When an AI feature that worked beautifully in demo starts skipping steps, looping on itself, or giving different answers to the same question on a Tuesday, the instinct is to blame the model. The instinct is wrong. The actual cause — and the architectural shift that fixes it for good.
AI features that drift in production almost never have a model problem. They have an architecture problem: the model was made the workflow orchestrator, so when input distribution widens, every drift in understanding becomes a drift in workflow. The durable fix moves orchestration into typed code and keeps the model only for what it is uniquely good at.
In nearly every case we audit, the model did not change — the input distribution did. New customer segments, new partner formats, new internal code paths all expand the inputs the system sees beyond what the original prompt was tuned against. When the model is also the workflow orchestrator, every drift in understanding becomes a drift in what the system does.
Prompts that grow scar tissue — adding clauses for every past incident — eventually optimise for tail cases at the centre's expense. We have seen prompts grow from five paragraphs to forty over a year, with the system measurably less reliable at the end. More instructions produce more places for attention to leak, not more obedience.
Less than teams expect. The shift is to move workflow orchestration out of the model and into typed code, while keeping the model for the steps that genuinely benefit from probabilistic reasoning. It is a redesign of where decisions live, not a rewrite of the feature. Most engagements move the architecture in two to four weeks.