The rebuild instinct is usually wrong
Teams assume adding AI means re-architecting, and the assumption kills the project on cost before it starts. In practice most first features need one place where a model can read some context and return something a human then uses.
Find the step where judgment happens
Somewhere in your product a person reads something and decides. That step is the seam — it already has the inputs assembled and an output that gets used, which is why designing the workflow first and inserting the model after works on existing systems too.
You’re not adding AI to a product. You’re adding it to one step of one workflow.
Suggest before you replace
The lowest-risk insertion presents a suggestion the existing user accepts or edits, leaving every current path intact. Nothing breaks if the feature is switched off, which makes it safe to roll out gradually.
The retrieval layer is the real work
Getting the right context to the model — the relevant records, the current policy, the customer history — is usually harder than the model call, and it is where permissions have to be enforced rather than assumed.
You will discover what wasn’t recorded
Most retrofits stall on the fact that the system stored outcomes and not decisions. That gap is the instrumentation that should have existed, and closing it for one workflow is a smaller job than it sounds.
Ship it narrow and keep the seam
One workflow, one team, real traffic. The infrastructure built for it — retrieval, logging, evaluation — is what makes the second feature cheap, which is the argument for sequencing features so they compound.