Compliance and Regulation

The idea in one minute#

Regulation and standards for AI ask engineering teams for five things, in different words: know what you run (an inventory, with risk classes), manage its risks (a documented process), control its data (governance and privacy), keep humans in charge (oversight and transparency), and be able to prove it (records, logs, testing evidence). Nearly all of that evidence is produced by controls this course has already covered. Compliance, done well, is mostly a matter of collecting what a well-run system emits anyway — and done badly, is a pile of documents describing controls that do not exist.

Dates and statuses below were checked on 4 October 2026. This is engineering orientation, not legal advice.

A picture#

flowchart LR
  subgraph REQ["What is asked"]
    direction LR
    R1[":i-scale: <b>EU AI Act</b><br/><small>law, by risk class</small>"]
    R2[":nist: <b>NIST AI RMF</b><br/><small>voluntary framework</small>"]
    R3[":i-scroll-text: <b>ISO/IEC 42001</b><br/><small>certifiable management system</small>"]
    R4[":i-lock: <b>Privacy and sector law</b><br/><small>GDPR, health, finance</small>"]
    R5[":i-building-2: <b>Customer assurance</b><br/><small>SOC 2, questionnaires</small>"]
  end
  REQ --> MAP[":i-list-checks: <b>One control set</b><br/><small>each control mapped<br/>to every requirement</small>"]
  MAP --> EVID
  subgraph EVID["Evidence, produced by the system"]
    direction LR
    E1[":i-database: <b>AI inventory, AI-BOM</b>"]
    E2[":i-scale: <b>Evaluation and red-team results</b>"]
    E3[":i-scroll-text: <b>Audit logs</b>"]
    E4[":github: <b>Change history</b><br/><small>prompts, policies, models</small>"]
    E5[":i-siren: <b>Incident records</b>"]
  end
  class R1,R4 memory
  class R2,R3,R5 queue
  class MAP compute
  class E1,E2,E3,E4,E5 io

How it really works#

The EU AI Act for engineers#

The Act regulates by role and risk class. First find your role:

Fine-tuning a model substantially, or putting your name on a system, can move you from deployer to provider.

Then the risk class:

The schedule, after the Digital Omnibus agreement of May 2026 (formal adoption was still pending when checked):

2 Feb 2025   prohibitions; AI literacy
2 Aug 2025   general-purpose model obligations
2 Aug 2026   transparency obligations (Article 50)
2 Dec 2027   high-risk obligations, standalone systems (Annex III)   ← deferred from Aug 2026
2 Aug 2028   high-risk obligations, AI in regulated products (Annex I) ← deferred from Aug 2027

High-risk requirements, mapped to engineering#

The cybersecurity article explicitly names attacks this course has covered. A system built to the earlier lessons has the substance; the remaining work is assembling the evidence.

NIST AI RMF, as a working structure#

The COSAiS overlays apply SP 800-53 controls to AI use cases, including agents; the agent-specific overlays were still in development when checked. Organisations already operating SP 800-53 or FedRAMP controls should expect to extend, not replace, them.

ISO/IEC 42001#

A management-system standard for AI, structured like ISO 27001: policy, roles, risk and impact assessment, controls, internal audit, continual improvement, with a certificate at the end. It is what customers increasingly ask for as proof, in the way they ask for ISO 27001 or SOC 2 for security generally. It certifies that you have and follow a process — not that any particular model is safe.

Privacy law still applies#

AI did not suspend data-protection law. The recurring engineering issues:

The AI inventory#

Everything else depends on knowing what exists. Keep one register:

system         name, owner, purpose, users
classification AI Act role and risk class; internal risk tier
models         identifiers and versions, hosted or self-hosted, provider terms
data           categories processed; personal data? residency; retention
capabilities   tools and MCP servers; autonomy level by action class
tenancy        who shares it
controls       links to threat model, evaluation suite, policies, runbooks
status         last review, last red-team exercise, open risks

“Shadow AI” — tools and agents adopted without review — is the gap this closes. The practical mechanism is the paved path: when the governed gateway is the easiest way to get a model key, most usage registers itself.

Evidence without ceremony#

Auditors, customers and regulators all ask for the same artefacts. Produce them from the running system, not by hand:

A control that leaves no evidence will be treated as absent. Designing the evidence in — a log line, a CI artefact, a signed record — is part of designing the control.

A sensible order of work#

1. Build the inventory and classify each system.
2. Stop anything prohibited; plan dates for anything high-risk.
3. Adopt one control set (this course’s lifecycle) and map it once to each framework you answer to.
4. Wire the evidence sources.
5. Close gaps in order of risk, not in order of the standard’s clauses.
6. Review on a schedule and on every material change.

Remember this#

• Five asks in every framework: inventory, risk management, data control, human oversight, evidence.
• EU AI Act: find your role and risk class; high-risk deadlines moved to December 2027 and August 2028, transparency duties did not.
• The Act’s cybersecurity requirement names poisoning, adversarial inputs and confidentiality attacks explicitly.
• Privacy law applies in full; erasure is the hard one.
• One control set, mapped to many frameworks, with evidence emitted by the system.

Try it#

1. Write the inventory entry for one AI system you know.
2. Decide its AI Act role and risk class, and justify it in two sentences.
3. For three controls you rely on, name the evidence each produces. Which produce none?

Check yourself#

1. What can turn a deployer into a provider under the AI Act?
2. Which data-protection obligation is hardest to meet for fine-tuned models, and what design choice avoids the problem?
3. Why should evidence be generated by the system rather than written afterwards?

Sources#

• EU AI Act omnibus agreement: postponed high-risk deadlines and other key changes
• EU approves delays and other amendments to certain EU AI Act obligations
• NIST AI agent standards: research note