01The claim, and what it is not
One engineer, part-time, twelve weeks: software developed under the process for the highest safety class of medical device software, with the evidence a regulator would ask for. What that proves, and what it does not.
Claims about AI in software development tend to come in two kinds: a demo built in an afternoon, or a company saying that a large share of its code is now written by AI. Neither tells you whether that code can be trusted where an error has consequences. Mine is narrower, and it comes with numbers.
The claim
Since early August 2026, working alone and part-time, I have been building a clinical software product for hospitals with Claude and the BMAD method. Part of it is a medical device, and the whole product is developed under the process IEC 62304 sets for Class C, the class for software whose failure could cause serious injury or death. By the project's own records, my working time came to about 24 days of eight hours, between 16 and 29 depending on how the records are read, and somewhat more in reality, since they see only the time spent working with Claude Code (Appendix A). On this project the team was one person; wherever a lesson applies more widely, I write "the team".
172
Stories delivered through the same chain: requirement, story, implementation, code review, tests, Definition of Done. For 23 of them, the story records no code review; I cannot say whether one took place, so they count as not reviewed (recorded in the quality system as open item OPEN-114).
55,000
Lines of Java, comments included, more than half of them tests: about 24,000 lines of product code and 31,000 lines of test code, covered by about 1,650 automated tests.
Nearly 900
Findings raised by code review and recorded in the stories, each fixed, deferred with a reason, dismissed with a reason, or escalated to me.
€20
A month: a Claude Pro subscription, with Claude Opus 5.5 and Claude Sonnet 5 behind all of it. BMAD itself is open source.
Figure 1. The numbers at a glance.
Alongside the code sits the documentation a regulated product needs, written as the work happened: three product requirement documents, three architectures with their review rounds, a risk register of about thirty hazards, around a hundred open items, each dated and traced, and 62 verification logs recording what was checked, on which build, and what deviated.
The clinical application, the medical device in the strict sense, accounts so far for 13 stories and about 5,900 lines, and its first slice deliberately leaves out the clinical safety controls (chapter 03). Most of the work went into what it stands on: the configuration tool, with 113 of the 172 stories and about 14,500 of the 24,000 lines of product code, and access control. They are the product's load-bearing walls, and the least visible part of it. I designed both to sit outside the device (chapter 08), but that classification is for the notified body to decide, and every story, inside the device or not, went through the same chain under the same process.
So the claim is this: with a method around it and a quality system behind it, generative AI makes it possible to develop a product under a Class C process, with the design-control evidence that process demands, from a one-person team: the cost is that person's time, and an AI subscription. It is a possibility shown once, under conditions, and the rest of the article is about those conditions: the practices that made use of it, and what stayed with people.
I will not claim a measured saving against a traditional team. I have no measured baseline, and GenAI Beyond the Hype is clear about results without one. The figures above are measured, and readers who have run regulated projects can compare them with their own. For the configuration tool, the largest part of the product with 113 of the 172 stories, Appendix A offers an estimate instead, labelled as one, of what manual work would have taken, with the method stated so that it can be checked.
Then there is the cost. The subscription was Claude Pro, €20 a month, throughout, and it came with limits. Claude Code keeps its session logs for thirty days, so what survives covers thirteen working days between 25 September and 26 October, all with Claude Opus 5.5. In the six days from 25 to 30 September alone, I hit the session limit seven times and the weekly limit once, which locked me out until 4 October. Priced at the API's list rates, those thirteen days come to about $310 to $380 of tokens, more than ten times the month's subscription. A flat subscription hides the real consumption, and billing is moving towards usage (GenAI Beyond the Hype). The business case does not rest on the flat fee: even at list price, those thirteen days of AI cost less than one day of a senior developer.
Claude wrote fewer than two million tokens in that time; 97% of the volume was context it read again from the cache, and most of my own time went to reading what it wrote. I use AI without deferring to it, and I try to be frugal with it before its price catches up.
What the logs record, 25 September to 26 October 2026
| Tokens | At API list price | |
|---|---|---|
| Written by Claude (output) | 1.95 million | about $39 |
| Context written to the cache | 22.2 million | about $111 to $177, depending on how long it was cached |
| Context read back from the cache | 818 million | about $164 |
| Fresh input | about 6,000 | under $0.10 |
| Total | about $310 to $380 |
List prices for Claude Opus 5.5 as of October 2026: $4 per million input tokens, $20 per million output tokens, $0.20 per million tokens read from the cache. The counts include the review agents that Claude Code started on its own. Most stories before 25 September were developed with Claude Sonnet 5, at half those prices; the logs of that period no longer exist.
What it is not
The claim is narrow on purpose, because readers who work in regulated industries will test it, and they should. The work is in progress, and this article reports it from a gate (chapter 10).
The product is not certified
No notified body or regulator has reviewed it, and nothing has been released.
It is not clinically validated
No clinician has used it in a ward, and the practitioners' day-to-day pain points have yet to be investigated in a structured way.
The quality system is not complete
Parts of the software development file required by IEC 62304 are still templates. Chapter 07 tells how one of its main mechanisms did not work for ten weeks without anyone noticing.
It has not been independently reviewed
One person wrote and approved every story, and is the only person who checked any of them: for 23 stories no review is recorded, and for 51 no check by a person (OPEN-114). The three AI reviewers of chapter 04 found real defects, but one model playing several roles is not the independent review a quality system requires before release, and that review is still to come. Twice, Claude also sorted the findings about its own work: the surviving mutants of Appendix B and the sample of review findings in Appendix D. My check of a sample of each is pending.
The business case is not proven
I have a thesis about what a new product can bring to a mature market (chapter 02), and I am building the case step by step, the way Beyond the Demo sets out.
AI did not take the helm
What the product must do, which constraints it meets and the fundamental architecture choices stayed with the team, and the article shows several places where letting the AI decide went wrong.
Why it still matters
A regulator would not accept any of this as a finished device, and would expect to find this kind of evidence on the way to one: requirements traced to tests, risks traced to controls, decisions recorded with their reasons, deviations logged as they occurred. Producing that evidence has always been the slow, expensive part of regulated software. What I want to show is that GenAI, shepherded by a method and held in check by a quality system, can produce it alongside the code, at a pace a single engineer can sustain. The last chapter turns to what that means for people, and for the leaders who will have to manage it.
For your own project
- When you report what AI did for a project, separate what was measured from what was estimated. If there is no baseline, give the figures and say so.
- Count the evidence as well as the code: requirements traced to tests, risks, decision records, verification logs. In regulated work, that is where AI's contribution is easiest to overlook.
- State what the result is not: certified, validated, proven as a business. Readers who work under audit will look for those limits, and they should find them stated by you.