Product, operations, or nowhere
A large share of AI ideas that arrive as product requests are actually operational opportunities. Summarizing internal tickets, drafting responses for your own support team, and classifying inbound requests improve your margins — they don't necessarily improve your customer's outcome, and shipping them as customer-facing features adds surface area for no gain.
The sorting question is simple: does the customer's outcome change, or does our cost to serve change? Both are worth pursuing. They belong on different roadmaps with different owners.
Five questions before it enters the roadmap
- Which customer problem does this solve, in their words? If the answer only exists in our language, discovery hasn't happened yet.
- What evidence do we have? Named conversations, observed workarounds, or usage data — not a competitor's launch post.
- Is the data reality good enough? Coverage, freshness, permissions, and quality. Most disappointing AI features are data problems wearing a model costume.
- How will we evaluate quality? A test set, a review process, and a threshold agreed before the build, not a subjective read at demo time.
- What happens when it's wrong? Wrong confidently, in front of a customer, at scale. If the answer is unclear, the feature isn't ready.
Evaluation and failure behaviour
Deterministic features either work or they don't. Probabilistic ones are correct most of the time, which means quality is a distribution rather than a state, and it has to be measured deliberately.
Practically, that means three things in the spec: a held-out set of realistic examples with expected outputs, a defined acceptable error rate for that use case, and a designed failure path — visible uncertainty, an easy correction, and a human route when confidence is low. Features that let users see and fix errors survive imperfect accuracy. Features that assert wrong answers confidently lose trust once and rarely get it back.
Why competitive parity is a weak reason
"A competitor shipped it" tells you they made a bet. It doesn't tell you the bet worked, and it certainly doesn't tell you it fits your segment or your data. Parity features consistently ship, demo, and then go unused — while consuming the roadmap capacity that a well-evidenced bet needed.
If a board or sales conversation makes an AI capability commercially necessary regardless of evidence, treat it as an explicit commercial commitment and say so in the operating review. Name it, size it, and move something. That's an honest override, not a product decision dressed up as one.