EffektAI, the platform we sell to clients, was built using the same method we sell alongside it. Nine weeks, one small team, and an AI that wrote most of the code. Here is what actually happened, including the parts that went wrong.
Key takeaways
- AI wrote 71 per cent of the shipped codebase by line count — the remaining 29 per cent was architecture, contract logic, and the quality checks themselves.
- Our own quality rig caught four contract math regressions before they reached a client sandbox — all four would have silently over-billed on outcome milestones.
- The bottleneck was never generation speed. It was review throughput — which is why the rig, not the model, is the product.
- We shipped in nine weeks with three people. The same scope, built the old way, was quoted at seven months by two of our own principals before we started.
Why we built it on ourselves first
Every vendor claims their AI delivery method is fast and safe. Most of that claim rests on a demo, not a production system with real money moving through it. We did not want to sell a method we had not stress-tested on something that would embarrass us if it broke. EffektAI needed to calculate outcome-based invoices, track milestone acceptance, and reconcile scope changes against a signed contract. Get that wrong and someone gets billed incorrectly. That is not a forgiving place to find out your process has a hole in it.
The 71 per cent number, and what it hides
Seventy-one per cent of shipped lines were AI-generated. That number gets quoted a lot in vendor pitches, usually to imply the work is nearly automatic. It is not the useful number. The useful number is where the remaining 29 per cent went. Almost none of it was boilerplate the AI "couldn't handle." It was the contract state machine, the milestone acceptance logic, and the test suite that pins both down. We wrote those by hand, then had the AI extend them, because that is the part where a subtle mistake costs a client real money rather than a broken button.
The split matters for how you staff a project like this. You do not need fewer engineers because AI writes code faster. You need engineers who are better at reading code quickly, because review becomes the job.
Four regressions the rig caught
Our quality rig runs every change against a fixed set of contract scenarios — part-completed milestones, mid-contract scope variations, currency rounding at the cent level, that kind of thing. Across nine weeks it flagged four regressions where AI-generated changes altered invoice totals in ways that passed every existing unit test but broke a scenario the rig specifically modelled. Two were rounding-direction changes that would have shorted us, not the client. Two would have double-counted a partially delivered milestone. None of the four were caught by code review alone before the rig existed as a gate — they were the kind of change that looks correct on the diff and is wrong three steps downstream.
That is the argument for a rig over a reviewer's judgement: a person reviewing a diff is checking whether the change looks right. A rig is checking whether the system still behaves right, against scenarios nobody re-derives from memory on a Tuesday afternoon.
What we would change next time
We under-invested in the rig for the first two weeks, treating it as something to build once the "real" feature work was further along. That was backwards. The two regressions that were closest to shipping undetected both landed in that early window, before the contract-scenario coverage was complete. On the next build we start the rig in week one, even if it only covers a fraction of the surface area at first. A thin rig from day one beats a thorough rig from week three.
The honest summary
Nine weeks against a seven-month baseline is a real result, not a marketing rounding error. It is also not "AI did it for us." It is three people, a model that writes fast, and a quality rig that was strict enough to catch four mistakes that would have cost us client trust. Take away any one of those three and the number does not hold.
Outcome-priced from day one
See what this would cost at Effektiv pace.
Pick a project that finished or stalled. Show us a quote you've received or an invoice you've paid. We'll price the same scope on outcomes, not hours.