A minimum viable product is the smallest version of your product that real users can use, built to test whether the idea works before you spend on the whole thing. For a startup, MVP development is less about writing code quickly and more about choosing what not to build. Get the scope right and an MVP answers the question that matters most, whether people want this, in months rather than years, on a budget you can survive being wrong with. This guide covers how startups should scope an MVP, how long MVP development takes, what drives the cost, which technology to choose, and what to do once it is in users’ hands.
What an MVP is, and what it is not
An MVP is a tool for learning. It exists to test your riskiest assumption with real people, and it is viable because those people can actually use it to get something done. It is not:
- A demo or clickable prototype. Those show what the product could be; an MVP has users doing real work with it.
- Version one of everything. If your MVP has every feature on the roadmap at lower quality, it is a rushed full product.
- Throwaway code. A good MVP is small, but it is built so the next release can grow from it rather than start again.
Not every MVP needs software at all. Depending on what you need to learn, one of these may come first:
| Type | What it tests | Needs development? |
|---|---|---|
| Landing page or waitlist | Whether people want it enough to sign up or pre-order | Barely |
| Concierge MVP | Whether the service is valuable, delivered by hand to a few customers | No |
| “Wizard of Oz” MVP | Whether users value the outcome, with people doing behind the scenes what software will do later | A front end only |
| No-code MVP | Whether a simple workflow product gets used | Configuration, little code |
| Single-feature MVP | Whether the core feature solves the problem well enough to keep users | Yes |
| AI MVP | Whether the AI can do the job well enough, and whether users trust it | Yes |
The rest of this guide is about the last two, where you are building real software.
Start with the riskiest assumption
Every startup idea rests on a stack of assumptions: this problem is painful, these people have it, they will change how they work, they will pay, we can build it. The MVP should test the one that would kill the business if it were wrong, and that is usually about demand, not technology.
Write it down as a sentence with a number in it, before any design work: “At least 15 of the 40 accounting firms in our pilot will use the approval flow every week for a month.” That sentence decides the scope, tells you which metrics to build in, and stops the argument at the end about whether the MVP worked.
How to scope an MVP
- Name one user and one job. Not “small businesses”, but “the office manager who approves supplier invoices”. Not “manage finances”, but “get an invoice approved without chasing anyone”.
- Map the shortest path through that job from the user’s first visit to the moment they get the result. Every screen on that path is in scope.
- List everything else the product could do, then move it to a “later” list. Be strict: if the job can be done without it, it waits.
- Replace features with manual work where you can. An admin can set up accounts by hand, reports can be exported to a spreadsheet, and edge cases can be handled by email for the first hundred users.
- Keep the things you cannot add later cheaply: secure sign-in and access control, careful handling of user data, a data model that can grow, automated deployment, and analytics for the metric in your assumption.
For the invoice-approval example, the scope might look like this:
| In the MVP | Later |
|---|---|
| Upload or email in an invoice | Integrations with every accounting system |
| Route it to one approver, with reminders | Multi-level approval rules |
| Approve or reject from a link or phone | Native mobile apps |
| Export approved invoices to a spreadsheet | Automatic sync and payment |
| Sign-in, roles for approver and submitter, audit log | Single sign-on, custom roles |
| Usage tracking for weekly active approvers | A reporting dashboard |
How long MVP development takes
A focused MVP for one user type and one core job typically takes around two to four months from kickoff to first users. The time splits roughly like this:
| Phase | Typical length | What happens |
|---|---|---|
| Discovery and scope | 1–2 weeks | The riskiest assumption, the one job, the in/out list, technical risks |
| Design | 1–3 weeks | Core screens, data model and architecture, often overlapping with the start of the build |
| Build | 4–10 weeks | Working features in short cycles, reviewed with you as they land |
| Test and launch | 1–2 weeks | Testing the key flows, production setup, monitoring, onboarding the first users |
What makes it longer: more than one platform at launch, several user roles, integrations with other systems (especially ones without good APIs), regulated data such as health or financial records, an AI component whose quality is still unknown, and slow decisions. The last one is the most common. A founder who can answer questions within a day saves more time than any technology choice.
What MVP development costs
MVP cost is mostly people’s time, so the most useful way to compare estimates is in person-weeks, not currency. Rates differ several times over between countries and between freelancers, agencies and in-house hires, but the amount of work for a given scope does not.
Build cost ≈ team size × weeks × weekly rate. A team of three (say, two engineers and a designer who also tests) working for ten weeks is 30 person-weeks. Multiply by the rate of whoever you are considering, and compare quotes on the same scope.
The things that move the number most:
| Driver | Smaller MVP | Larger MVP |
|---|---|---|
| Platforms | A responsive web app | Web plus native iOS and Android |
| User roles | One or two | Several, with different permissions and screens |
| Integrations | None, or one well-documented API | Several systems, payments, legacy software |
| AI | None, or a hosted model for one step | AI at the core, with retrieval, evaluation and tuning |
| Compliance | Ordinary business data | Health, financial or children’s data, audits |
| Design | Clean and clear, using a component library | Custom visual design and animation |
Then there are running costs, small at MVP scale but worth listing: hosting and database, email and SMS services, error tracking and analytics, AI usage if the product calls a model, and app store accounts if you publish mobile apps (Apple charges an annual developer fee, Google a one-time one). Budget, too, for the work after launch, because the first weeks with real users always produce changes.
Choosing the technology
- Choose proven, common tools your future team will know. An MVP is the wrong place to try a new framework.
- Start with web unless the product needs the phone: the camera, location, offline use, push notifications or an app-store presence. A responsive web app reaches everyone and ships updates instantly.
- If you need mobile, consider one cross-platform codebase such as React Native or Flutter, so each feature is written once for iOS and Android. Our mobile development page covers when native is worth it.
- Use managed cloud services for hosting, databases and sign-in, rather than running your own servers.
- Keep the architecture simple: one application and one database is right for almost every MVP. Microservices solve problems you do not have yet.
- Use no-code tools where they fit, for internal admin, landing pages or the whole MVP if the workflow is simple, and know where their limits are before you depend on them.
If your MVP includes AI
AI MVPs carry an extra risk: you do not know in advance whether the model can do the job well enough. Test that first, before building the product around it.
- Try the AI on real data in the first week or two. Run 50 to 100 real examples through a simple prototype and check the results by hand. If the quality is not there, you have learned it cheaply.
- Decide what “good enough” means as a number, and keep a test set to measure it on every change. Our guide to LLM evaluation metrics explains which ones to use.
- Design for the AI being wrong sometimes: let users see sources, correct outputs and send things to a person.
- Track cost per use from day one. A model call that costs cents is fine at MVP scale and may not be at a hundred times the volume.
- Use a hosted model first. Training or fine-tuning your own model is rarely the right first step for an MVP.
If the product answers questions from documents, see how to build a RAG chatbot; if it automates a business process, what AI workflow automation is.
Who should build it
| Option | Works well when | Watch out for |
|---|---|---|
| A technical co-founder or your own team | You have the skills, and want to own the product long term | Hiring takes months; founders building alone burn out |
| Freelancers | The scope is small and someone on your side can manage technical work | Coordination falls on you; gaps between specialists |
| A development partner | You need a complete team now, including design, testing and cloud | Code ownership, handover and whether they push back on scope |
Whoever builds it, settle two things in writing before work starts: who owns the code and the accounts, and how the work will be handed over to your own team later.
After launch: measure, then decide
Launch is the start of the experiment, not the end of the project. In the first weeks:
- Track the metric from your assumption, plus activation (did new users reach the result once?) and retention (did they come back?).
- Talk to users, especially those who signed up and stopped. Numbers tell you what happened; conversations tell you why.
- Fix what blocks the core job quickly, and resist adding features from the “later” list until the core job works.
- Decide on purpose: keep building, change direction, or stop. All three are good outcomes if the decision is made on evidence and early.
Common mistakes
- Building for everyone. Several user types at once means several products, each done badly.
- Skipping the success metric. Without it, every result can be read as encouraging.
- Polishing before validating. Custom design and animation can wait until people want the product.
- Cutting security and data handling. A breach or a lost dataset at MVP stage can end the company.
- Treating the MVP as disposable. If it works, you will build on it, so build it to be built on.
- Launching to nobody. Line up the first users before the build finishes, not after.
Before you talk to a developer
You will get a better estimate, from us or anyone else, if you can answer these:
- Who is the first user, and what job will the MVP help them do?
- What is the riskiest assumption, and what number would prove it?
- Which platform do they need: web, mobile or both?
- Which other systems does it have to work with?
- Does it handle sensitive data, or include AI?
- Where will the first 20 to 50 users come from?
- Who on your side can make decisions quickly during the build?
Conclusion
MVP development for startups comes down to three disciplines: test the assumption that could sink the business, cut the scope to one user and one job, and build that small thing properly enough to grow from. A focused MVP is a matter of months, its cost follows its scope, and the real result is not the software but the decision it lets you make.
We build MVPs for founders and product teams, including AI MVPs, and we have taken our own products from idea to production. See MVP development for what we include and how we run it.