Real-World Stories

Patterns and anti-patterns from shipping AI into real systems.

From the field

Rather than tell you specific stories, I want to share the patterns that keep showing up when teams ship AI features. None of these are unique. All of them are worth knowing about before you hit them yourself.

Patterns that tend to work

Suggest, don't act. When the AI proposes and a human approves, the existing review culture does the safety work for you.
Route through a PR. If the agent's output is a pull request, you already have audit, review, rollback and CODEOWNERS for free.
Scope tightly. A small, well-defined task with a narrow blast radius almost always lands better than a broad, ambitious one.
Show your work. Citations, plans and diffs in the UI build more trust than higher accuracy ever will.
Have an off switch. Per feature, per tenant, per tool. Tested.

Anti-patterns to watch for

Approval theatre. A human is "in the loop" but couldn't realistically review every item in the stream.
No blast-radius thinking. An automation is given the power to take an action it can't reasonably undo.
Silent drift. A model or prompt change ships and nobody notices the behaviour shifted because there were no evals.
Hidden AI. Users can't tell they're talking to an AI, or can't tell which parts of the experience were generated.
One-way decisions. The system can act, but the user has no path to challenge, correct or appeal.

A useful frame

The system is the unit of reliability, not the model.

The same model can be safe in one product and unsafe in another. The thing you're shipping is the system around the model: the prompts, the tools, the guardrails, the UI, the review process, the rollback. That is what you are accountable for.