AI coding pitfall
Architectural drift failure pattern
Each AI session optimises the local problem. Over weeks, the system becomes a collage of styles, data models, and integration patterns with no single coherent design.
How it shows up
- Multiple competing ways to do the same thing across services and modules.
- Data models that diverge from the domain; events and APIs that no longer match.
- Teams cannot predict impact of change, so delivery slows despite "faster coding".
Who this pattern hits
- Teams whose system became a collage of AI session styles and data models.
- Architects who need one coherent spine again before the next wave of features.
- Leaders who see local optimisations destroying global coherence.
When it matters: When every local AI change made the global system harder to reason about and change safely.
How we recover
Our solution
Re-establish one coherent architecture
We map what was built, define target boundaries, and realign services, data, and APIs toward a single model. Drift stops when local AI changes are judged against an explicit architecture, not the last chat session.
- Document current vs target architecture and prioritise the highest-risk seams.
- Converge on shared patterns for data, events, and integration contracts.
- Guardrails so new AI-assisted work cannot reintroduce competing designs.
We re-establish one coherent architecture and a path to extract or retire the rest, with explicit boundaries for future AI-assisted work.
More →Talk through this pattern
Start on the home page with a short email to Robert, free and no obligation.
Book free recovery discussionWhat this failure mode is
Architectural drift is when multi-session AI delivery creates competing designs for the same concerns — data access, events, errors, and integration — until no single model of the system remains true.
Symptoms you can observe
- Three ways to call the same dependency, none fully documented.
- Event names and payloads evolve without versioning or consumers that still work.
- New features avoid existing modules because “the AI suggested a cleaner approach”.
- Onboarding requires tribal knowledge of which pattern is current.
Why AI-accelerated delivery makes it worse
Each AI session maximises local coherence. Across months, that produces global incoherence unless humans enforce convergence.
How Mayordomo recovers it
- Document current versus target architecture and prioritise the highest-risk seams.
- Converge on shared patterns for data, events, and integration contracts.
- Add guardrails in review and tooling so new work cannot reintroduce competing designs.
Related services: AI Project Recovery & Realignment · Modernisation & rescue. See also the failure patterns hub.