Skip to content
All posts

How long does it take to build an AI MVP? A week-by-week plan

A realistic AI MVP timeline: what happens in each of six weeks, what has to be ready before week one, and what usually causes delays.

Shubham Somani 4 min read

Founders usually ask two questions about an AI MVP: what will it cost, and how long will it take? We covered cost in how much it costs to build an AI MVP. This post covers the AI MVP timeline: a realistic week-by-week plan for a focused first version, and what makes it slip.

The short answer is that a focused AI MVP can launch in about six weeks. “Focused” is doing most of the work in that sentence.

Before week one: what needs to be true

Six weeks only works if a few things are settled before the clock starts:

  • One user and one job. Not a platform, but a specific person doing a specific task. See picking your first AI feature.
  • Access to real data. Sample documents, records or conversations the feature will run on, plus permission to use them.
  • A decision-maker who’s available. Someone who can answer product questions within a day, not a week.
  • A rough idea of “good enough”. What a useful answer looks like, and what the feature must never do.

If these aren’t in place, spend a week or two getting them in place first. That’s what a short discovery sprint is for.

Weeks 1 and 2: prove it works on your data

  • Build a rough prototype of the core AI step against real data.
  • Write the first eval set: around 30 real examples, including the cases it must refuse. More on that in how to test an LLM feature.
  • Try two or three approaches to prompting and retrieval, and keep the one that scores best.
  • Estimate the running cost per request.

The checkpoint at the end of week 2: does the core step work well enough on real examples to be worth building around? If it doesn’t, you’ve learned that cheaply, before any product is built.

Weeks 3 and 4: build the product around it

  • Build the screens the user actually needs, and nothing more.
  • Add auth, data access with the right permissions, and any integrations the workflow depends on.
  • Turn the prototype into production code: retries, timeouts, a fallback when the model fails, and structured logging.
  • Keep running the evals on every change so quality doesn’t slip while you build.

The checkpoint at the end of week 4: someone on your team can use it end to end on real data.

Week 5: harden it

  • Add guardrails: input limits, output checks, and a human step before anything irreversible.
  • Set up monitoring for errors, latency and cost per request.
  • Test the edge cases from the eval set, along with anything new the team found in week 4.
  • Write a short runbook covering what to check and who to call when something looks wrong.

Week 6: launch to real users

  • Release to a small group of pilot users first.
  • Watch real usage daily, and add every bad answer to the eval set.
  • Fix the top issues, then widen access.
  • Agree what happens next: the improvements, the metrics to watch, and who owns them.

What usually causes delays

  • Data that arrives late or turns out to be messier than expected.
  • Scope that grows mid-build: a second user type, another integration, “while we’re at it”.
  • Slow decisions on product questions that block the team for days.
  • No agreed quality bar, so “is it good enough?” turns into a debate instead of a number.

Each of these is cheaper to fix before week one than during week four.

The short version

  • A focused AI MVP can launch in about six weeks.
  • Settle the user, the job, the data and the quality bar before you start.
  • Prove the core AI step on real data in the first two weeks.
  • Build the product in weeks 3 and 4, harden it in week 5, and launch to a pilot group in week 6.
  • Most delays come from data, scope and decisions, not from the model.

Our AI MVP follows this six-week shape. If you’d like to see whether your idea fits, book a 30-minute call.