Perspectives, engineering notes and field reports on what it takes to make AI software delivery something you can agree on, build once, and prove later.
The industry has accepted that AI-generated code broke human review — the numbers are brutal, and the argument has moved on to what replaces it: more human scrutiny, or verified delivery pipelines. It's a real debate. But both answers inspect the code after it exists, and neither asks the question that actually determines whether the pull request should have been written at all.
A company commissions a €200k application. Two years later it needs changing. The original team is gone, the spec is stale, and nobody remembers why the architecture is the way it is. So a new supplier spends weeks rediscovering a system the company already paid to have understood once. That second bill is avoidable — and it's bigger than anyone measures.
If you run a software house, your margin doesn't leak where you think it does. It leaks before a single line is written — in the estimate you rushed, the scope you left fuzzy, and the change request you never charged for. AI writes the code faster now. It does nothing for any of that.
For fifty years the hard part was writing software. Cursor, Copilot and Claude Code solved that — and every tool we built now points at a solved problem.
Models assume cheaper code means cheaper software. Only the first half is true. Cost didn't vanish — it moved from construction to coordination.
We review every pull request and never review the spec it implements. When AI writes flawless code for the wrong requirement, code review approves it.
We hunt debt in the code. The most expensive debt is in the decisions nobody wrote down — and AI-generated code just raised the interest rate on all of it.
Every AI-delivery pitch leads with speed because speed demos well. Boards don't optimise for velocity — they optimise for the absence of nasty discoveries.
Speed doesn't make ungoverned delivery slightly riskier — it changes the maths of failure. Slow construction caught disagreement. AI removed the slowness.
Everyone is lowering cost per token. Wrong number. The expensive tokens are spent building the wrong thing, and a cheaper model won't save you from that.
AI isn't replacing developers. It moved the work upstream, to the hardest part: getting an organisation to agree what to build before code is written.
Your tools track activity — commits, tickets, pages. None track commitment. In spec-driven development the agreement, not the code, is the real artifact.
The lone-developer myth was always a myth. Enterprise software is built by organisations agreeing — and now AI writes the code, agreement is the hard part.
Talk to our team about bringing Propel to your organization.