AI SDLC transformation
Flagship case, ~100 people, work in progress.
Where it started
The obvious move failed. Hand developers AI agents, expect a 10x jump in time-to-market - and instead you get masses of unread code, the team’s shared context torn apart, quality dropping and documentation falling behind. This is the failure mode most AI adoption is walking into right now.
The turn
The counterintuitive fix: development gets AI last. AI goes first into feature preparation and planning, where it compounds instead of eroding.
The pipeline, in order:
- 1Record and transcribe every grooming session.
- 2Agents break down the artefacts and enrich the context.
- 3Documentation review speeds up - summaries, proofreading, cross-checks against sources, role-based review.
- 4Tasks are generated straight from artefacts through MCP / skills / connectors, with manual effort near zero.
- 5Architecture is created and validated - paper sketches become mermaid / PlantUML in seconds, with tests and cross-checks.
- 6Helm charts for operations are generated from the artefacts.
- 7Test cases and automated tests are generated for QA.
- 8Only at the end - agentic development.
Why it worked
The method is BMAD, layered on top of DDD and docs-as-code that were already in place:
- BMAD coordinates agents through artefacts - and those artefacts already lived in git under team-wide review.
- DDD gave the agents an unambiguous language: the ubiquitous language doubles as project context.
- BMAD’s two phases - Agentic Planning, then Context-Engineered Development - matched my own decision to give development AI last.
- The failed naïve stage was a documented failure mode - exactly what BMAD is designed against.
Timeline and scale
R&D and pilots on micro-teams took 4 months; the full transition of a stream took 3 more months after R&D. Scale so far: about 100 people, with the rollout still going - this is honestly work in progress.
Results
- Quarter planning used to be two-to-three weeks of intense work layered on top of ongoing development. Now it’s part of the ordinary process: a roadmap plus feature skeletons in about a day of work by the delivery manager and leads with agents.
- Delivery speed: roughly 2x on greenfield, +60% on brownfield.
- Overtime and incidents have almost disappeared.
- Predictability is up.
- Two or three hypotheses are worked in parallel (record: four) instead of one.
The business no longer chooses which hypothesis to sacrifice - it can choose all of them, and get a production-ready result without replanning or halting sprints.
The through-line
This rollout worked because the foundation was already there - the processing & anti-fraud turnaround is where DDD, docs-as-code and ADR/RFC went in. And the speed instinct predates AI entirely: see the greenfield digital banking PoC, built end-to-end in 10 weeks with no AI at all.