What to put in an AI MVP spec: a one-page template for founders
A practical AI MVP spec template: the user, the job, the data, what good looks like and what the feature must never do. Copy it before you talk to any builder.
Most AI MVPs that go over time or budget don’t fail on the model. They fail because nobody wrote down what the feature is for, what data it runs on and what “good enough” means. A short AI MVP spec fixes most of that before anyone writes code.
You don’t need a 40-page requirements document. You need one page that answers the questions an engineer will ask in the first week anyway. Here’s the template we’d want from any founder, with notes on each section.
1. The user and the job
- Who uses it? One specific type of person, such as “a support agent handling refund requests”, not “our users”.
- What job are they doing? Describe the task as it happens today, step by step.
- What changes with AI? Which step gets faster, cheaper or possible for the first time.
If you can’t name one user and one job, start with picking your first AI feature before writing the spec.
2. Inputs and data
- What goes in? Emails, PDFs, database records, free-text questions, images.
- Where does it live, and who can give the team access?
- Ten real examples. Include messy ones. They’re the most useful thing you can hand a builder.
- What’s sensitive? Personal data, customer contracts, anything that can’t leave a region. See keeping customer data safe in LLM features.
3. Outputs and where they go
- What comes out? A draft reply, a summary, a classification, filled-in fields, an action.
- Who sees it before it counts? A human reviews it, the user accepts or edits it, or it runs on its own.
- Where does it land? Your app, a CRM, a Slack channel, an email.
Starting with a human in the loop is usually the cheapest way to launch safely.
4. What good looks like
This is the section most specs skip, and the one that matters most.
- Five to ten example inputs with the answer you’d accept. These become your first tests.
- What’s acceptable to get wrong, and how often, in plain words.
- How fast it needs to feel. Instant, a few seconds, or “done by tomorrow morning” are very different builds.
Without this, “is it good enough?” becomes a debate instead of a check. We cover turning these examples into tests in test your LLM feature before users do.
5. What it must never do
- Make up prices, policies or legal statements
- Act without confirmation on anything that costs money or can’t be undone
- Show one customer’s data to another
- Answer questions outside its scope instead of handing off
Write these down explicitly. They shape the prompts, the guardrails and the tests.
6. Scope: in, out and later
- In: the one workflow above, for one user type.
- Out: everything else, written down so it doesn’t creep back in mid-build.
- Later: good ideas you’re deliberately parking for version two.
7. Constraints
- Deadline, and what it’s tied to: a launch, a customer, a fundraise.
- Budget range for the build and a rough ceiling for monthly running costs. See what an AI feature costs to run.
- Existing stack the feature has to fit into.
- Who decides product questions, and how quickly they can answer.
Copy-paste template
AI MVP spec: <feature name>
User: <one specific person>
Job today: <steps>
What AI changes: <one sentence>
Inputs and data source: <types, location, access owner>
Sensitive data: <what, and any rules>
Output and destination: <what, where, who reviews>
Good looks like: <5-10 example inputs with accepted answers>
Acceptable errors: <plain words>
Speed: <instant / seconds / overnight>
Must never: <list>
In scope: <one workflow>
Out of scope / later: <list>
Deadline and why: <date, reason>
Budget: <build range, monthly running ceiling>
Decision-maker: <name, response time>The short version
A good AI MVP spec names one user and one job, hands over real data, defines what good looks like with examples, and lists what the feature must never do. Fill in the template before your first call with any builder and you’ll get better estimates and a faster start.
If you’d like help turning a spec like this into a working product, see our fixed-scope offers or book a 30-minute call.