AI use case risk classification, explained.

AI use case risk classification means deciding, idea by idea, how much review an AI use needs and why

One AI idea, from a sentence to the inventory

1. Someone has an ideaA lab head describes what the AI would do, in her own words, and attaches a design, a model card (the vendor's summary of what the model is and does) and the vendor's terms.
2. The facts are checkedSomeone compares the form with the attachments. What does the design say about who acts on the output and what data is sent?
3. The idea is classifiedFive questions, each answered against a named clause, lead to a proposed review tier and a list of reviewers. This is the step this guide is about.
4. People review and decideQuality, privacy and security confirm their parts. A review board approves, approves with conditions, sends back or rejects, and signs.
5. The use is registered and watchedAn inventory entry records the owner, the conditions and the next review. A rule change or a new model version sends it back to step 3.

Step 3 is the subject of this guide. The people, company and figures in the example are fictional.

Why a "minimal risk" AI can still be critical in a GMP plant, and three more surprises

Provider, deployer, critical GMP use, human in the loop: the words this work runs on

Fourteen terms are defined in the one-page summary.

Five questions sort almost every AI idea in a drug company, and each has its own rulebook

In shortNo single rule covers an AI idea in life sciences. Teams answer five questions: its AI Act category, its GxP criticality, its model risk, what personal data it uses, and what the vendor's terms allow. GxP is the family of good-practice rules for manufacturing, laboratories, clinical work and distribution, and GMP is its manufacturing member. A company's review tier usually follows the highest-risk answer on any of them.

The EU AI Act's four groups, and where a pharma company's AI ideas tend to land

In shortThe Act bans a short list of practices, sets heavy duties for high-risk systems, requires honesty about some others and sets no rules for the rest. "Minimal risk" is the Commission's plain-words name for that last group, and it says most AI in use falls there. We found no split for life sciences. In a drug company, hiring, worker-management and device uses are the likely exceptions.

How a lawyer walks an idea down Article 6 and the Annex III list

Four in ten systems could not be placed with confidence in the best-known study

The share of AI ideas in each group is less settled than vendor slides suggest. We found no published split for life sciences. The most cited study is appliedAI's review of 106 enterprise AI systems, published in March 2023: 18% came out high-risk, 42% low risk, and 40% unclear.16

It read an earlier draft of the Act, before the Article 6(3) filter existed, so treat it as a warning about ambiguity and leave the percentages out of your budget.17 Its authors note that even the 18% sat above the 15% maximum they say the Commission's impact assessment assumed.16 The Commission says today only that most AI systems in use fall into the minimal group.13

Provider or deployer: why one sentence in Article 25 matters to a sponsor

A company that buys a tool is usually a deployer. Article 25(1)(c) says that anyone who modifies the intended purpose of an AI system, "including a general-purpose AI system", that was not classified as high-risk, so that it becomes high-risk, is treated as a provider.1 A sponsor who points a general chat tool at job applications may have just made the company a provider of a high-risk system. A review that checks the tool and skips the use will miss this.

Reach matters too. The Act covers deployers located in the Union and, for providers and deployers elsewhere, cases where "the output produced by the AI system is used in the Union".1 A US team's tool may or may not fall under it. That is a question for counsel, fact by fact. The Commission's draft guidelines themselves say that only the Court of Justice can give an authoritative interpretation.15

What the AI Omnibus changed: high-risk rules move to December 2027 and August 2028

In shortThe delay is law. Regulation (EU) 2026/1744 was published July 24, 2026 and took effect July 27, 2026.8 Bans and general-purpose model rules already apply, and transparency has applied since August 2, 2026. Anything you read that calls the delay a proposal, or that gives August 2026 for high-risk duties, is out of date.

Article 4 AI literacy became a duty to support, with no guaranteed level

Many company policies and training decks still quote the original Article 4, which asked for a "sufficient level" of AI literacy. The Omnibus replaced it. Providers and deployers now "shall take measures to support the development of AI literacy" of their staff, and the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual".8 A company may still set a training bar of its own, and GMP training rules still apply in a plant. Update the quotation.

Medical devices stay in Annex I, Section A, for now

The Omnibus moved machinery out of Section A. It left medical devices and in vitro diagnostics there.8 The Commission's separate December 2025 proposal to revise the device regulations would move them to Section B, which would leave AI in devices mostly to the device rules. Parliament's file on that proposal, read November 3, 2026, says "Awaiting committee decision", so stay close to it if you build AI into a device.18,19

The Commission's high-risk guidelines are a late draft

Article 6(5) asked for guidelines by February 2, 2026. The Commission published a draft in three parts on May 19, 2026, and updated the page on July 23, 2026, with no final text as of November 3, 2026.15 One law firm reports that the consultation closed on June 23, 2026.20 The draft is useful for its examples: it says that not every AI system used in recruitment or at work is high-risk, and it walks through a recruitment tool that loses the filter because it profiles recruiters.15

Fines under Article 99: the top tier is reserved for banned practices

Draft EU GMP Annex 22: what it says about generative AI, and why its wording may move

In shortThe draft covers static, deterministic models (they do not change in use, and the same input gives the same output) in critical manufacturing uses and says generative AI and large language models "should not be used in critical GMP applications". It is a draft. It has no legal force, the consultation pushed back, and EMA's own report says there "may be viable pathways". Plan for it, and keep your policy easy to change.

Section 1's two key sentences: the line on generative AI and the human-in-the-loop line

The draft applies to "all types of computerised systems used in the manufacturing of medicinal products and active substances, where Artificial Intelligence models are used in critical applications with direct impact on patient safety, product quality or data integrity".2 It applies to static models that do not adapt during use, and to models with deterministic output, meaning identical inputs give identical outputs. It then says: "the document does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications."2

On non-critical uses it says that "personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use, i.e. a human-in-the-loop (HITL)".2

What the draft would ask of a model in a critical use

Summary of the draft's sections.2 Read the text before relying on any line.

From consultation to workshop: where the argument has gone

EMA's report, EMA/156789/2026, records that the drafting group "has therefore discussed the scope of Annex 22 to allow dynamic or adaptive models, probabilistic models, and Generative AI/LLMs in GMP applications", provided they meet the annex under a documented, risk-based control strategy, and that "concerns remain" about what that strategy needs.4

One session's examples, some of them pilots or theoretical frameworks, included an AI system supporting deviation management and an LLM supporting a GMP risk assessment.4 Those are industry positions that a drafting group will now weigh. The report closes on the drafting group's next step, and that step is still open.

Batch release has its own rules, whatever the final annex says

An idea that releases batches with no quality sign-off fails before anyone opens Annex 22. In the US, 21 CFR 211.22(a) gives the quality control unit responsibility and authority to approve or reject drug products.6 In the EU, Annex 16 sets out certification by a Qualified Person, the named expert who certifies batches, and batch release, in operation since April 15, 2016.3

FDA's model influence and decision consequence grid, and what EMA's paper and the GAMP AI guide add

In shortFDA's January 2025 draft guidance gives a seven-step way to establish that an AI output is credible, and rates model risk by influence and consequence. It covers AI whose output supports a regulatory decision on safety, effectiveness or quality. Many companies borrow its grid for internal triage, which is sensible, as long as nobody mistakes the result for an FDA finding.

FDA's AI credibility framework: what the draft covers, and what it leaves out

The draft addresses AI used "to produce information or data intended to support regulatory decision-making regarding safety, effectiveness, or quality for drugs". It leaves out drug discovery. It also leaves out "operational efficiencies (e.g., internal workflows, resource allocation, drafting/writing a regulatory submission) that do not impact patient safety, drug quality, or the reliability of results from a nonclinical or clinical study".7 It is nonbinding, marked not for implementation, and still a draft on its docket page, FDA-2024-D-4689.7 FDA's drug center, CDER, counts over 500 submissions with AI components from 2016 to 2023.23

The grid in one picture: influence across, consequence up

FDA's manufacturing example is a vision system checking fill volume in vials. Fill volume bears directly on quality (FDA calls it a critical quality attribute), so the consequence is high. Release testing also measures fill volume on a sample from each batch, so the model's influence is low, and the model risk comes out medium.7 The lesson for a sponsor: adding an independent check lowers influence and leaves the consequence where it was.

What FDA warning letter 722591 says about one firm's AI-written documents

The letter is addressed to one firm, on that firm's facts, alongside other findings under current good manufacturing practice (CGMP), FDA's manufacturing rules. It cites rules already in force. Read it for what it says to that firm: we draw no general FDA policy from it, and the letter states none.

EMA's paper, the EMA and FDA principles, GAMP and CSA: how far to lean on each

Personal data under GDPR and the vendor's training terms: where design documents contradict forms

In shortPrivacy and contract questions change the outcome more often than sponsors expect. A deviation report can name an analyst. A support chat can reveal a patient's illness. A vendor may train on your inputs unless an option is signed. These facts sit in the design and the contract, which is why those documents need reading.

What to read in a vendor's terms before the idea goes further

No law sets these answers for you. Your own policy does. Five facts come up every time:

• whether inputs and outputs may be used to improve the vendor's services unless you opt out;
• which region processes the data;
• how long inputs are kept;
• how much notice you get before the default model version changes;
• whether the use runs through a contract you already hold for another purpose.

A tool bought for meeting notes is the classic way a generative model reaches a regulated process, since the contract and the login already exist.

Digital Omnibus changes to GDPR are still a proposal: the Council agreed its mandate in October 2026, with Parliament's position and negotiations ahead.12 Work from the GDPR as it stands.

Sets or suggests: how one verb changes a deviation assistant's whole classification

In shortWork the example from the top of this page down the five questions, twice: once as designed, once with a QA reviewer deciding each class. The AI Act answer stays put. The GMP and model risk answers move, and so does the review it needs. The names and the contract are separate problems with their own fixes. Every answer here is a proposal until the company's reviewers and board decide.

The second route, a static validated model, and why this sponsor takes the first

A team could also replace the generative model with a static one and validate it for the critical use. Draft Annex 22 is written for exactly that: test data held apart, acceptance criteria at least as good as today's manual process, explainability during testing, change control.2 The work takes much longer. In this fictional example the company's own plan compares about 6 weeks for the suggestion route with 4 to 6 months for a static, validated model.

A person who reviews 40 suggestions a month and agrees with nearly all of them can drift into approving without reading. That is why the example's conditions include a quarterly review of agreement and override rates, and why draft section 3.3 asks that an operator's consistent performance be monitored like any other manual process.2

Eight ways AI classification goes wrong in pharma, and how each one shows up

In shortMost errors come from reading too little, such as taking a form at its word and leaving the design unopened. None has a published rate that we could find, so the table says how to look for each one in your own files.

Reasonable readers can reach very different answers on the same systems

appliedAI's 2023 study also showed how far readings can spread. Depending on how the unclear cases fell, the high-risk share of its sample ran from 18% to 58%, with doubt clustering in critical infrastructure, jobs, law enforcement and product safety.16 The text has moved since then, and the filter in Article 6(3) now exists. Consistency across reviewers still takes a written method.

What a wrong AI classification can cost: findings, fines, breaches and a stalled idea

In shortFour costs sit apart. A GMP finding reaches product and license. An AI Act fine is large, and national enforcement is new. A data breach has published figures, though from all sectors. The everyday cost is the good idea that waits, and nobody publishes that one.

What belongs in an AI inventory or AI register, so that a rule change takes hours

In shortAn AI inventory is a list of every AI use, each with enough facts to answer a regulator, an auditor or your own board. The AI Act does not ask deployers for a general company register, so this is your own control. Its most useful column records the clauses each decision relied on.

Frameworks that can shape your method, none of them required

NIST's AI Risk Management Framework, released January 26, 2023, and its Generative AI Profile of July 26, 2024, offer voluntary structure.30 ISO/IEC 42005, published in May 2025, offers process guidance for assessing an AI system's impact on people and society, without certification,31 and ISO/IEC 42001 covers AI management systems.31 A company can adopt them, and a management system standard can sit on top of this work. None of them replaces the five questions.

AI governance in life sciences: training is ahead of board attention, and US state rules are in flux

In shortAI budgets are rising and few companies have scaled their AI use. Surveys show training in place at about two in three companies, and governance at the board in about half. Outside Europe the picture shifts faster: Colorado replaced its AI law before it took effect. Software for this work is a young market.

More ideas are coming, and few have reached production. An intake built for today's volume will meet many more once pilots turn into programs.

Gartner's release of February 17, 2026 puts AI governance spending at $492 million in 2026, passing $1 billion by 2030. Its release of August 26, 2026 lists AI governance platforms at $275 million for 2026, inside a wider market for securing AI, so read any market size with its definition. We saw excerpts of both releases and could not read their full text.33 We found no public figure for how many AI ideas a pharma company receives a year, or how long a decision takes. Count your own.

Colorado's rewrite shows why a US team should not hard-code state rules

Colorado's 2024 AI Act never took effect. SB 26-189 repeals and replaces it, and narrows the law to transparency and rights around consequential automated decisions. The Attorney General has not finally adopted the implementing rules, and a federal court order in xAI v. Weiser holds back enforcement, so when enforcement starts is open.34,35 Keep rules like this in a library you can change, each with its status and date.

Who touches a new AI idea today: sponsor, AI office, IT Quality, privacy, security and the board

In shortA long form, then separate reviews, then a committee. No public study gives the time or the cost, so the split below is an illustration to replace. The hand-offs and the waiting are the part most teams recognize.

The six roles around one AI idea, and what each one cares about

Form and spreadsheet, a risk module, or a dedicated platform: how teams build intake

From "sets the class" to inventory entry REG-0413: a deviation assistant's whole classification path

In shortDr. Maya Chen's deviation triage assistant, AIU-2041, is proposed as minimal risk under the AI Act and critical for GMP. The proposed tier reads "High risk, not allowed as designed", with two redesigns offered. Dana Okafor, who leads the AI governance office and chairs the review board, picks the one where a QA reviewer decides. The path lights again, she approves with five conditions and signs, and the use enters the inventory.

1

Five branches light up, each with its reason and clause

The left of the screen holds Maya's intake. Each answer carries a tag saying where it came from: "read from the description", "Answer 2", or "design §3.1" with a link to the passage. Her description says the assistant "sets" the class. The GxP branch reads that, finds a generative model on the model card, and lights red.

The AI Act branch stays green, minimal risk, because this use sits outside Article 5 and Annex III. The personal data branch turns amber: design section 3.1 adds analyst names, where the answer said none. The vendor branch turns amber because the no-training option is not signed. All five wire into one proposed tier, routed to Sam Patel (IT Quality), Lena Ortiz (privacy), Omar Haddad (security), Hannah Brooks (vendor terms) and Dana.

2

The contradiction shows up on day one, with its page

Opening the full use case puts the intake conversation beside the evidence. The answer "No personal data" is outlined in amber. To the right, page 4 of the design reads: "analyst name and reviewer initials". Section 2.3 is highlighted too: "The assistant sets the deviation class". Maya wrote "suggest" in one answer and "ideally it sets the class" in the next. The record keeps both, and nobody has to find this at the privacy review.

3

One redesign and the path lights again

Dana chooses "Suggestion only, a QA reviewer decides". The GxP branch moves from red to amber: non-critical, person decides, with the operator's duties written down. Model risk drops from high to medium, since a person decides on the full record while the consequence stays high. The proposed tier becomes Enhanced review: approvable, 5 conditions, 6 assessments.

The move to "non-critical" is the sample company's own reading, under its policy 6.5. The draft Annex does not say a person in the loop makes a critical use non-critical. That is why Sam Patel has to confirm this branch and why the decision is re-checked when the final text arrives.

4

Five conditions, and a signature that says what it means

Dana reads the five conditions: suggestion only; a qualified QA reviewer confirms or changes every suggestion, and the choice is recorded; names and initials removed before any text leaves the company; the no-training option signed and data kept in the EU region; a quarterly review of agreement and override rates, with a re-check on every model version change. Her signature records its meaning: "I approve this use case with the conditions above."

5

Into the inventory, with six assessments sent

The use becomes REG-0413 and the inventory moves from 412 to 413. Six assessments go to their owners with dates: an AI validation plan to Sam Patel by October 30, a security review to Omar Haddad by October 21, vendor terms to Hannah Brooks by October 16, the privacy note to Lena Ortiz by October 17, AI literacy training for 12 QA reviewers to Maya by November 6, and a monitoring plan to Sam before go-live.

Classifying an AI idea as it is typed: automated batch release, read from one sentence

In shortA new idea can be sorted while the sponsor is still writing. Sofia Marin, a QA operations manager, types one sentence. The path reads the process and who acts from her words and labels them as read. Once she picks a generative model, the GxP branch lights critical. The consequence of "sets" against "suggests" is visible before she finishes the form.

The AI register in daily use: use case list, risk map, rule library, dashboard and settings

In shortAround the path sit the screens an AI office uses every day: every idea sorted by what needs a decision, a map of the portfolio, the rule library with what cites each clause, the numbers, and the reviewers and routing rules your team can change.

Six specialist agents read and route, reviewers confirm and the AI review board decides

In shortEach agent has one job and written limits. None of them approves a use. Reviewers confirm their own branch. The board chair signs, and only a signed decision, or the narrow self-service class your board has defined, creates an inventory entry.

Intake interviewer to assessment planner: each agent's job and its limit

Who answers for what: board, reviewers, sponsor, auditors

Maker-checker separation (one person prepares, another approves), audit export and role-based access are part of the design. The platform is SOC 2 Type II audited and its information security is managed to ISO 27001. GxP, 21 CFR Part 11 (FDA's electronic records rule) and EU Annex 11 (the EU's annex on computerized systems) are matters for your validated use: the product offers controls that support your compliance, and your quality team validates its intended use.

The classifier is an AI use too, so put it in your own inventory. On the path we show, it supports a decision that people make and record. Your counsel and quality team decide how to classify it for your company.

From a typed idea, design documents and vendor terms to a cited tier and an inventory entry

In shortThe inputs are what an AI idea already comes with. The outputs are a classification with citations, a checklist of assessments, an approval record and an inventory entry. Outcome numbers below are design targets, the goals the live solution is measured against.

Systems an AI office connects it to, by kind

Document management and collaboration tools for designs; the quality management system for procedures and deviations; the contract system for vendor terms; identity and access systems for roles; ticketing and email for reminders; and your data warehouse for inventory exports. Connections follow your security review, and the solution handles the document volumes of a global enterprise.

What your AI review backlog and unwritten reasons cost, and how to measure both in a week

In shortTwo kinds of cost. The everyday one is staff hours and good ideas waiting, and you can measure it this week. The rare one is a GMP finding, a fine or a breach, and nobody can predict it for you. We found no public figure for the hours or days a pharma AI review takes, so your first number is your own.

Timing ten past AI ideas and comparing their forms with their designs: your baseline in a week

The sponsor who stops asking, the wrong precedent and other costs of AI review no calculator counts

What faster AI classification returns when you count reviewer hours, waiting days and re-checks

In shortPut in your own numbers. Every default is a placeholder. The calculator counts staff time and shows waiting days separately, and it takes away what software would cost you. It gives no value to avoiding a finding or a fine, because nobody can promise that.

Reviewer hours, waiting days, re-checks: where automating AI intake can pay

Your yearly reviewer hours back on AI ideas and rule re-checks

Which AI ideas to automate first: high volume and clear rules

Is your AI governance office ready to automate intake? Eight questions and a reflection

In shortAnswer honestly, 0 to 2 on each. Your score suggests whether to build a baseline, pilot on closed ideas, or plan a rollout. Nothing you choose leaves this page.

What to ask any AI governance platform about clauses, GMP, drafts and who decides

In shortAsk every vendor the same questions, us included. The answers show whether the tool reads your documents, whether it knows what is draft and what is law, what it does when it is unsure, and whether a person stays in charge.

Replaying closed AI ideas: a ten-week pilot with planted contradictions and a written decision

In shortPilot on closed ideas first, so nothing real depends on the tool. Plant known problems. Score it and your own reviewers side by side. Then run a few live ideas in parallel, and decide in writing.

Planted contradictions, clause checks and override rates: how to score an AI classification pilot

EU AI Act risk classification, draft Annex 22 and FDA's AI guidance: questions people ask

In shortShort answers to the questions people search. Each stands on its own. None of them is legal advice.