Tiering criteria, and writing the reasoning down

Classification is a written, reasoned decision, not an attribute. The tier decides the duties, and what gets challenged later is never the tier on its own. It is the reasoning, or its absence. A wrong answer with recorded reasoning fares considerably better than an undocumented right one.

Classify against regimes, not against a house scale

The common failure is an internal three-band scale, high, medium and low, applied to every system, with the regimes mapped on afterwards. It feels tidy and it produces a record that answers no statutory question. A system is high-risk under the AI Act or it is not. It engages Article 27 or it does not. It is a solely automated decision under revFADP Article 21 or it is not. Those are separate determinations with separate answers, and an internal band cannot stand in for any of them.

So the classification record carries, per system: the regimes considered and why each does or does not apply; the tier assigned under each applicable regime; the specific criterion, annex entry or use-case category relied on; the reasoning including why adjacent tiers were rejected; whether anything makes you a provider rather than a deployer; residual uncertainty and where legal advice was taken; and the classifier, the reviewer and the date.

An internal tier is still useful for allocating your own effort, and it should sit alongside the regime determinations rather than replacing them.

The criteria that actually drive the answer

Five questions settle most classifications.

What does the output decide, and for whom? Decisions about identified individuals with legal or significant effects sit at the top of every regime. Decisions about aggregates, portfolios or internal processes generally do not.

Does the use case appear in a listed category? Annex III is a list, and the discipline is to read the entry rather than the summary of the entry. The insurance entry is drawn around risk assessment and pricing in life and health insurance; the credit entry around creditworthiness and credit scoring.

Is there meaningful human involvement in the decision, and can you evidence it? This determines the revFADP Article 21 and GDPR Article 22 position, both of which turn on the decision being based exclusively on automated processing. Note that an approver who accepts recommendations at volume without the information or time to judge them may not supply the human element in substance, and the throughput figures are what reveal it.

Are you provider or deployer, and could that change? Branding, substantial modification and a changed intended purpose are the three Article 25 routes, and each should be a standing question rather than a one-time finding.

Where does the output land? Article 2(1) reaches third-country providers and deployers where the output is used in the Union. Geography of the customer, not the server.

Write reasoning that anticipates the challenge

The reasoning field has three jobs, and most records do only the first.

State the conclusion and the criterion relied on. "Not Annex III, because the Annex III insurance entry covers risk assessment and pricing in life and health insurance, and this system triages motor claims for handling route."

Reject the adjacent tier explicitly. "The adjacent reading is that routing to investigate delays settlement and therefore affects the claimant materially. We considered this. The entry is drawn around risk assessment and pricing, and the routing decision does not set a price or assess insured risk. If the model were extended to influence settlement amounts, the analysis would change."

Record the uncertainty and what would change the answer. That sentence is what makes a record defensible. A classification with no stated uncertainty reads as either an easy case or an unconsidered one, and the reader cannot tell which.

Reclassification triggers

Classifying once at launch and never again is the standard failure. A new purpose, a swapped foundation model, a new affected population or a vendor change can each move a system between tiers, and nothing prompts you to notice. Write the triggers into the record itself: change of intended purpose, extension to a new population or jurisdiction, substantial modification, change of underlying model or vendor, and any regulatory change affecting the criterion relied on. Each trigger names the event, not a review date, because a calendar review catches the change up to twelve months late.

Worked example: two entries on the same day

A bank classifies two systems in one sitting.

A credit decision model for unsecured personal lending. AI Act: high-risk, Annex III creditworthiness and credit scoring. Role: deployer, with a note that the vendor model is used as supplied and any fine-tuning would require reassessment under Article 25. Article 27 fundamental rights impact assessment: applies, because deployers of Annex III creditworthiness systems are inside the narrow population. revFADP Article 21: engaged for Swiss applicants where the decline is automated. Duties dated to 2 December 2027 for the Annex III set, with the Swiss duties live now. Uncertainty: none material.

A model forecasting aggregate default rates for capital planning. AI Act: not Annex III. Criterion: the credit entry concerns creditworthiness evaluation and credit scoring of natural persons; this model produces portfolio-level forecasts and makes no determination about any individual. Adjacent tier rejected: if outputs were later used to set individual pricing or limits, the entry would be engaged, and that is written in as a reclassification trigger. Model risk management: in scope and tiered material. Uncertainty: recorded, because the boundary between portfolio calibration and individual effect is the point a challenge would target.

Same afternoon, same team, different regimes engaged. The second record is the more valuable of the two, because the first was never going to be argued about.

As of 24 September 2026.