Everyone got their own AI. The BA has one, the developer has one, the tester has one, and each of them is genuinely faster at their own desk. Projects still ship on roughly the dates they used to. We spent a while trying to work out why, and the answer turned out to be structural rather than cultural: each of those assistants runs in its own context window, and none of them can hand what they learned to the next one.
Everything gets worked out more than once
An assistant is only as useful as the context it is holding. The BA’s spends an afternoon getting across a domain, the stakeholder’s real intent and the edge cases that matter, and produces a sharp requirement. Then the work moves to the next desk, and all of that context stays behind. The developer’s assistant starts from nothing. The tester’s starts from nothing again, reading the same requirement to work out what “done” was supposed to mean.
Do that across four handoffs and the same understanding gets reconstructed three or four times. Sometimes differently each time, which is worse than not doing it at all.
When every person runs their own AI in their own context window, what gets understood in one seat never reaches the next one. The gains stay local, and the team keeps paying for the same work.
Why local speed isn’t team speed
Individual productivity is easy to feel and easy to measure. Team throughput is neither, and it is the one that pays the bills. The gap between them sits in the handoff.
- Rework disguised as work. The tester reads the whole requirement to infer acceptance criteria the BA had already settled. That job gets done twice, by two people who never compare answers.
- Handoffs leak. Every one compresses a rich understanding into a document and then decompresses it imperfectly at the other end. Detail goes missing at each step, and nobody notices until testing.
- Nothing carries forward. Effort adds up at best. It never multiplies, because each seat starts from a blank page.
“My AI made me twice as fast. We shipped on exactly the same date.”
What we did differently
ADIS is built on one inversion. Whatever a stage produces is what the next stage consumes, as a structured artifact rather than a conversation someone has to have again from scratch. Context gets captured once and carried forward instead of evaporating at the door.
The requirement handoff is the clearest example. The BA works with AI to shape a requirement, and that work ends in unambiguous acceptance criteria. Those criteria are the handoff. The tester doesn’t reinterpret anything, because the same criteria drive test case generation directly.
What moves between seats is crisp acceptance criteria, not a document the next person has to read closely and interpret for themselves.
Test cases generate straight from the acceptance criteria. Build scope, tests and evidence all trace back to the same source.
The traceability grid holds what each stage produced and makes it available to every other stage, which is the opposite of a room full of private context windows.
Someone signs off before an artifact moves on, so speed comes from not repeating work rather than from removing the review.
Individual AI is addition. Everyone gets a bit quicker and the totals move a little. What we were after is the other thing: the BA’s clarity making the tester’s job smaller, the tester’s coverage making the release safer, and nobody in the chain paying twice for the same understanding.