It has never been easier to build an impressive AI demo. It is still hard to build an AI product — something people rely on every day. The gap between the two is where most AI projects stall.
We've shipped AI features across recruiting, healthcare, events and photography — including Slingshot, 2Care AI, running.pictures and Vivah.pictures. Here is what we've learned about getting from idea to a usable AI MVP.
1. Start with a job, not a model
The strongest AI products do one clearly defined job better than the old way:
- Slingshot filters and matches candidates for marketing roles, so companies close positions within about a week instead of the two to four weeks a hire typically takes.
- running.pictures and Vivah.pictures let people find every photo of themselves in thousands of event images by uploading a selfie.
- 2Care AI remembers the medical documents a patient shares once and can reference them for life.
Notice that none of these are "a chatbot for X". Each one removes a specific, painful step. Write your job down in one sentence before choosing any technology.
2. Decide what "good enough" means
AI is probabilistic — it will sometimes be wrong. Before building, agree on:
- What accuracy is acceptable for the first release, and how you'll measure it.
- What happens when the model is unsure. Show fewer, better results? Ask the user? Hand off to a person?
- What a mistake costs. A slightly-off photo match is harmless; a wrong medical summary is not. Higher stakes mean more human review and tighter guardrails.
Build a small evaluation set — 50 to 200 real examples with the right answers — early. It turns "it feels better" into a number you can track.
3. Choose the simplest model that works
You rarely need to train your own model for an MVP. Options, from simplest to most involved:
- Hosted large language models for text understanding, extraction, summarisation and chat.
- Specialised APIs and open models for vision, face recognition, speech and search.
- Retrieval (RAG) to ground answers in your own documents and data.
- Fine-tuning or custom models — only when the above can't hit your quality bar.
Pick based on accuracy on your evaluation set, cost per request at your expected volume, latency, and data-privacy requirements.
4. The product around the model matters most
In most AI products, the model is a small part of the code. The rest is ordinary, essential product work:
- Data flows — getting documents, images or records in cleanly, and keeping them secure.
- Interfaces that build trust — show sources, confidence, and an easy way to correct mistakes.
- Fallbacks — sensible behaviour when the model or API is slow or unavailable.
- Cost controls — caching, batching and limits so usage growth doesn't surprise you.
- Feedback loops — capture corrections so the product gets better over time.
5. Scope an AI MVP like any MVP
The same rules apply as for any first release: one user, one job, one core flow. A realistic AI MVP usually takes 8–12 weeks; products with sensitive data, several models, or deep integrations take 3–6 months. See our MVP cost guide for budget ranges.
A checklist before you start
- One-sentence job the AI does
- 50+ real examples to evaluate against
- Agreed accuracy bar for launch
- Plan for low-confidence results
- Data privacy requirements written down
- Estimated cost per user per month
Ready to build?
If you have an AI idea and want to know what a first version would take, tell us about it. You can also read more about our AI product development work.