02The barriers AI lowers
Some software stays out of reach of a small team, however sound the idea, because of how much must be built before it can offer anything new. Generative AI, with a method around it, can lower that barrier a long way. It leaves the others where they were.
Tools that let people lay out screens on a grid, drop fields into them and decide which are read-only, editable or mandatory were once products in their own right, built and maintained by whole companies. The product in this article needed one: a page composer with which a hospital administrator lays out any page of the application and decides who may do what on it. None of it is complex, but every interaction, down to a simple drag and drop, takes a great deal of work to build and as much again to test. With Claude and BMAD, a first working version took about three weeks.
Without generative AI and a method to steer it, I would not have contemplated building this product at all. This chapter is about what made the difference, and what did not change.
What kept such work out of reach?
Its volume, far more than its difficulty. A new system must first do what its users already expect, or nobody will consider it, and only then can it offer something more. In a field where hospitals have bought software for decades, users expect a great deal. Many organisations know the same situation from the inside, with an in-house system judged too big to replace because of everything built into it over the years.
That first front is mostly well-understood work, clearly specified and often repetitive, expensive because there is so much of it. Until now it meant years of a team's work before anything new could be shown, and many attempts ran out of money or patience on the way.
What did AI make possible?
The volume became affordable
Well-specified, repetitive work is what an AI does best. Held to clear acceptance criteria, reviewed and tested before anything is accepted (chapter 04), it can bring that work within reach of one person: here, a first working page composer in about three weeks, and 172 stories in about 24 days of recorded time, and somewhat more in reality (chapter 01). Readers who have built such things can judge what a team would have needed. For those who want a figure, Appendix A estimates it for the configuration tool alone, with its method stated: on figures chosen to favour manual work, twelve to fifteen times the effort for that one component. What one person no longer spends on that volume can go to the part that makes a product worth having. Domain-Driven Design asks where to invest the best effort, and Drawing the Lines gives its answer: in the core subdomain, where a product differentiates itself, while supporting subdomains are necessary but not distinctive and generic ones are bought or reused. AI makes the supporting work cheap to build, which leaves more attention for the core. Here the core is the page composer and the configuration it produces, as chapter 08 sets out.
Questions could be answered by building
The product rests on four design questions: how to make a highly configurable medical device safe, how modular it should be, which functions are indispensable, and how to integrate with hospital systems that never look alike. A Kano analysis, which sorts the features users take for granted from the ones that make a difference, answered the third on paper, as Agile Planning describes for ordering a backlog. The others needed something built, and with Claude an option could be built, tested and thrown away if it was wrong.
One person could reach something worth showing
A project needs a product owner. Here one person could hold that role and still put a working product in front of the medical staff who will use it, which is the only way to learn what they need.
These are possibilities, and none of them comes guaranteed. Each held only as long as the conditions around it held, and chapter 07 shows what happened when one did not.
What does it leave unchanged?
Deciding what to build stayed with experience: of the domain, where good practice is documented well enough for a first set of requirements, and of software, to turn expected capabilities into epics and stories and to judge whether the result was right. The daily experience of the medical staff still has to come from them. And the barriers outside the code are untouched: hospitals buy on references, certification, support and the cost of switching. Whether the product is worth building remains a thesis until the numbers exist, and such a case is built step by step, each stage funded once the previous one has answered its question (chapter 10).
The risk that comes with cheap building
Cheap building brings a risk of its own. When a feature costs hours, features pile up quickly, each tested on its own, and nobody finds out until late whether they work together. The page composer was an example: three weeks in, it worked well on its own terms and had never fed a real application. Scrum already has a practice for this risk, and chapter 03 shows how it was applied.
For your own project
- Revisit the "too big to replace" verdicts and the build-or-buy decisions made when building was expensive. Some rested on a cost that has since fallen.
- Separate the features users take for granted from the ones that make a difference, and estimate each again at today's cost of building.
- Treat each saving as conditional: write the conditions down, and check that they still hold.