The AI feature your roadmap doesn't need yet
Voice mode, multi-agent orchestration, and "autonomous workflows" sound great in pitch meetings. Your users might just want search that works.
We get a lot of MVP enquiries that open with a feature list. AI voice assistant. Slack bot. Auto-generated reports. Something with "agents" in the name. When we ask what job the user is hiring the product for, the room goes quiet for a second. That's not a knock on anyone — the hype cycle is loud right now.
Start with the job, not the stack
The MVPs we've shipped fastest share a pattern: one painful manual step, one measurable outcome. A property manager spending six hours a week copying inspection notes into emails. A clinic coordinator retyping referral details from faxes. A founder answering the same onboarding questions in DMs. The AI piece is usually extraction, classification, or draft-and-review — not a general-purpose copilot wearing every hat.
Compare that to a dashboard with six AI tabs where each tab does something slightly different and none connect to a billing event. Those projects drift. They demo well and stall in beta because nobody owns the feedback loop.
Red flags we gently push back on
- "We need multi-agent" before anyone has defined agent number one
- Voice as v1 when 95% of usage will be async text
- Building your own model hosting before you've validated retention
- AI for every screen instead of the one step users complain about loudest
- No plan for wrong answers — legal, medical, and money-adjacent products need this on day one
What a sane first release looks like
Pick a workflow. Narrow the input types. Show your work (what did the system read, what did it propose). Let humans approve before anything external sends. Log everything. Ship to ten real users, not a hundred beta signups you'll never interview.
The best AI MVPs we've shipped could be described in one sentence without the word revolutionary.
That doesn't mean boring. It means legible. Your investors and your users can both understand what it does. You can count whether it saved time. You can rip it out if it didn't.
If you're early stage and reading this because you're writing a spec: strip one feature from the list. Then strip another. What's left is probably where we should start talking.