AI in a real workflow is mostly the failure path

Capability is the easy screenshot

It is not hard to make a model look useful. Give it a tidy prompt, a clean document, and an audience that already knows the answer. The meeting goes well. Then the same workflow meets a half-filled form, a customer who used the wrong word, a policy that changed last Tuesday, and a model that is confident about all of it.

We take AI work when the model sits inside a path that already matters: a decision, a draft that will be sent, a classification that will move money or access. In that setting the interesting design is not “can it.” It is “what happens next when it cannot.”

Write the failure path before you pick a model

Before anyone compares vendors, write down the cases in which the workflow must not accept a model’s answer. Those cases are the product. The model is a component that sometimes fills them and sometimes must be skipped.

• What must never be invented: prices, permissions, medical or legal claims, anything that will be treated as a record.
• Who is allowed to approve a borderline result, and how long they have.
• What the user sees when the model is slow or down. A spinner is not a fallback.
• How you will know the model has drifted: a held-out set, a sample of production, a cost or latency budget that, once crossed, turns the feature off.

Evaluation is not a vibe

“It seems better” is not a release criterion. Build a small set of real inputs — the ugly ones, not the brochure ones — and decide in advance what good looks like. Score the cases that would hurt if they were wrong more heavily than the cases that would merely look clever.

Revisit the set. The distribution of real inputs will move. A workflow that was safe in April can become unsafe in August because the business started selling something new, or because a vendor changed a default.

If a change cannot be assessed twice, the team is guessing. Guessing is fine in a prototype. It is not a production practice.

Cost and latency are product decisions

A path that is right and takes eight seconds will be bypassed. A path that is cheap and wrong will be used until it is a scandal. Put numbers on both before the model is wired into the critical path: a latency budget the user can feel, a cost budget the business can survive at ten times today’s volume, and a fallback that keeps the core transaction possible.

Sometimes the honest answer is not to put a model in the loop. A deterministic rule, a better search, or a human with a good form will beat a generative step that cannot be checked. We will say that. It is cheaper than a chatbot that operations quietly stop using.

This is the work we mean by “AI in your product”

When someone asks us to “add AI,” we translate that into a workflow, an evaluation, a fallback, and an owner. If those cannot be named, we are not ready to pick a model. We are ready for a short advisory conversation.

If you already have a demo that impressed a room, bring that. Bring also the case that would embarrass you. That second artefact is where the real design starts.