and incident response.
ISO/IEC 42001:2023 was published in December 2023. It has the same high-level structure as ISO 27001 and ISO 9001 (Clauses 4 to 10), with 38 reference controls in Annex A across nine sections. Its key requirements include an AI policy (5.2), AI risk assessment and treatment (6.1.2 and 6.1.3), an AI system impact assessment (6.1.4), a Statement of Applicability, internal audit (9.2) and management review (9.3). These come from Advisera's guide to ISO 42001, a secondary source: check the standard before quoting clause numbers to an auditor. Say it in every board meeting: an ISO 42001 certificate is scaffolding, not EU AI Act compliance.
Size the framework to the fleet: every size needs the register; only some need the certificate.
What does the EU AI Act ask of a deployer?
Mostly, to run high-risk systems with trained oversight, keep their logs, tell your workforce, and sometimes assess the impact on people's rights first. A provider develops an AI system and puts it on the market or into service under its own name. A deployer uses an AI system under its own authority. Buy a CV screening tool and the vendor is the provider, you the deployer; build a chatbot in-house on a vendor's model and you are its provider. In the fleet, the manufacturer answers for the brakes and the driver for the driving.
For high-risk systems, Article 26 of the AI Act (not amended by the Omnibus) requires deployers to:
- use the system in line with the provider's instructions for use;
- assign human oversight to people with "the necessary competence, training and authority";
- keep the logs the system generates automatically, where they are under the deployer's control, for a period suited to the purpose and "of at least six months" unless other law says otherwise;
- as employers, inform "workers' representatives and the affected workers" before putting a high-risk system into use at the workplace.
These EU AI Act deployer obligations arrive with the Annex III date, 2 December 2027. A reviewer with no authority to overrule the system is not oversight.
A fundamental rights impact assessment (FRIA) is the deployer's assessment, made before deploying a high-risk system, of how its use could affect people's fundamental rights. Article 27 covers only some deployers: bodies governed by public law, private entities providing public services, and deployers of systems for credit scoring or creditworthiness of individuals (fraud detection excluded) or for risk assessment and pricing in life and health insurance. Critical-infrastructure systems are excluded.
A data protection impact assessment (DPIA) is the GDPR assessment your privacy team runs when processing personal data is likely to pose a high risk to people. Under Article 27(4), as amended by the Omnibus, the FRIA can cross-refer to or incorporate the DPIA, and Article 27(5) has the AI Office provide a questionnaire template. Run them as one exercise: one evidence pack, the DPO and the governance owner together, two sign-offs.
One deployer duty is not waiting for 2027. Under Article 50(4), a deployer that uses AI to generate or manipulate image, audio or video deepfakes must disclose that the content is artificially generated or manipulated, and AI text published to inform the public on matters of public interest must be disclosed unless it had human review under someone's editorial responsibility. Both have applied since 2 August 2026.
What can banking model risk rules teach everyone else?
They give you the best list of inventory fields available, and some humility. On 17 April 2026 the Federal Reserve, OCC and FDIC issued revised model risk management guidance, SR 26-2 (with OCC Bulletin 2026-13), superseding SR 11-7. As summarised by Forvis Mazars on 4 June 2026, it is risk-based, introduces model materiality, and is most relevant to institutions with more than $30 billion in assets.
Two lessons. First, its inventory expectations are a ready-made template: purpose, ownership, inputs and outputs, assumptions, materiality, validation status, monitoring and limitations, with validation generally before first use. Second, the guidance leaves generative and agentic AI out of scope, because such models "are novel and rapidly evolving", and the agencies said a request for information on AI model risk would follow. In October 2026 Reuters reported, as relayed by Rajesh Beri's D(AI)LY Brief on 13 October 2026, that the agencies had begun asking banks how they would shut down AI agents and whether they can audit agents' decisions, with that request still pending. Both are inventory fields: who holds the off switch, and where the audit trail is. Stop waiting for a finished rule for LLMs and borrow the fields.
The Workshop: Build Your AI Inventory This Week
You need a half-day workshop in two blocks, then a one-hour committee meeting a week later. You end with a named artifact: the AI System Inventory with Risk Tier and Control Plan, plus a one-page committee charter.
Who is in the room: this method is about counting and roles, so seat the people who know what exists and who answers for it. First block (steps 1 to 3): the governance owner (general counsel, chief risk officer, CISO or a named head of AI governance), procurement, the CIO or a delegate, and the data protection officer, because many rows hold personal data. Second block (steps 4 to 10): add legal or compliance, HR (workforce tools and the duty to inform workers) and two business owners of live systems. Security and finance need not attend: send them the rows linked to their Lesson 13 and 14 artifacts for correction.
What to bring: the shadow AI findings from the shadow AI lesson, the runbooks (Lesson 11), data packs (Lesson 12), threat maps (Lesson 13) and unit economics sheets (Lesson 14), procurement's vendor list, the instructions for use from your AI vendors, and any DPIAs already on file. Most of your evidence already exists; today you index it.
The ten steps
- Register every AI system. One row each, including vendor AI features and agents. Merge three lists that have never met: IT's project list, procurement's vendor list and the shadow AI findings. Do not debate risk yet. You are counting vehicles.
- Fill the core fields. Owner, purpose, users, data used, model or vendor, inputs and outputs, known limitations, and autonomy level. Autonomy is one of four levels: suggests, drafts, acts with approval, acts alone. An owner is a named person, not a team and not "IT".
- Record your role. Provider or deployer, row by row. Most vendor rows make you the deployer; anything you built and run under your own name makes you the provider. Every dated duty depends on this cell; mark doubtful ones "legal to confirm".
- Classify against the EU tiers. Prohibited (including the two new prohibitions from 2 December 2026), high risk, transparency or minimal, with the date each duty applies. A working call with a flag beats an empty cell.
- Assign an internal materiality tier. High, medium or low, based on who is affected and what happens if the system is wrong. Ask: "If this is wrong for a month and nobody notices, who gets hurt and how badly?"
- List the deployer duties each row triggers. For high-risk deployer rows: the named oversight person and their authority to overrule, a log retention setting of at least six months, the date workers and their representatives are informed, a FRIA decision (required or not, and why) and the DPIA reference. For transparency rows: the disclosure rule, and whose duty it is.
- Attach controls by tier. State what each tier requires (impact assessment, oversight, testing before first use, monitoring, disclosure) and link each control to evidence from Lessons 11 to 14. A control with no link is a wish.
- Set the review triggers. Review before first use and on change: a new model, new data, new users or a new purpose. This is the van that starts carrying chemicals. Write the triggers on the row.
- Stand up the committee. Name the chair, the members, the meeting rhythm and what gets escalated. The inventory is its standing agenda: new, changed and overdue rows.
- Map to a framework and book the audit. Tag each activity you now do as Govern, Map, Measure or Manage; a function with no tags is your gap. Decide on purpose whether ISO 42001 certification is a goal or a reference. Book a date within the year for someone outside the team to sample three rows and ask for the evidence. At Level 5, continuous AI assurance with red teaming and audit-ready evidence replaces that annual sample with tests that run on every change.
A filled-in example
Illustrative only: picture a 2,000-person staffing and facilities company operating in three EU countries. Everything here is invented, and the tier and role calls are working calls, not legal advice.
| System (autonomy) | Owner | Role | EU tier (working call) | Materiality | Controls and deployer duties | Next date |
|---|---|---|---|---|---|---|
| CV screening tool from a vendor (ranks, human decides) | Head of recruitment | Deployer | High risk (recruitment is an Annex III area); legal to confirm | High | Trained reviewers with authority to overrule; logs kept at least six months; DPIA with the DPO; no FRIA on our working call (private recruitment) | Deployer duties from 2 December 2027 |
| Candidate matching under a public employment service contract (suggests) | Public contracts lead | Deployer | High risk; legal to confirm | High | FRIA (here we are a private entity providing a public service), built on the DPIA evidence | FRIA before deploying under the 2 December 2027 regime |
| Shift allocation tool for facilities staff (allocates, supervisor approves) | Facilities operations director | Deployer | High risk (allocating work and monitoring performance); legal to confirm | High | Inform workers' representatives and staff before use; trained supervisors; logs six months or more | 2 December 2027; worker briefing set for spring 2027 |
| Marketing image generator (vendor tool, placed on the market before 2 August 2026; drafts) | Marketing lead | Deployer | Transparency (synthetic content) | Low | Ask the provider for its Article 50(2) marking plan, due by 2 December 2026; our own deepfake disclosure rule for published images | Our disclosure duty: since 2 August 2026 |
| Invoice-matching agent, built in-house (acts with approval above a threshold) | Finance operations lead | Provider | Minimal | High (money moves) | Iteration and spend caps, approval rule, override log, runbook | Review on any model change |
The CV screening tool was never on IT's list: procurement bought it, recruitment runs it. It carries the high-risk duties, and they are deployer duties: trained oversight, kept logs, a DPIA.
The candidate matching row does similar work, yet it needs a FRIA and the CV tool does not. The difference is not the technology but who the company serves: under that contract it is a private entity providing a public service. The FRIA follows the deployer's situation, so it lives on the row, not in the vendor's brochure.
The image generator looks trivial: low materiality, drafts only. But it carries two dates. Its provider must mark output as synthetic by 2 December 2026, because the tool was on the market before 2 August 2026. Your own deployer duty, to disclose deepfakes you publish, has applied since 2 August 2026 with no grace period. The role column tells you whose duty is whose.
The invoice-matching agent is the reverse. The law's working tier is minimal, yet internal materiality is high, because money moves. Record only the legal tier and your riskiest operational system gets the lightest review.
In our illustration the team ends the half-day with 17 systems against the 4 that IT had listed, and 3 with no owner. Those three rows open the committee's first agenda.
The template: what every row holds
Copy this AI inventory template into your register tool; a spreadsheet is fine to start.
| Group | Fields per row |
|---|---|
| Identity | System name; type (built, bought, embedded feature, agent); business owner; purpose |
| Reach | Users and people affected; data used; model or vendor; inputs and outputs; assumptions and limitations; autonomy level |
| Classification | Role (provider or deployer); EU tier and applicable dates; internal materiality tier |
| Deployer duties | Oversight person and authority; log retention period; worker information date; FRIA decision and reason; DPIA reference; disclosure rule |
| Assurance | Validation status and date; monitoring in place (link to runbook); data definitions and sources (link to data foundation pack); security status (link to threat map); cost owner (link to unit economics sheet) |
| Control | Controls required and evidence; for agents, off-switch holder and audit trail; review triggers; last review; committee decision |
Then write the one-page committee charter:
- Chair: the governance owner, by name.
- Members: legal or compliance, the DPO, IT, security, finance, procurement, HR and rotating business owners.
- Rhythm: how often it meets, and the standing agenda.
- Escalation rules: what must come to it before go-live (anything high materiality or high-risk) and what goes above it.
Back to the director's question. With this register, the answer is a filter on two columns, role and EU tier, and a covering note. It names the three high-risk calls still with legal and links each one's evidence by name: the release and rollback runbook from Lesson 11, the data foundation pack from Lesson 12, the threat map from Lesson 13 and the unit economics sheet from Lesson 14, plus the deployer duties the row owes from 2 December 2027. That is an afternoon, and the director gets a list instead of a promise.
Failure Modes and the Metrics to Track
Governance fails quietly: the documents exist, the committee meets, and nothing is controlled. Each failure has a one-conversation test.
How do you spot AI governance that only exists on paper?
Ask for a number, a date or a story:
| Failure mode | What it looks like | How to spot it |
|---|---|---|
| Policy without inventory | A signed policy and no list of what it governs. | Ask how many AI systems are in use. A guess is the finding. |
| Incomplete register: vendor features | It covers what the AI team built and misses AI switched on inside bought products. | Compare the register with procurement's vendor list and the shadow AI findings. |
| Incomplete register: agents | Agents that run on a schedule or trigger appear in no row. | Ask who holds the off switch for each agent running today. An agent with no row has no answer. |
| Role confusion | Duties land on the wrong party: you wait on a vendor's deadline while your own deployer duty is already live. | Ask whose duty each dated cell is. A row with no role has no answer. |
| Classify once | Tiers set at approval, never revisited when the system, users or law changed. | No row has a review date later than its approval date. |
| Deadline theatre | Everything is planned around one future date. | The plan has no entry for duties already in force. |
| Oversight on paper | A named human reviewer who never sees the outputs, or cannot overrule them. | Ask the reviewer to describe the last output they rejected. |
| Certificate as compliance | A framework stands in for the work. | ISO 42001 is presented to the board as meeting the EU AI Act. |
The oversight test generalises: if nobody can recall a single thing the policy stopped or changed, you have a document, not a control.
Deadline theatre deserves one more line, because the 2026 postponement fed it. Your plan should show duties in three groups: already in force (the original prohibitions, AI literacy, Article 50 transparency), dated 2 December 2026 (the new prohibitions and your providers' marking), and dated 2 December 2027 or 2 August 2028 (the high-risk duties).
Which metrics show the inventory is alive?
Track seven, and report them at every meeting.
| Metric | How to measure it |
|---|---|
| Inventory coverage | Systems registered divided by systems found by procurement, network and survey checks. |
| Row completeness | Share of rows with a named owner, a role, an EU tier and a materiality tier. |
| Deployer duty readiness | Share of high-risk deployer rows with a named oversight person, logs kept at least six months, and a recorded FRIA decision and DPIA reference. |
| Validation before first use | Share of high-materiality systems validated before go-live. |
| Agent control | Share of agent rows with a named off-switch holder and a working audit-trail link. |
| Overdue reviews | Count of rows past their review date. |
| Committee decisions per quarter | Number of decisions, including refusals and conditions. |
The last is the honest one: a committee that has never refused or attached a condition is an audience, not a control.
Frequently Asked Questions
Is ISO 42001 certification enough for EU AI Act compliance?
No. ISO/IEC 42001 certifies that you run an AI management system; the EU AI Act sets legal duties per system, role and date, and a certificate does not discharge them. Use 42001 as scaffolding for policy, risk assessment and audit, and keep the register as the place where each legal duty is tracked.
When do the EU AI Act high-risk obligations apply?
Under Regulation (EU) 2026/1744, in force since 27 July 2026, Annex III high-risk duties apply from 2 December 2027 and Annex I product-embedded duties from 2 August 2028. Confirm against the Official Journal text before relying on any date.
Which EU AI Act duties apply before the high-risk deadline?
The original prohibitions and AI literacy have applied since 2 February 2025, general-purpose AI model obligations since 2 August 2025, and Article 50 transparency duties, including deployer deepfake disclosure, since 2 August 2026. Two new prohibitions added by the Omnibus, on non-consensual intimate imagery and AI-generated child sexual abuse material, apply from 2 December 2026.
Does every deployer of a high-risk AI system need a fundamental rights impact assessment?
No. Article 27 requires a FRIA from deployers that are bodies governed by public law or private entities providing public services, and from deployers of credit scoring or life and health insurance pricing systems. Where you also owe a GDPR DPIA, the FRIA can cross-refer to it or incorporate it.
Who should own AI governance in a mid-sized company?
One named executive accountable for AI risk, such as the general counsel, chief risk officer, CISO or a head of AI governance, who chairs the committee. Each system then has its own named business owner on its row, so accountability is never left with "IT".
Key Takeaways
- Start with the list, not the policy: you cannot control AI systems you have not counted, including vendor features and agents.
- Give every row a role (provider or deployer) and two tiers: the EU AI Act tier and your own materiality tier.
- Treat the 2026 postponement, in Jones Walker's words, as "a reprieve rather than a repeal". Article 50 has applied since 2 August 2026, new prohibitions apply from 2 December 2026, Annex III from 2 December 2027.
- Deployers owe their own duties: trained oversight with authority, logs kept at least six months, informing workers, and for some deployers a FRIA that can build on the GDPR DPIA.
- The EU AI Act is law, NIST AI RMF is a way of working (Govern, Map, Measure, Manage), and an ISO 42001 certificate does not equal legal compliance.
- Review before first use and whenever model, data, users or purpose change; test oversight by asking what the reviewer last rejected.
- Level 3 ends with artifacts per system and one register across them, built by hand. That breaks at twenty systems, which is why Lesson 16 builds a shared internal platform.
Skill.re