
Key takeaway
An auditor pulled one requirement and every document was signed off, yet the trail between them was broken. The story we heard on our stage, why AI on the reviewer's side makes such gaps easier to find, and how to read your own file first.
Short answer: You keep it audit-ready by making sure every requirement can be followed to its test and its result before anyone else looks. At the 2nd MedTech Meets AI Conference, Utkarsh Chopra, Director R&D, Innovation Realization at Philips, told the story of an auditor who pulled a single requirement and asked how it had been tested. Every document was signed off, yet the documents did not agree with each other, and it took a war room to rebuild the trail. Gaps like that used to surface only when a reviewer happened to sample the right page. Regulators and notified bodies are now equipping themselves with AI that can read the whole file, so the manufacturer has to read it first.
Chopra's keynote in Barcelona in October 2026 was titled "Auditable AI: Finding the Gaps in Your Evidence Before the Auditor Does". He opened with a question from antiquity: "Plutarch's question: is a ship rebuilt plank by plank still the same ship?" His point was that a medical device which has been on the market for years is that ship. Suppliers change, software is updated, standards are revised, labels are rewritten and production moves from one site to another. The product is still sold under the same name, but each replaced plank brought its own evidence with it, and that evidence was filed wherever the team responsible happened to keep it. One of his slides put the consequence in six words: "After launch, the evidence starts to age".
He then told what happened when this met an audit, without naming the business. An auditor asked about a platform that had been on the market for years and pulled one requirement: "Show me how you tested this, and what happened." In his words: "Not next week. Now." The evidence existed, but it sat in four systems and in people's heads, and nobody could read across them quickly enough. So the auditor did it for them, one thread at a time, and the result was a finding, a corrective action and a war room in which people rebuilt the trail by hand. The finding was eventually closed. What stayed with him was a different lesson: "a trail rebuilt by hand starts to age the day you finish it."
To show how this happens to careful teams, he gave the room a quiz with four illustrative pages. The requirement, kept in the product lifecycle system, says the alarm must sound within 2 seconds. The test protocol, kept in the document system, says the device passes if the alarm sounds within 5 seconds. The test result, on a lab drive, records an alarm after 4.1 seconds and a pass. The verification report, on a team site, states that all requirements are verified. All four pages are signed off.
"Every page is right on its own. The trail breaks between them."
Utkarsh Chopra, Director R&D, Innovation Realization, Philips, 2nd MedTech Meets AI Conference, October 2026
The requirement asks for 2 seconds and the protocol allows 5, so a result of 4.1 seconds passes the test, fails the requirement, and still ends up in a report that says verified. Nobody made a mistake that a review of any single document would catch. This, he argued, is the harder half of the problem. People are good at noticing what is missing. They are much less good at noticing two pages that each look right and contradict each other, especially when the pages live in different systems and the product keeps changing.
Regulators, he went on, find exactly these breaks. He showed four phrases from three FDA warning letters sent in 2025, one for each link in the chain. On the requirement: "Your design inputs do not clearly define all of these requirements" (March 2025). On the protocol: "did not test product sterility as part of the shelf-life validations" (April 2025). On the result: "documented the results as pass/fail, rather than documenting the measured results" (March 2025). And on change: "These actions were taken outside of the CAPA system" (September 2025). His reading of them was short: "None of these says the device failed. All four are about the evidence."
We checked the four quotes against the public letters, and the wider numbers point the same way. Covington counted 59 device warning letters in calendar 2025 against 45 in 2024, and of the 38 that cited the quality system regulation, 26 cited design controls and 26 corrective and preventive action. In FDA's inspection data for fiscal 2025, corrective and preventive action was the most-cited device provision (GMP Insiders, January 2026, trade analysis of FDA data).
Up to this point the story could have been told ten years ago. What makes it urgent, in Chopra's argument, is that review has always been rationed by the time a human reviewer has, and that this limit is disappearing: "AI now sits on both sides of the table", and "The sample is becoming a census." How far is that true today?
In the United States it is well advanced. The FDA launched Elsa, a generative AI tool, for all staff in June 2025, and listed the identification of high-priority inspection targets among its uses. In December 2025 it added agentic AI for pre-market reviews, post-market surveillance, and inspections and compliance, and Elsa 4.0 followed in May 2026. The agency has not said how far these tools are used in device reviews specifically, so the census is a direction of travel and not yet a documented fact. The same period brought a new inspection regime: the US quality rule took effect in February 2026 with a risk-based programme under which management reviews, internal audits and supplier audit reports can be inspected (Ropes & Gray, February 2026, law firm).
Europe is earlier in the same movement. In May 2026, TÜV SÜD described an internal system that runs completeness checks, compares text blocks for consistency and checks references to standards, and said that machine-to-machine exchange with manufacturers' own AI is being considered (RAPS, May 2026). In September 2026 the notified bodies' association launched DIGICAP, a three-year EU-funded project to build structured, machine-readable conformity assessment data, and the proposed revision of the device regulations would allow structured technical documentation (Osborne Clarke, January 2026, law firm).
What these machines will find is often weak, and the notified bodies say so themselves. Chopra cited their figures, and they hold up with one nuance. In Team-NB's 2025 survey of 41 notified bodies, 65% of responses put MDR submissions at half complete or less, and 48% said a new certificate takes 13 to 18 months (Team-NB, May 2026). These are shares of notified body responses, not of all files. The European Commission's survey of all 53 notified bodies gives a similar picture: only 14 of 52 said that more than half of MDR submissions are complete (European Commission, 2026). Team-NB's guidance on technical documentation (April 2026) asks for precisely the links his quiz showed breaking: verification reports tied to risk IDs and a matrix of design inputs and outputs.
The team's first answer to the war room was to stop waiting for the audit. They now run product quality reviews before the auditor arrives, reading the file for design controls, supplier controls and regulatory compliance. But a review is still people reading, and the product keeps changing underneath them, which is why he concluded: "A war room rebuilds the trail once. Keeping it true needs a machine."
He showed how his own department already uses AI for the reading. Before a stage gate, an agent reads the project file against the master list of deliverables and reports what is missing, so that the project director walks in prepared. He set two rules for that system which would serve any tool: every answer shows its source, its owner and how fresh it is, and when the system cannot ground an answer, it declines. The division of labour is deliberate: "The agents do the reading and the paperwork. The directors keep the judgement."
Most of the industry is not there yet. In a Deloitte survey of 100 MedTech leaders (July 2026), 45% reported agentic AI in production, but only 7% of those called it fully embedded and audit-ready, and one in three said AI outputs are used informally, without updated procedures or validation. In a vendor benchmark of 130 regulatory professionals, only 17% rated their regulatory data quality as excellent for AI (Veeva MedTech, July 2025, vendor). An AI that reads a file built on inconsistent records will reproduce the inconsistency, so the data has to be put in order first.
His closing instructions bring the argument back to practice: "Assume a machine will read every page you submit." "Make every requirement find its result before you file." And run the reviewer's read on your own file first, because "Your file will be read by a machine either way. The question is whose machine reads it first." In day-to-day terms, that means tracing each requirement to its protocol, its measured result and its report by identifier, with acceptance criteria that match; recording measurements and not only verdicts, since a pass with no number behind it cannot be checked against the requirement; and re-checking the trail after every change, because each change is a plank replaced.
For the organisational side of getting AI out of pilots, see Why MedTech AI Fails to Scale.
The 2nd MedTech Meets AI Conference, organised by UBIQ Events, took place in Barcelona on 14–15 October 2026. Dates for the next edition will be announced on the conference website.
The FDA has given generative and agentic AI tools to all staff and names pre-market reviews, post-market surveillance and inspections among their uses. It has not published how far they are used in device reviews specifically.
Elsa is the FDA's generative AI tool, launched agency-wide on 2 June 2025. Version 4.0, released in May 2026, added customisable agents, document creation and data analysis. The FDA says it is not trained on data submitted by industry.
Some are starting to. TÜV SÜD described an internal tool in 2026 that checks completeness and consistency of technical documentation, and a three-year EU project led by Team-NB began in September 2026 to make conformity assessment data machine-readable.
A break in the chain from requirement to test protocol to result to report. Each document can be correct on its own while the chain fails, for example when a protocol's acceptance criterion is looser than the requirement it is meant to verify.
Often not very. In a 2025 survey of 41 notified bodies, 65% of responses put submissions at half complete or less, and the most common time to a new MDR certificate was 13 to 18 months.
From 2 August 2028, after an amendment adopted in 2026 moved the date from 2 August 2027.
The MedTech Meets AI Conference, organised by UBIQ Events, brings together R&D, quality, regulatory and digital leaders from medical device companies across Europe.