Launch — and the long life after launch

Lesson 4 of 5 in The AI Lifecycle: From Data to Deployment to Retirement.

Deployment is not a moment; it is a set of design decisions. How a system is released shapes what can go wrong and what can be undone:

• Staged rollout — release to 1% of users, watch, expand. Limits blast radius; standard practice for consequential systems.
• API-gated release — the model stays on the provider’s servers behind usage policies, rate limits, and abuse monitoring. Control is retained; misuse can be cut off.
• Open-weight release — the model’s parameters are published for anyone to download, inspect, and fine-tune. Maximum transparency and access — and irreversible: there is no recall button for weights already on a million laptops.

Deployment is also where three obligations become concrete: configuring human oversight (who reviews what, with what authority), disclosure (people must know they are interacting with AI where the law requires it — EU AI Act Article 50 being the flagship), and integration safeguards (the guardrails between the model and the systems it can touch).

Then begins stage 6 — usually the longest and most neglected. A model is a snapshot of the world at training time, and the world does not stand still. Customer behaviour shifts, fraudsters adapt, language changes, a pandemic rewrites every pattern at once. The model’s accuracy decays — silently, because a wrong prediction looks exactly like a right one from the inside. This is model drift, driven by distribution shift, and it is the reason “we tested it before launch” is never a complete answer.

Operating well means five running disciplines: monitoring performance (disaggregated, against live outcomes where possible), logging decisions so incidents can be reconstructed, incident detection and response with clear escalation paths, feedback-loop awareness (are the model’s own outputs reshaping the data it will be retrained on?), and retraining triggers — pre-agreed thresholds at which the loop back to stage 3 fires.

Regulation has caught up with this reality. The EU AI Act requires providers of high-risk systems to run a post-market monitoring system and report serious incidents; medical-AI and automotive regimes have demanded the same for years. Deployment, in every modern framework, is where obligations intensify, not where they end.

Key terms: model drift, distribution shift, post-market monitoring, open-weight model, human oversight

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.