Founders ask me what an AI MVP costs and they want one number. The honest answer is that it ranges from 30,000 euros to 250,000 euros depending on five variables, and the variables are predictable. Once you know which bracket you are in, you can negotiate, plan, and avoid the most expensive surprises. Here is the real cost picture as of 2026.
The five variables that determine the bracket
- Workflow scope. One narrow named workflow versus three to five connected workflows.
- Data preparation. Clean structured data versus unstructured PDFs and scanned documents.
- Integrations. Standalone web app versus deep CRM, ERP, or industry-specific integrations.
- Compliance. Standard SaaS versus HIPAA, SOC 2, GDPR sensitive, or industry regulated.
- Eval rigor. Manual spot checks versus automated eval harness with regression suite.
Score yourself honestly on the five and you can predict the bracket within 20 percent.
Bracket 1. The narrow MVP. 30,000 to 60,000 euros, 6 to 10 weeks
One workflow, one customer surface, foundation model API, lightweight retrieval. Examples I have shipped at this size include a draft generator for compliance memos, a meeting summary tool tied to a CRM, an internal search assistant for a 20,000 document corpus.
Line items at this bracket. One senior product owner part-time across the engagement, 12,000 to 20,000. Two engineers for six to eight weeks, 18,000 to 30,000. API and infra during build, 1,000 to 3,000. Eval harness, 2,000 to 4,000. Design, 1,500 to 4,000. Total runway including a small contingency, 35,000 to 65,000 euros, with the high end almost always reflecting underestimated data work.
Ongoing cost after launch lands at 800 to 3,000 euros per month for hosting, model API, and observability for a customer base under 500 users. Add an engineer at 25 percent capacity for maintenance. The MVP is not free after launch and founders forget this every time.
Bracket 2. The integrated MVP. 60,000 to 120,000 euros, 10 to 16 weeks
Two to three connected workflows, integration with at least one customer system, light multi-tenant. Examples at this size from my work include the Sainni real-time sales coaching MVP and the early ConstructionOS proposal engine.
The cost driver here is integration. Connecting cleanly to a Salesforce, a HubSpot, an SAP, an industry-specific dispatch system, or a healthcare EHR adds 15,000 to 40,000 euros depending on the API surface and the customer's IT cooperation. Customers always say their integrations are easy. Plan for 1.5 times the estimate they give you.
Line items. Senior product owner full-time, 25,000 to 40,000. Three engineers for 10 to 14 weeks, 50,000 to 80,000. API and infra during build, 3,000 to 6,000. Eval harness with regression suite, 5,000 to 10,000. Design and front-end polish, 5,000 to 12,000. Compliance review if needed, 3,000 to 8,000. Total 70,000 to 130,000 euros.
Bracket 3. The platform MVP. 120,000 to 250,000 euros, 14 to 24 weeks
Multi-tenant from day one, three to five connected workflows, deep integrations, regulated data, custom evaluation tooling. This is the bracket for most enterprise AI MVPs and for startups whose MVP is itself the company's first paid product.
The honest version of this bracket needs a small full team. Senior AI product manager or AI architect leading from end to end. Two to three product engineers. One platform or infra engineer. One designer. Part-time legal or compliance. The team runs for four to six months and ships a real product, not a demo.
Line items. Leadership, 50,000 to 80,000. Engineering, 90,000 to 150,000. Infra and API during build, 8,000 to 18,000. Eval and observability tooling, 10,000 to 25,000. Design and front-end, 10,000 to 25,000. Compliance and legal, 8,000 to 25,000. Total 175,000 to 320,000, with most projects landing in the 140,000 to 220,000 band when scope is held.
The hidden cost lines founders miss
Five line items show up late and blow the budget if not surfaced early.
Data preparation
If your data is in PDFs, scanned documents, voice recordings, or photo evidence, the work to prepare it for retrieval is real. OCR, layout extraction, transcription, deduplication. Allocate 10 to 25 percent of the build budget to this and reduce as you discover otherwise.
Eval and labeling
Building 50 to 200 high-quality labeled examples for the eval harness takes a domain expert 20 to 60 hours. If the founder thinks this can be skipped, the project will fail later. Pay for this up front.
Customer integration support
If your MVP integrates into a customer system, the customer's IT team will not move at startup speed. Two to four weeks of slack across the build is normal. Either pay for a customer success engineer or accept the timeline.
Token cost surprises
Founders estimate token cost based on a happy-path interaction. Real usage has retries, multi-turn conversations, retrieval prompts, and embeddings. The realistic per-active-user cost in 2026 is 0.50 to 4.00 euros per month for chat-style products and 0.05 to 0.50 per workflow for batch or pipeline products. Estimate, then double, and use that for pricing.
Observability and on-call
You cannot operate an AI product without logs of every prompt, every response, every retrieval hit, and every user feedback signal. The tooling is 200 to 1,500 euros per month at startup scale, plus the engineering time to instrument it. Skip this and you will not know why the product broke when it breaks.
The cost of underspending
The most expensive AI MVP is the one that ships at half the budget needed for the bracket it actually belonged in. The 40,000 euro version of a 120,000 euro project does not get you 33 percent of the value. It gets you a brittle demo that the customer will not adopt and that you cannot extend. The right answer is to either fund the bracket properly or descope into a smaller bracket where the budget works.
How to use these numbers
Take your scope. Score it honestly on the five variables. Pick the bracket. Add 15 percent contingency. Negotiate from there. If a vendor or a fractional team quotes you 50 percent below the bracket, ask which line items they removed and what the implication is. There are no free AI MVPs. There are honest budgets and there are budgets that fail later.
If you want a second opinion on a quote you have received or a budget you are sizing, write to me. I respond within 48 hours.