How much does it cost to build an AI MVP?
Short answer
The cost of an AI MVP is mostly people time: the engineer-months it takes, multiplied by what each month costs you. Model fees and hosting are usually a small share at MVP volumes, because a pilot makes thousands of model calls, not millions. In the worked examples below, an API-based MVP comes to about £47,000 and a custom-model MVP to about £153,000, on stated assumptions you should replace with your own.
By Dr. Dulan Dias, co-founder and CTO of Tellme AI
Key facts
- As of 4 November 2026, Anthropic listed Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens.
- Anthropic's Batch API charges 50% less for input and output tokens on requests that can be processed asynchronously.
- On 4 November 2026, OpenAI's pricing page listed gpt-6-luna at $0.10 and gpt-6-astra at $10 per million input tokens at standard short-context rates.
- As of November 2026, registering a limited company online with Companies House costs £100.
- As of November 2026, the UK data protection fee is £52 for tier 1 organisations (turnover up to £632,000 or no more than 10 staff) and £78 for tier 2.
- As of November 2026, Innovate UK's Smart grants are closed to new applications.
- Tellme AI's production MVP, covering front end, backend, admin portal, CI/CD and AWS infrastructure, was built end to end within 4 to 6 months.
- From Model to Product, Chapter 15, treats cost as a technical constraint: every inference, API call, GPU and database query has a cost that multiplies as usage grows.
Definitions
- MVP (minimum viable product)
- The smallest version of a product that real users can rely on to test whether it solves their problem. For an AI product it includes enough evaluation and monitoring to show whether the AI part works.
- Engineer-month
- One engineer working full time for one month. Engineer-months multiplied by a fully loaded monthly cost give the team line of a budget.
- Fully loaded cost
- What an employee costs the business rather than their salary alone: salary plus employer's National Insurance, pension contributions, equipment and software. Contractors are usually costed by day rate instead.
- Inference cost
- What you pay each time a model produces an output. Hosted providers charge per million tokens, with separate prices for input and output.
What makes up the cost of an AI MVP?
An AI MVP has six cost lines. People are usually most of the total; the others matter because they are the ones founders forget or underestimate.
| Cost line | What it covers | How to estimate it |
|---|---|---|
| Team | Engineering, product and design time to build, test and launch | Engineer-months × fully loaded monthly cost |
| Model and inference | API fees during development, evaluation runs and the pilot | Calls per month × tokens per call × provider price |
| Infrastructure | Hosting, database, storage, logging and error tracking | Monthly bills for the services you choose |
| Data | Collecting, cleaning, labelling and licensing data; building an evaluation set | Days of expert time plus any licence fees |
| Compliance | Company registration, data protection fee, DPIA, contracts, security review | Fixed fees plus days of internal or legal time |
| Contingency | Rework when the first approach misses the quality bar | A share of the team line, larger when the problem is new |
Worked example 1: an API-based document assistant
A B2B startup builds an assistant that answers staff questions from a client's policy documents, using a hosted model and retrieval over those documents. The total comes to about £47,000, and about 96% of it is people. Every figure is an assumption for illustration, not market data; replace each one with your own.
- Two engineers for three months (6 engineer-months) at an assumed fully loaded £7,500 per engineer-month. Founder time is not costed.
- A pilot of 50 users asking 20 questions each working day, about 20,000 questions a month.
- Each question sends 3,000 input tokens (instructions, retrieved passages and the question) and receives 400 output tokens.
- Claude Sonnet 5.5 at its listed $2 and $10 per million input and output tokens, checked on 4 November 2026: $0.010 a question, or $200 a month at pilot volume.
- Hosting, a managed PostgreSQL database with pgvector, logging and error tracking at an assumed £300 a month.
| Line | Basis | Three-month cost |
|---|---|---|
| Team | 6 engineer-months × £7,500 | £45,000 |
| Model fees | An allowance of $300 a month for development, evaluation runs and the pilot | $900, under £1,000 |
| Infrastructure | £300 a month | £900 |
| Fees | Online company registration £100; tier 1 data protection fee £52 | £152 |
| Total | About £47,000 |
Worked example 2: a custom vision model with an app around it
A startup builds a phone app that recognises items from the camera with its own trained model, then answers follow-up questions with a small language model. Custom models bring data work and a longer timeline, so the total is about £153,000, and people are still about 95% of it. Again, every figure is an assumption.
- Three engineers for six months (18 engineer-months) at the same assumed £7,500.
- A domain expert spends two months collecting, labelling and reviewing images and writing evaluation cases, at an assumed £5,000 a month.
- An assumed £2,000 of cloud GPU time for training experiments.
- 5,000 follow-up questions a month at 1,500 input and 200 output tokens on Claude Haiku 5.5, listed at $0.10 and $0.50 per million tokens: about $1.25 a month.
- Hosting at an assumed £400 a month, and an assumed £3,000 for a one-off legal review of terms, privacy notice and the data protection impact assessment.
| Line | Basis | Six-month cost |
|---|---|---|
| Team | 18 engineer-months × £7,500 | £135,000 |
| Data and evaluation | 2 months of expert time × £5,000 | £10,000 |
| Training compute | Assumed cloud GPU budget | £2,000 |
| Language model fees | About $1.25 a month at pilot volume | Under £10 |
| Infrastructure | £400 a month × 6 | £2,400 |
| Legal review and fees | £3,000 review; £100 registration; £52 data protection fee | £3,152 |
| Total | About £153,000 |
What moves the number most
The model price is rarely what makes an AI MVP expensive. These are the drivers that do, roughly in order of impact:
- Scope. One workflow for one type of user costs a fraction of a platform, and every extra integration, role or channel adds engineer-weeks.
- Custom models. Training your own adds data collection, labelling, experiments and a longer path to acceptable quality. Start with a hosted model unless a measured gap rules it out; rules, machine learning or an LLM helps decide.
- Data. Messy, scattered or unlicensed data can take longer to fix than the product takes to build, and evaluation sets need expert time.
- Integrations. Connecting to legacy systems, single sign-on or client data sources is often the slowest work in a B2B MVP.
- Regulated domains. Health, finance, hiring and education bring sector rules and more documentation, and products sold into the EU for uses such as hiring or credit face the EU AI Act's high-risk obligations from 2 December 2027. See UK AI regulation for startups.
- Who builds it. A founder who can build, an in-house team, freelancers and an agency have very different cash costs and very different ownership afterwards; do you need a technical co-founder weighs the options.
Model and inference costs: small now, central later
At MVP volume, model fees are a rounding error in both examples: $200 a month for a mid-tier model answering 20,000 questions, about $1.25 for a small model answering 5,000. The spread between models is wide, though. On 4 November 2026 Anthropic's prices for current models ran from $0.10 per million input tokens for Claude Haiku 5.5 to $10 for Claude Fable 5.1, and OpenAI listed gpt-6-luna at $0.10 and gpt-6-astra at $10.
That changes after launch. If the first example grows from 50 users to 5,000 with the same usage, the model line becomes $20,000 a month. Design for that before it arrives: route easy requests to smaller models, cache repeated answers, trim prompts and use batch pricing for work that can wait. LLM cost per user gives the formula and the levers.
Chapter 12 of From Model to Product puts the principle plainly: "If the system works but cannot be paid for, the architecture is wrong." Chapter 15 extends it, treating cost as a technical constraint that multiplies with every request.
From practice: how long an AI MVP takes
Time is the hardest line to estimate, so one real data point helps. As founding engineer at Tellme AI, Dr. Dulan Dias built the production MVP end to end within 4 to 6 months: the Next.js front end, the FastAPI and MongoDB backend, an admin portal, CI/CD and the AWS infrastructure. Tellme is now live at three national museums.
Two lessons carry over. One engineer who can work across the whole stack removes coordination cost, so the team line of a small MVP can be smaller than a headcount plan suggests, at the price of depending on that person. And production work such as deployment pipelines, an admin portal and monitoring is part of an MVP that real users rely on, not an extra to budget later.
How to spend less without building something worse
Most savings come from doing less, not from cheaper inputs:
- Write down the one job the MVP must do and the quality bar it must meet, and cut everything else.
- Try rules, search or an existing product first, and build only the part nothing else does.
- Use a hosted model through an API before considering training or self-hosting.
- Start retrieval in the database you already run, and add a vector database only when measured load requires it.
- Build a small evaluation set from real cases in the first week, so quality debates are settled by data.
- Use managed services and one cloud region, and skip Kubernetes and microservices until a boundary is real.
- Set spend limits and per-user caps on model calls from the first deploy.
- If you prototyped with AI coding tools, budget for the hardening the prototype skipped; taking a vibe-coded app to production lists what that involves.
Frequently asked questions
Can you build an AI MVP for under £10,000?
In cash terms, yes, if a founder builds it and their time is unpaid. In the first worked example, the cash cost without salaries is model fees, hosting and fees: under £2,000 over three months. The real cost is the builder's time, and the risk is that production work such as security, monitoring and data protection gets skipped to stay inside the budget.
How long does it take to build an AI MVP?
It depends on scope, data and who builds it. The worked examples on this page assume three months for an API-based product and six for one with a custom model; those are assumptions, not benchmarks. One real data point: Tellme AI's production MVP, including front end, backend, admin portal, CI/CD and cloud infrastructure, was built end to end by its founding engineer, Dr. Dulan Dias, within 4 to 6 months.
Is an agency cheaper than hiring engineers for an AI MVP?
Sometimes for the build, rarely for what follows. An agency can start quickly and absorbs hiring risk, but you pay its margin, and knowledge of how the system works can leave when the contract ends. AI products need continuing evaluation, prompt and model changes and cost tuning after launch, so whoever builds the MVP should either keep running it or hand it over with tests and documentation.
How much should I budget for LLM API fees while building?
Less than most founders fear. Estimate calls per month, tokens per call and the provider's price; in the first worked example, 20,000 questions a month on a mid-tier model cost about $200. Add an allowance for evaluation runs and experiments, set a hard monthly spend limit with your provider, and log tokens per call from the start so the estimate becomes a measurement.
Does R&D tax relief cover the cost of building an AI MVP?
Only the parts that qualify as research and development. HMRC guidance requires a project to seek an advance in science or technology by resolving scientific or technological uncertainty. Wiring a hosted model into a standard product will not usually qualify; novel technical work might. Check the current HMRC guidance and take advice from an R&D tax specialist before counting on it.
Are there UK grants for building an AI MVP?
Fewer than older articles suggest. As of November 2026, Innovate UK's Smart grants, its open grant competition, are closed to new applications, and its remaining funding runs through specific competitions and programmes listed on its funding pages. Check those for anything that matches your sector. How a UK small business should start using AI lists the wider support that exists.
What hidden costs do AI MVPs have?
Four come up repeatedly: building and maintaining an evaluation set, cleaning the data the product depends on, the security and data protection work needed before real customer data goes in, and monitoring once users arrive. None of them show in a demo, and all of them are needed before an AI product can be trusted. Budget them as part of the MVP rather than as phase two.
Sources
About the author
Dr. Dulan Dias is co-founder and Chief Technology Officer of Tellme AI in Belfast; founder of EchonLabs, grown from 2017 to more than 60 people before its acquisition in 2026; holder of a PhD in AI and data science from Queen’s University Belfast; and the author of From Model to Product.