Before you ask if your AI is high risk, list it
You cannot tier what you have not listed: a plain case for an AI use-case inventory first, with twelve fields, three internal control tiers and a separate EU AI Act legal-review flag.
Status as of October 3, 2026. This is a practical opinion, not legal advice. Check dates and sources before you rely on any of it.
Opinion summary
You cannot tier what you have not listed. Life sciences teams ask “is this AI high risk?” before they can say what AI they have. A tier is a judgment about a named use. The name, the purpose, and the setting have to sit on a record before the judgment means anything.
Three plain claims.
First, list the use before you label the risk. NIST AI RMF 1.0 says outright that mechanisms to inventory AI systems should be in place. GOVERN 1.6 states, verbatim: “Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities.”
Second, the EU AI Act (Regulation (EU) 2024/1689) makes high-risk classification depend on intended purpose and Article 6, including its Annex I and Annex III routes.
Third, a simple inventory plus an internal tiering scheme gives Quality and AI leaders one place to hold the inputs the law asks for. It does not replace the legal classification.
A skeptical Quality reader says we already have a computerized system inventory and GxP risk assessments. My view: do not build another list. Add AI fields to the record you have. Keep the internal control tier and the legal status as two separate fields on that same record.
Why the inventory comes first
Someone asks how serious the AI is, and the room cannot say what the system is for, who uses it, or what the output touches. Intended purpose, EU role, and context are the facts. The inventory is where those facts have to live before a tier means anything.
NIST AI RMF 1.0 MAP 1.1 asks teams to document, in this excerpt, “Intended purposes, potentially beneficial uses, context-specific laws, norms and expectations”. Write that in one or two sentences. Name the user and the setting. An illustrative sentence, not taken from any company, is: “Drafts first-pass summaries of deviation records for QA reviewers; output is edited before filing.”
The January 2025 FDA draft guidance on AI to support regulatory decision-making for drugs and biologics places context of use at Step 2 of its seven-step framework, after defining the question of interest. I use only that context-of-use idea. The guidance is a draft.
The Act also turns on your role. Article 3(3) defines provider. Article 3(4) defines deployer as “a natural or legal person, public authority, agency or other body using an AI system under its authority”, except where the system is used in the course of a personal non-professional activity. Field 4 records provider, deployer or another operator under Article 3, and it allows “not yet assessed”.
My view: tie each use case to the computerized system record you already keep. Intended purpose belongs to the use, not to the vendor product name. Illustrative only: one hosted model drafts deviation text for one team and ranks job applicants for another.
MAP 1.5 states, verbatim, “Organizational risk tolerances are determined and documented.” My recommendation: put GxP relevance on the row as well. Does the output affect a GxP record, decision or process, and which process? I use that answer later as a tier driver. That use is my recommendation.
What the EU AI Act actually asks
EU AI Act status needs legal review; this piece is not legal advice.
Whether an AI system is high risk under the EU AI Act depends on its intended purpose and Article 6, including the Annex I and Annex III routes. The original text is Regulation (EU) 2024/1689. The AI Omnibus, Regulation (EU) 2026/1744, amends Articles 4, 6 and 113. Primary EUR-Lex text checked on October 3, 2026: the amending act was made on July 8, 2026, published on July 24, 2026, and entered into force on July 27, 2026. The Commission AI Act page says the AI Omnibus “entered into force on 27 July 2026”.
The Annex I route needs both conditions in Article 6(1), “where both of the following conditions are fulfilled”. The AI is intended as a safety component of a product, or is itself a product, covered by Annex I legislation. And that product “is required to undergo a third-party conformity assessment”. The Annex I route can apply to AI that is, or is a safety component of, a medical device or IVD that must undergo third-party conformity assessment, for example under the medical device and IVD rules.
Article 6(1a) and 6(1b) were inserted by Regulation (EU) 2026/1744, Article 1(8). Article 6(1a) says that AI used solely for non-safety-related aspects of quality control does not qualify as a safety component, subject to Article 6(1b) where failure or malfunction would endanger health and safety. This does not exempt anything from the Act as a whole. Paragraph 1c says a product that needs third-party conformity assessment solely for risks other than health and safety does not meet Article 6(1), point (b).
The Annex III route is Article 6(2): “AI systems referred to in Annex III shall be considered to be high-risk.” Annex III areas, for example employment, are use areas. If an area applies and you believe the system is not high risk, record the Article 6(3) exception, its profiling override, and the provider’s Article 6(4) assessment and registration duties. Article 6(3) says the system “shall always be considered to be high-risk where the AI system performs profiling of natural persons”.
As of October 2026, Annex III high-risk rules (systems classified under Article 6(2) and Annex III) apply from December 2, 2027. Annex I rules (systems classified under Article 6(1) and Annex I) apply from August 2, 2028.
Article 4, as replaced by Article 1(5) of the amending act, says that providers and deployers “shall take measures to support the development of AI literacy of their staff”.
Where NIST and ISO fit
NIST AI RMF 1.0 (NIST AI 100-1, January 2023) is voluntary. The PDF says, verbatim, that it is “voluntary, rights-preserving, non-sector-specific, and use-case agnostic”.
GOVERN 1.3, in excerpt, is the line I use for tiers: “determine the needed level of risk management activities based on the organization’s risk tolerance”.
MAP 5.1’s excerpt points at “Likelihood and magnitude of each identified impact”. MANAGE 1.2’s excerpt says treatment is “prioritized based on impact, likelihood, and available resources or methods”. MEASURE 2.1 states, verbatim, “Test sets, metrics, and details about the tools used during TEVV are documented.” I use that last line for the heaviest internal tier, where fit-for-use evidence has to exist. MAP 4.1’s excerpt, “including the use of third-party data or software”, is why data and vendor sit on the same row as the use.
The NIST AI RMF page says the NIST AI 600-1 GenAI profile was released on July 26, 2024, and that AI RMF 1.0 “is being revised”. The quotes above are from the January 2023 PDF, opened on October 3, 2026.
ISO/IEC 42001 is an AI management-system standard. I have not read its text. I make no claim about what it requires.
Internal tiers, not legal labels
An internal control tier sets the minimum controls you apply. My view: keep that tier and the EU AI Act legal status as two separate fields on one record. The legal-review flag is the second field. Counsel settles the legal classification.
The Act decides high-risk status by intended purpose and Article 6, including the Annex I and Annex III routes. Where an Annex III area applies, the Article 6(3) exception, its profiling override, and the provider’s Article 6(4) assessment and registration duties come into play. A use can carry intensive internal controls and still sit outside a high-risk classification. The reverse can happen too: an HR screening tool in an Annex III area, for example employment, with little GxP impact.
The placing question I use is: “How much does the AI output shape the decision, and how bad is a wrong decision?” It borrows the idea of the January 2025 FDA draft guidance, which says “Model risk is a combination of two factors”: model influence and decision consequence. That guidance is a draft. It covers AI used to support regulatory decision-making for drugs and biologics. It does not set up an enterprise tier scheme. The borrow is an analogy only, for the placing question.
My recommendation: use GxP relevance as a tier driver, beside how far a qualified person checks the output before anyone relies on it. My recommendation: every tier definition, placing question, and minimum control in the tool below is mine. The illustrative examples are mine too. They are not a standard.
Close
My view: put these AI fields on the computerized system or use-case record you already keep. One row can hold the intended purpose, the role, the GxP relevance, and the legal question as separate fields. The tool below sets out twelve inventory fields, three internal control tiers, and the legal-review flag.
The tool: twelve inventory fields
The example values are illustrative, not taken from any company. Field 5 records separate Yes, No, or Unsure answers to 5(a) and 5(b), so each Article 6(1) condition is answered. Both conditions must be met before the Annex I route applies. Field 7 is where you file the Article 6(3) exception, its profiling override, and the provider’s Article 6(4) assessment and registration duties, when an Annex III area applies and you believe the use is not high risk. Fields 9, 10 and 11 record data, human oversight, and the vendor. Field 12 is not an internal severity tier. The next section says when to set it.
One illustrative record with all twelve fields. Gold dots mark the inputs to the legal analysis; teal dots mark the two fields that drive the internal control tier.
| # | Field | What to record | Example |
|---|---|---|---|
| 1 | Use case name and ID | A short name and a unique ID for the use, not the vendor product | UC-014 Deviation summary drafting |
| 2 | Intended purpose | What the system is for, who uses it, and in what setting, in one or two sentences | Drafts first-pass summaries of deviation records for QA reviewers; output is edited before filing |
| 3 | Named business owner | The accountable person in the business, by role and name | QA operations lead, Site X (example) |
| 4 | EU role | Provider, deployer or other operator under Article 3; “not yet assessed” is allowed | Deployer |
| 5 | Annex I route (two conditions) | (a) Is the AI intended as a safety component of an Annex I product, or itself such a product? Yes / No / Unsure. (b) Is that product required by the relevant Annex I legislation to undergo third-party conformity assessment? Yes / No / Unsure. Both conditions in Article 6(1) must be met | (a) No; (b) No |
| 6 | Annex III area | Which Annex III area applies, if any (for example employment) | None. Or: Employment (CV screening tool) |
| 7 | Article 6(3) assessment recorded | If an Annex III area applies and you believe it is not high risk: who assessed it, when, and where the record is; whether the system performs profiling of natural persons (which keeps it high risk); and, if you are the provider, the status of the Article 6(4) assessment and registration duties | N/A. Or: Assessed September 12, 2026 by counsel; no profiling; record ref |
| 8 | GxP relevance | Does the output affect a GxP record, decision or process? Yes, with which process / No | Yes: deviation management (GMP) |
| 9 | Data used | Data sources, any personal or patient data, and whether data leaves the company | Deviation text; no patient data; vendor-hosted EU region |
| 10 | Human oversight | Who reviews the output, at what point, and whether they can override it | QA reviewer edits and approves every summary |
| 11 | Vendor or third party | Supplier, model, version, contract owner | Vendor A, hosted LLM, version pinned |
| 12 | EU AI Act legal-review flag | Yes / No / Unsure. Not an internal severity tier. See the next section for when to set it | Unsure |
The tool: internal control tiers and the legal-review flag
Internal control tier is not EU AI Act classification. The tiers set the minimum controls you apply. The Act decides high-risk status by intended purpose and Article 6, including the Annex I and Annex III routes. A use case can sit in internal Tier 3 and still not be legally high risk, and the reverse can happen too, for example an HR screening tool in an Annex III area with little GxP impact. The placing question borrows the idea from the January 2025 FDA draft guidance (model influence and decision consequence): “How much does the AI output shape the decision, and how bad is a wrong decision?” This is an analogy only. The guidance is a draft. It does not set up an enterprise tier scheme. My recommendation: the tier definitions, placing questions, illustrative examples, and minimum controls below are mine, not a standard.
Three questions from the tier table, asked in order, place a use case in an internal control tier. The scheme is mine, not a standard and not EU AI Act classification.
| Tier | Definition | Placing question | Illustrative examples | Minimum controls |
|---|---|---|---|---|
| 1 Basic controls | General productivity. No GxP record or decision, no patient data, a person uses the output at their own discretion | “If the output is wrong, is the worst case some lost time?” | Drafting internal emails; summarizing public literature for a team meeting | Inventory entry; approved tool list; AI literacy basics; acceptable use rules |
| 2 Enhanced controls | Supports work that feeds GxP or regulated content, and a qualified person reviews every output before use | “Does a qualified person check every output before it enters a record or decision?” | First-draft deviation summaries; literature screening suggestions for a pharmacovigilance reviewer | Tier 1 controls, plus named owner, documented intended purpose, defined human review step, vendor assessment, periodic spot checks |
| 3 Intensive controls | The output materially influences a GxP, patient safety, product quality or regulatory submission decision, or people rely on it without full review | “Would a wrong output likely reach a batch, patient, or regulator without being caught?” | Model output used as evidence in a submission; predictive model guiding a process control decision | Tier 2 controls, plus fit-for-use evidence (test sets, metrics), documented risk assessment, change control, performance monitoring, defined escalation; consider the FDA draft credibility steps where relevant |
Each tier keeps every control from the tier below and adds its own. These minimum controls are mine, not a standard.
The EU AI Act legal-review flag is field 12, not a tier. Record Yes, No, or Unsure. Set it to Yes or Unsure when the use may meet both conditions in Article 6(1), for example AI that is, or is a safety component of, a medical device or IVD that must undergo third-party conformity assessment, or when it may match an Annex III area. Screening question: “Could this meet both conditions in Article 6(1), for example under the medical device or IVD rules, or does it match an Annex III area?” Illustrative examples: software intended for diagnosis (a possible medical device or IVD route); AI ranking job applicants (an Annex III area). When the flag is Yes or Unsure, stop and record the legal analysis before go-live, with counsel: the role, the two Annex I route conditions, Annex III, and the Article 6(3) exception, its profiling override, and the provider’s Article 6(4) assessment and registration duties. The flag is separate from the control tier and can be set at any of Tiers 1 to 3.
How field 12 is set. It sits beside the control tier, not inside it, and counsel settles the legal classification. Not legal advice.
The tool: framework mapping
These rows were checked on October 3, 2026. Each quoted phrase is marked verbatim or excerpt, and the words link to the source.
I work with Quality and AI leaders on questions about using GenAI and agents in GxP workflows defensibly. If you want to talk through a use-case inventory for your team, get in touch. You can also build your AI inventory and risk tiers.