Most AI roadmaps contain ten feature ideas. In our experience roughly four of them will reach production and stay there. The other six die somewhere between the demo and the incident review. The pattern of which is which is remarkably consistent.
The four that ship
Retrieval over your own documents. Support articles, policies, contracts, runbooks. The corpus is bounded, the ground truth exists, and a wrong answer is visibly wrong. This is the workhorse, and it's why it's the first thing we bid.
Structured extraction. Invoices to fields, emails to tickets, PDFs to rows. Unglamorous, measurable, and the model is graded against a schema, not a vibe. Payback is usually measured in weeks.
Drafting with a human on the send button. Reply suggestions, report first-passes, summaries a person edits before it leaves the building. The human absorbs the tail risk, so the feature survives its first bad output.
Classification and routing. Priority, sentiment, department, escalation. Low ceremony, easy to eval, and wrong answers cost a re-route rather than a customer.
The six that don't
The fully autonomous agent that acts on production systems without review. The chatbot over data nobody cleaned. The feature whose success metric was never defined past "wow." The one that needs real-time data the org doesn't actually have. The one where a hallucination is a legal event. And the one built because a competitor announced it.
None of these are impossible. They're just not first projects. Each one either lacks bounded inputs, lacks a definition of correct, or puts the model's worst output directly in front of a customer or a regulator.
The triage question
One question sorts the list faster than any framework: when the model is wrong, who notices, and what does it cost? If the answer is "a person notices during their normal work, and it costs an edit" — build it. If the answer is "nobody notices until a customer does" — build the eval harness first, or build something else.
Not sure which side of that question your feature falls on? That is the question a diagnostic sprint exists to answer: one week, fixed fee, a written verdict against your real data.