Before a medical device goes into a premarket submission, someone in regulatory affairs opens a long workbook of recognised consensus standards and decides for every row: is this potentially applicable to our device, why, who needs to hold the standard, and do we need to buy our own copy. The decisions are expert work and the recording is clerical work, and the two are interleaved for days.
We built an assistant that does the first pass. It is worth describing not because the assistant is clever but because of the three constraints the regulatory team placed on it, each of which would be good advice for anyone putting a model into expert review.
Show the tools
The assistant is a chat. A reviewer drops the device profile PDF into the conversation and talks to it. Behind the chat, the model works through tools: search the workbook by keyword, fetch one standard, list the task groups, read the device profile, correct the profile from the conversation, run the rules, start a screening run, fetch results, edit one row.
Every tool call appears as a chip under the answer. The reviewer can see that the assistant searched for “sterilization” and fetched three standards before answering, or that it did not. This is the single feature that turned the team from sceptical to willing. An assistant that explains itself in prose is persuasive; an assistant that shows what it looked at is checkable.
A second opinion from rules
The team’s conventions were already partly explicit. Radiation-dosimetry standards are part of the sterilisation service and depend on the method. Cleanroom and packaging-seal standards belong to the contract manufacturer. The biocompatibility series goes to the biological test house. Material specifications for the sponsor’s own parts stay with the sponsor. Radio-frequency immunity guides apply to every powered device.
We wrote these as a deterministic screener that runs beside the model. For any standard the reviewer names, the rules give their verdict and the model gives its own, and the two are shown together. Where they agree, the reviewer moves on. Where they disagree, the reviewer has found either a gap in the rules or a mistake in the model, and both are worth knowing. With no API key at all, the screening run uses the rules alone, which means the tool is never entirely unavailable.
Give the workbook back as you found it
The output is not a report. It is the same workbook, with the four screening columns filled for every row. The Excel table, its style, the column widths and the summary block of count formulas under the table are preserved, so the counts by “needed” label keep working. The team’s downstream process, which reads that workbook, did not change.
This sounds like a small thing and was the hardest engineering in the project. Spreadsheet libraries are happy to read a workbook and unhappy to write one back without disturbing it. It mattered because the alternative, a new file the team has to reconcile with the old one, is a new piece of clerical work replacing the one we removed.
What a run does
The assistant reads the PDF and builds a structured device profile: energy source, laser or ionising or lamp, patient contact and its duration, blood contact, implantability, materials, sterility and method, software, and which work is contracted out to whom. The reviewer corrects it in conversation (“we do EMC testing in-house”) and the profile updates.
Then, on request, the assistant screens every standard in the background and fills the four columns: potentially applicable (yes, no, or a question mark), a one-sentence reviewer-style comment naming the component, material or process and who owns it, who needs to hold the standard (the sponsor, nobody, maybe, the system level, or the named contract party), and whether the sponsor must purchase its own copy because it will claim conformance in the submission.
The reviewer filters the results by task group, by label, by keyword, by rows the model was unsure about, and edits any row. Edits are kept, and the final workbook is written out.
The result
An afternoon instead of days for the first pass, and the expert’s time concentrated on the rows marked “maybe” and on the rows where the rules and the model disagree. The team did not stop reviewing. They stopped typing.
The general pattern is one we recommend wherever a model enters an expert workflow: make the model show its sources, run a rule-based opinion beside it, and never change the file the experts already use.
About the authors. Neugent Consulting designs and builds AI systems for operators. The examples in this piece come from client work; clients are anonymised and details are simplified.
