Risk tiers and the regulatory map

A working internal tier scheme (EU-template, four tiers)

1. Tier 0 — No-go — Statutory prohibitions + internal red lines

   Everything on the EU AI Act Article 5 list, plus the organisation’s own refusals (many firms add: no emotion inference on customers, no social scoring of any kind, no fully automated dismissal decisions). Intake gates catch these; only the executive risk committee can hear an appeal — and for the statutory items, not even them.

2. Tier 1 — High scrutiny — Consequential decisions about people; regulated categories

   Anything making or materially influencing decisions with legal or similarly significant effect — hiring, credit, insurance pricing, housing, health, education — plus whatever maps to EU Annex III areas or state-law "consequential decision" definitions. Full assessment stack, independent validation, committee approval, enhanced monitoring, quarterly attestation.

3. Tier 2 — Standard review — Meaningful but bounded risk

   Customer-facing systems without consequential-decision effect, internal genAI on sensitive data, systems whose failure is costly but reversible. Fast-lane approval against a checklist, standard documentation, routine monitoring, annual attestation.

4. Tier 3 — Minimal — Register and move on

   Low-stakes, reversible, human-mediated uses. One registry record, acceptable-use policy applies, no further per-system obligations. The tier that makes wide registry scope affordable — and the reason nobody needs a committee meeting about a grammar checker.

Enterprise chatbot

A deployed assistant your organisation offers — internal helpdesk or customer-facing. Customer-facing versions pick up disclosure duties (EU AI Act Art 50: people must know they are talking to a machine, applicable since Aug 2026) and full accountability for outputs — Moffatt v. Air Canada made a company honour the refund policy its chatbot invented. Triage: Tier 2 minimum; Tier 1 if it can commit the organisation or touch consequential decisions.

Copilot on internal data

Embedded assistants over your documents, code, or records. The risk centre of gravity is data: access-permission inheritance (does the copilot read documents the user could not?), confidential data in prompts to external APIs, and output reuse. Triage keys on the sensitivity of the data corpus, not the tool brand.

Embedded genAI feature

GenAI that appeared inside existing SaaS — the summariser in the meeting tool, the "write with AI" button in the CRM. Same embedded-AI problem as before, now with content generation attached: where do prompts go, is customer data used for vendor training, can the feature be disabled tenant-wide? Triage output is often a contractual action item as much as a tier.

Employee BYO-AI

Individual use of consumer tools on work content. Not a system you can tier — a behaviour you can channel. The controls are the acceptable-use policy (what data classes may never leave), a sanctioned alternative that is genuinely good, DLP on the egress paths you can see, and training. Registered in the inventory as a usage category with named policy controls, not as a system record.

Triage desk: assign the tier

Interactive decision tree — outcomes:

• Out of scope

  Conventional software — no registry entry. Note the reasoning in the intake log so the decision is reproducible if the system later gains AI features.

• Track as automated decision system

  Not AI, but not ignorable: rule-based automation of consequential decisions is how Robodebt happened. Register as an ADS with an owner and a legal-basis check — some regimes (and every ombudsman) reach it regardless of the AI definition.

• Tier 0 — stopped at the gate

  The request ends here, with documented reasons. Statutory prohibitions (EU Art 5, applicable since February 2025, penalties up to €35M or 7% of worldwide turnover) admit no internal appeal; internal red lines can be appealed to the executive risk committee only.

• Tier 1 — high scrutiny, regimes mapped

  Full assessment stack, independent validation, committee approval, enhanced monitoring — plus the regulatory flags recorded now: EU Annex III classification (obligations from 2 December 2027, but classification and gap work start immediately), Colorado deployer duties (in force since 30 June 2026), NYC LL144 annual bias audit if it screens NYC candidates, and any sectoral duties, which apply regardless of AI law.

• Tier 1 — high scrutiny on internal standards

  No named statute today does not mean no risk: consequential decisions about people take the full Tier 1 treatment on internal standards alone. Regime mapping is re-checked at every attestation — the map changes faster than your systems do.

• Tier 2 — approved only with conditions

  Standard review, but expect mandatory conditions: EU Art 50 disclosure that users face a machine (applicable since August 2026), grounding on approved sources, output monitoring, and clear limits on what the system may promise. Moffatt v. Air Canada is the case to cite when someone asks why — the chatbot’s invented refund policy bound the airline.

• Tier 2 — standard review, fast lane

  Checklist-based delegated approval within SLA, standard documentation, routine monitoring, annual attestation. The workhorse tier — most genuine business AI lands here.

• Tier 3 — register and move on

  One registry record, acceptable-use policy applies, done. Proportionality means being genuinely fast here — this outcome existing (and being reachable in a day) is what keeps the front door busier than the back.