Where AI Belongs in Your MVP (and Where It Doesn't)

"Is it AI-powered?" became the wrong question in 2023. It's still being asked in 2026.

Here's the thing: AI is a tool. A brilliant one. But bolting it onto an MVP because it's trendy is a great way to spend your seed round on a feature nobody uses.

We've shipped AI into plenty of products — and talked founders out of it in plenty more. Here's how we decide.

The rule: AI earns its place when it replaces effort

Every AI feature in an MVP should answer one question: what human effort does this eliminate?

Good answers

• "It summarises a 50-message WhatsApp thread so the user doesn't have to scroll."
• "It drafts a parody CV from a LinkedIn URL in 4 seconds instead of 20 minutes."
• "It classifies incoming leads so ops doesn't sort them by hand."

Bad answers

• "It feels modern."
• "Investors like it."
• "Our competitor has it."

If you can't finish the sentence with a specific hour of human work it kills, don't build it yet.

Where AI almost always earns its place in an MVP

Four patterns we keep seeing pay off:

1. Summarisation

Chat logs, documents, meeting notes, customer tickets. High-volume text distilled to output. Cheap to build, obvious value.

2. Classification & routing

Leads, support tickets, user-generated content. Turning unstructured input into structured decisions.

3. Generation with a narrow prompt

First-draft emails, product descriptions, ad copy variations. Human reviews; AI produces the raw material.

4. Search & retrieval

Semantic search over your own content. Users find things faster. RAG is mature enough in 2026 that this is now a solved pattern.

Where AI is usually a mistake in an MVP

1. Core differentiation you haven't validated

If your whole pitch is "we use AI," you don't have a product — you have a feature. Validate the workflow manually first.

2. Anything requiring ~100% accuracy

Medical, legal, financial calculations. Hallucinations are a feature, not a bug, of LLMs. If your MVP fails when the model is wrong, don't build on LLMs yet.

3. Chatbots with no specific job

The "talk to our data" chatbot is 2023's homepage slider. Users don't want to type — they want an answer in one click.

4. Anywhere cost per call exceeds the value

Run the math. A $0.02 API call × 10,000 users × 30 sessions/month = $6,000/month for a feature they might not use. Free-tier economics get scary fast.

The test we use with clients

Before any AI feature ships in an MVP, we ask three questions:

• What does the user do less of because of this? (If the answer is vague, cut it.)
• What happens when the model is wrong? (If the answer is "user gets hurt," cut it.)
• Can we ship the MVP without it? (If yes, defer to v2 and use the runway to validate core.)