When to embed AI in your product (and when not to)
AI features can shave hours off a workflow — or quietly torch your unit economics. Three filters we use before greenlighting an AI feature.
Every product brief in 2026 includes the line 'and we should add AI somewhere.' Half of them shouldn't. Here's the test we run.
Filter 1: Is the user already iterating?
AI shines when the user is already cycling through drafts: writing, designing, planning, debugging. If the workflow is single-shot (look up a record, file a complaint, pay a bill), an LLM in the loop usually adds latency and cost without changing the outcome.
Filter 2: Can the user tell when it's wrong?
If the answer is no, you have a verification problem. Either constrain the output (structured fields, citations, tool calls with deterministic backends) or don't ship the feature. 'Plausible-sounding nonsense at scale' is the failure mode that erodes trust faster than any bug.
Filter 3: Does the margin work?
Run the math on a per-interaction basis: token cost × interactions per session × sessions per user × users. Compare to the revenue or retention impact. We've killed three planned AI features in the last year because the unit economics didn't survive the projection.
What 'good' looks like
Some examples from our portfolio: AI-powered search and listing matching in Dealio (constrained output, clear verification, sub-cent token cost); first-draft visit notes in DocDoc (heavy iteration loop, the doctor reviews every word); and campaign drafting in Veev, where AI proposes the copy and a human always approves before anything goes live.