Skip to content

AI Agents for Medical Device Makers: Documentation, Quality and MDR

For most medical device manufacturers, the bottleneck is not the production line but the paperwork behind it. The EU Medical Device Regulation (MDR), and in the US the FDA's quality system rules, require living technical files, post-market surveillance, safety reports and incident reporting against hard deadlines. AI agents can draft and check much of that text. Once they do, they become part of the quality management system and must be treated like any other QMS software.

What are AI agents in medical device manufacturing used for?

An AI agent is a language model that can use tools. It reads documents from your document control system, searches complaint records or literature databases, compares content against a requirement and files a draft with references. For a device maker it is a regulatory affairs and quality tool, not part of the product.

That distinction matters more than anything else in this topic. Article 2 of the MDR defines a medical device by the medical purpose its manufacturer intends: diagnosis, monitoring, treatment and so on. An agent that checks the chapters of a technical file for gaps has no such purpose. It is software used in the quality system, so the rules for QMS software apply, not the rules for software as a medical device.

Typical starting points:

  • Technical documentation: find gaps against Annex II, check cross-references, flag sections affected by a design change.
  • Complaints and vigilance: structure incoming complaints, flag possible serious incidents, make deadlines visible.
  • Post-market surveillance: run literature and database searches, summarise hits, prepare sections of the periodic safety update report.
  • QMS upkeep: compare procedures against revised standards, check training records and CAPA entries for completeness.

For a general view of how such agents are built and which departments they suit, see the overview of AI agents for business teams.

Why is documentation the real bottleneck?

The MDR makes the manufacturer responsible for a complete technical file that stays current. Every design change, revised standard and finding from the field has to flow back into it. That workload grows with the number of products and variants, not with production volume.

What the regulation actually says:

  • Article 10(4) requires manufacturers to "draw up and keep up to date" technical documentation containing the elements of Annexes II and III.
  • Article 10(8) requires it to remain available for at least 10 years after the last device has been placed on the market, and at least 15 years for implantable devices.
  • Annex II asks for documentation presented in a "clear, organised, readily searchable and unambiguous manner", with cross-references to the exact location of evidence for each applicable safety and performance requirement.
  • Article 86 requires a periodic safety update report (PSUR) for class IIb and III devices updated at least annually, and for class IIa devices when necessary and at least every two years.

This is reading, matching, summarising and cross-referencing work, which plays to the strengths of language models. It is also work where an error surfaces in a notified body audit. So the rule in this sector is stricter than elsewhere: the agent prepares, a qualified person approves (human-in-the-loop), and both steps leave a record. For automotive suppliers, the focus is on sampling and the requirements of vehicle makers; device makers add statutory reporting deadlines towards authorities and a file that must be maintained for the whole product lifetime.

How does an agent work in a regulated environment?

A medical device agent uses the same building blocks as any other agent: a model, tools, instructions and memory. What it adds is a traceability layer. Every output must be traceable to the document versions it read, the instruction it followed and the model version that produced it.

The building blocks:

  1. Model: a language model accessed through a provider's API or run on your own hardware. If confidential design data must not leave the company, self-hosting is an option; the trade-offs are covered on the page about sovereign AI.
  2. Tools: read access to document control, the complaint system and the requirements checklist; search access to literature databases. Write access only to a draft area, never directly to controlled documents.
  3. Instructions and criteria: the requirements the agent checks against, kept as versioned files, for example an Annex II checklist or the criteria from your own vigilance procedure.
  4. Run log: for each run, the input documents and their versions, the instruction version, the model, the output and the reviewer's decision.

The fourth block is what separates a usable regulatory agent from a chat window. Without it, you cannot show an auditor what a draft was based on or who reviewed it.

Which use cases are there?

Four use cases cover most regulatory text work. They differ mainly in how time-critical they are and which data the agent needs. Each will get its own in-depth article in this section; this overview explains how they fit together.

MDR technical documentation. The agent checks an existing file against the structure of Annexes II and III and reports missing sections, outdated references to standards and cross-references that point nowhere. After a product change, it lists the chapters that need updating. This is the lowest-risk entry point because no deadline is running and every output passes through document control anyway.

Complaint handling and vigilance. Here time matters. Under Article 87, a serious incident must be reported no later than 15 days after the manufacturer becomes aware of it, no later than 10 days in the case of death or an unanticipated serious deterioration in health, and no later than 2 days for a serious public health threat. An agent can structure incoming complaints and immediately flag those that may meet the definition of a serious incident in Article 2(65). Article 87(7) says that a manufacturer who is unsure whether an incident is reportable must still report within the deadline, so an agent should only ever escalate, never clear.

Post-market surveillance. Article 83 calls for a system that gathers and analyses data on quality, performance and safety "actively and systematically" across the device lifetime. That includes literature searches, public incident databases and similar devices on the market. An agent can run searches to a fixed protocol, summarise the hits and prepare sections for the PSUR. Article 88 adds trend reporting for statistically significant increases in non-serious incidents; the method sits in your post-market surveillance plan, and an agent can apply it on a schedule and flag anomalies.

QMS validation. Every agent above is software used in the QMS and has to be validated. A companion checklist explains how to validate AI tools in an ISO 13485 quality system. A separate overview in the department section covers AI in quality management across industries, such as complaints, audits and inspection planning; this article focuses on what the MDR and ISO 13485 change.

What data and systems do the agents need?

Medical device agents depend on controlled, versioned documents and structured complaint records. If those are missing, start there: an agent can only be as traceable as the sources it reads from.

In practice:

  • Document control: technical file, risk management file, clinical evaluation, instructions for use and labelling, each with a clear version and approval status. Whether this lives in an eQMS, a PLM system or versioned SharePoint libraries matters less than consistent versioning.
  • Requirements checklist: the list of applicable general safety and performance requirements with references to evidence, as Annex II section 4 requires. It is the reference point for many agent tasks.
  • Complaint records: complaints from service, sales and support with date received, product, lot or UDI and free text. Without a date received, reporting deadlines cannot be tracked.
  • Literature and database access: the sources named in your post-market surveillance plan, plus public incident databases.
  • Procedures: your own SOPs for document control, vigilance and post-market surveillance. The agent checks against these, not against general knowledge.

In our own operations, preparing this data foundation has taken more effort than building the agent. How agents work across scattered document stores is covered on the page about knowledge management agents.

What does regulation require of the AI itself?

Two frameworks apply: ISO 13485 for validating QMS software, and the EU AI Act. For internal drafting and research agents, ISO 13485 is the one that shapes daily work. The AI Act matters mostly once AI becomes part of the device.

ISO 13485 and the FDA. Clause 4.1.6 of ISO 13485:2016, as quoted in industry guidance, requires documented procedures for validating the application of computer software used in the QMS, before initial use and, as appropriate, after changes to the software or its application. The approach must be proportionate to the risk of use, and records must be kept. For an agent, that means defining its purpose and limits, running test cases with known correct answers and recording the results. Changing the model version or the instructions counts as a change. For US-market manufacturers, the FDA's Quality Management System Regulation took effect on 2 February 2026 and incorporates ISO 13485:2016 by reference, so one validation approach can serve both markets.

EU AI Act. Under Article 6(1), an AI system is high-risk if it is a safety component of a product, or is itself a product, covered by the legislation listed in Annex I and subject to third-party conformity assessment. The MDR is on that list. An internal tool that drafts text for your own file is neither. According to the amended Article 113 as shown in the AI Act Explorer in October 2026, high-risk obligations under Annex I apply from 2 August 2028. Article 4, by contrast, has applied since 2 February 2025: deployers must take measures to ensure a sufficient level of AI literacy among the staff who use these systems.

This section is orientation, not legal advice. Whether a specific tool is a medical device or high-risk AI is a call for your regulatory affairs team to make and document.

How do you roll it out step by step?

The safe path runs from a narrow, non-deadline process to the time-critical ones. Each step ends with a validated, logged application before the next one starts.

  1. Pick and scope one process. One process, one product or family, one written requirement. Checking a technical file against Annex II works well because the output is a list a reviewer can assess quickly.
  2. Record the classification. Briefly justify why the tool is not a medical device and what risk its use poses within the QMS.
  3. Set up data access. Read access to controlled documents with version information, a draft area for outputs, a log for every run.
  4. Build test cases. Plant known gaps in a copy of the file and check the agent finds them; include clean files to measure false alarms.
  5. Validate and release. Record test results, define limits of use and release the tool under your own procedure.
  6. Monitor in use. Have specialists sample outputs regularly; rerun the test cases whenever the model or instructions change.
  7. Extend. Only then move on to complaint triage or post-market research, using the same structure.

What are the limits and risks?

Language models can produce plausible but wrong statements (hallucinations), attach the wrong source and give different answers to the same input. In a regulated setting that is a compliance risk, not an inconvenience. Agents therefore need to be designed so that errors surface before anything reaches a controlled document.

Specific risks and countermeasures:

  • Invented content: an agent that fills in missing evidence rather than reporting it is dangerous. Allow it to reference only existing documents, with a location for every claim.
  • False reassurance in vigilance: a missed serious incident can mean a missed deadline. The agent only escalates; a person still assesses every complaint.
  • Silent model updates: providers update models. Pin the model version where the provider allows it and revalidate on change.
  • Confidentiality: design data and complaints containing health data leave the building with cloud models. Check your data processing terms, minimise data or run models on your own infrastructure; this makes the setup one that can be operated in line with the GDPR, but does not replace your own assessment.
  • False confidence: if reviewers trust drafts blindly, the agent shifts responsibility instead of removing work. Use sampling, AI literacy training under Article 4 of the AI Act and clear roles.

When it does not fit: a small portfolio with rare changes, no working document control, or nobody with time to review outputs. In those cases, a cleaner file structure helps more than an agent.

How do you get started?

Start with an inventory: which regulatory texts does your company produce regularly, how much time goes into them, and where does the source data live? That almost always points to a first process with plenty of text, clear criteria and no running deadline.

Three questions help you choose:

  • Is there a written requirement to check against (an annex, an SOP, a checklist)?
  • Are the source documents versioned and readable by a tool?
  • Is there someone qualified to approve outputs who has time to do it?

If all three answers are yes, the process is a candidate. To start with regulatory affairs, quality and IT in the same room, a joint AI agent workshop helps to rank processes and assess data and validation effort. How department-level agents are built is described in the overview of AI agents.

Legal status as of October 2026.

Frequently asked questions

Is an internal documentation agent itself a medical device?

Usually not. Under Article 2 of the EU MDR, what counts is the medical purpose the manufacturer intends, and a tool that drafts text for your own technical file has none. Record that reasoning anyway.

Can an agent decide whether an incident is reportable?

No. It can sort complaints and surface deadlines, but the reporting decision under Article 87 stays with a responsible person in your vigilance process.

Does a cloud language model need validating under ISO 13485?

You validate its use in your QMS process, not the model in the abstract. Effort scales with risk, and a model or prompt change counts as a change that triggers re-assessment.

How does this relate to the FDA?

Since 2 February 2026 the FDA's Quality Management System Regulation incorporates ISO 13485:2016 by reference, so the same software validation logic applies to US-market QMS tools.

Does the EU AI Act apply?

The AI literacy duty in Article 4 already applies to anyone deploying AI. High-risk status under Annex I mainly concerns AI that is a safety component of a device, which an internal drafting tool normally is not.

Sources

  1. Regulation (EU) 2017/745 (MDR), Article 10 – General obligations of manufacturers (as adopted) legislation.gov.uk
  2. EU MDR, Annex II – Technical documentation legislation.gov.uk
  3. EU MDR, Article 2 – Definitions legislation.gov.uk
  4. EU MDR, Article 83 – Post-market surveillance system legislation.gov.uk
  5. EU MDR, Article 86 – Periodic safety update report legislation.gov.uk
  6. EU MDR, Article 87 – Reporting of serious incidents legislation.gov.uk
  7. EU MDR, Article 88 – Trend reporting legislation.gov.uk
  8. U.S. FDA: Quality Management System Regulation (QMSR) fda.gov
  9. Greenlight Guru: Software validation under ISO 13485:2016, clause 4.1.6 greenlight.guru
  10. AI Act Explorer, Article 6 – Classification rules for high-risk AI systems artificialintelligenceact.eu
  11. AI Act Explorer, Article 113 – Entry into force and application artificialintelligenceact.eu
  12. AI Act Explorer, Annex I – List of Union harmonisation legislation artificialintelligenceact.eu
  13. AI Act Explorer, Article 4 – AI literacy artificialintelligenceact.eu