How to scope an AI MVP you can ship in 8 weeks

An AI demo can be built in an afternoon. An AI product that people trust with real work takes longer — but not as long as most founders fear, if the scope is right. Here is the approach we use, with our own product SpecLens as the worked example.

1. Find the one painful loop

Write down the job your user does today, step by step, and circle the step that costs them the most time or the most mistakes. That step is your MVP. Everything else is version two.

For SpecLens, the job was comparing product specification sheets from multiple vendors. Procurement teams copied specs out of PDFs into spreadsheets, reconciled units, and chased missing fields. The first version did exactly one thing: take uploaded vendor documents and produce a comparison matrix.

2. Write the scope as input → output

If a feature does not change whether the output is correct and useful, it goes in the right-hand column.

3. Build the pipeline, not just the prompt

The prompt is rarely the bottleneck. In SpecLens, the hardest problem was turning research prototypes into a reliable production pipeline. A typical AI MVP pipeline has five stages:

1. Ingestion — accept messy real-world files and convert them to a common structure.
2. Extraction — use the model with a strict output schema, not free text.
3. Normalisation — convert units, names, and formats so like is compared with like.
4. Validation — check the output against rules and flag anything missing or inconsistent.
5. Presentation — show results with enough context for the user to trust or correct them.

4. Design for wrong answers

Every AI system is sometimes wrong. The products that succeed make errors cheap to spot and cheap to fix:

• Citations — link every extracted value to the page it came from.
• Confidence scores — highlight low-confidence values instead of hiding them.
• Editable output — let users correct a value in one click, and log the correction.
• An evaluation set — keep 20–50 real examples with known answers and re-run them whenever you change the model or prompt.

5. Pick boring technology around the model

Use a mainstream web stack, a managed database, and hosted model APIs. Keep the model behind an interface so you can switch providers as prices and capabilities change. Novel infrastructure is a cost you pay before you have learned anything from users.

6. A realistic 8-week plan

Common scoping mistakes

• Building an admin panel, roles, and settings before anyone has used the core feature.
• Promising "works with any document" instead of naming the formats you support.
• Skipping the evaluation set, then being unable to tell whether a change made things better or worse.
• Treating a chat interface as the product when users actually want a finished output.

Have an AI product idea you want to pressure-test? Tell us about it — we will help you cut it down to a version you can ship.

Frequently asked questions