Someone asks what an AI project costs. They get a range — maybe $80,000 to $250,000 for a mid-market production system. They budget for it, the project ships roughly on target, and everyone is satisfied.
Then month twelve arrives. Inference volume has grown. The model has drifted and needs retraining. Nobody owns the MLOps work, so a senior engineer is doing it between other things. The $180,000 build is running at $300,000 to $400,000 by month eighteen, and the person who approved it is working out how to explain that.
Nothing went wrong technically. The budget simply answered a question that was too narrow.
Proof of concept: $15K–$60K build, $20K–$80K first-year total. MVP on a managed API: $40K–$120K. Mid-market production system: $80K–$350K build, $150K–$600K first-year total. Enterprise deployment: $400K–$1M+ build, $800K–$2.5M+ first-year total. The gap between the two numbers in each pair is the part this article is about.
The Four Tiers, With Honest Totals
| Tier | Build cost | Year 1 total | Team and timeline |
|---|---|---|---|
| Proof of concept Single model, synthetic or sample data, no production infra | $15K–$60K | $20K–$80K | 2–3 engineers, 3–6 weeks |
| MVP First AI feature on a managed API, lightweight RAG | $40K–$120K | $70K–$180K | 3–4 engineers, 8–12 weeks |
| Mid-market production Real users, real integrations, monitoring, on-call | $80K–$350K | $150K–$600K | 4–7 engineers, 3–6 months |
| Enterprise Compliance, multi-system integration, governance | $400K–$1M+ | $800K–$2.5M+ | 8–15+ engineers, 6–18 months |
Those ranges are wide because the variable that moves them most is not the AI. It is how clean your data is and how many systems the thing has to talk to.
Where the Money Actually Goes
Across a typical AI project, the budget splits roughly like this:
| Phase | Share of budget | What surprises people |
|---|---|---|
| Discovery and scoping | 5–10% | Cutting it by $10–20K routinely pushes $80–150K of rework into deployment |
| Data preparation and annotation | 20–25% | Most teams budget 5–10% here, and hit overruns by month two |
| Model build and fine-tuning | 15% | This is the part that gets quoted. It is the smallest engineering line. |
| Testing and evaluation | 10% | Includes building the eval set, which most plans omit entirely |
| Deployment and integration | 15% | Scales with the number of systems, not the complexity of the model |
| Maintenance and monitoring | 25–30% | Lands after launch, when the project budget is usually closed |
Read that table twice. Data preparation costs more than building the model, and maintenance costs more than both. Yet the conversation that produces the budget is almost always about the model.
Underestimating data preparation. Teams assume their data is in better shape than it is, allocate 5–10% to cleaning and labelling, and discover in week three that half the records are inconsistent and nobody can explain what three of the columns mean. Poor data preparation can double project cost before a single model is trained — and unlike most overruns, it is entirely knowable in advance if someone looks at the data during discovery.
The Costs That Do Not Appear on a Quote
Four line items account for most of the gap between a quote and a three-year reality:
Post-market monitoring obligations under the EU AI Act add an estimated 15–25% to total project cost over three years for systems classified as high-risk. This is not optional and not a one-time compliance exercise. If your system touches employment, credit, education, essential services or biometrics, work out your classification before you budget, not after.
What Engineers Cost, by Region
| Region | Senior AI/ML engineer, hourly | What you trade |
|---|---|---|
| North America | $150–$280 | Nothing on timezone or communication; highest rate in the market |
| Western Europe | $110–$190 | Similar, with EU data-residency advantages for EU clients |
| Eastern Europe | $55–$110 | Strong engineering depth; partial overlap with US hours |
| Latin America | $45–$90 | Good US timezone overlap; smaller senior AI talent pool |
| South and Southeast Asia | $30–$75 | Lowest rate; overlap and seniority vary considerably by firm |
The honest caveat on that last row, and we say this as a firm that sits in it: the rate band is wide because the quality band is wide. A $30/hour engineer and a $75/hour engineer in the same city are not doing the same work. Rate alone tells you very little — what tells you more is whether the team has run something in production, carried the on-call pager for it, and can show you what broke.
Offshore rates look like a 60–70% saving against North America. In practice the saving is smaller, because a cheaper team that needs two extra months and a rework cycle costs more than a more expensive team that does not. Compare total delivered cost on a defined scope, not hourly rates — and ask both teams for a fixed-scope quote on the same brief to see which one actually understood it.
The Cheapest AI Project Is Often the One You Do Not Build
This section exists because we would rather say it than have a client discover it six months in.
A large share of AI projects that get custom-built would cost a fraction as much on a managed API with a well-designed prompt and retrieval layer. The custom build gets chosen because it feels more serious, or because someone wants the model weights, not because the economics support it.
- · Volume is moderate or unpredictable
- · A general-purpose model handles your task well
- · Your data can leave your environment
- · You have no MLOps capacity
- · Time to market matters more than cost per call
- · You are still learning what the product should do
- · Sustained high volume with predictable load
- · Domain language a general model gets wrong
- · Data cannot leave your infrastructure
- · Latency requirements an API cannot meet
- · The model itself is the differentiator
- · You have measured, not assumed, the volume
For reference: fine-tuning a 7B–13B model costs $2K–$30K, and a single training run on a 70B model runs $2K–$15K. Those are not large numbers. The cost of fine-tuning is rarely the problem. The cost of owning a fine-tuned model for three years is.
How to Get an Estimate You Can Trust
Any firm can produce a number. The question is whether the number survives contact with your actual data and systems. Five things separate a quote that holds from one that does not:
Did anyone look at your data before quoting?
Not asked about it — looked at it. Data condition is the largest single driver of cost variance, and a quote produced without seeing a sample is an informed guess at best. If a firm quotes a production system from a one-hour call, that number will move.
Does the quote include a year of running it?
Inference at projected volume, monitoring, retraining, and who owns MLOps. If those are absent, you have a build quote, not a project budget — and the difference between the two is 40–80% of your first year.
Is there a working prototype before the big commitment?
A prototype on your real data answers the question no proposal can: does this approach work on your problem? The cost of finding out in week one versus month four is the difference between a scoping exercise and a write-off.
What happens when accuracy is not good enough?
Every AI system has a quality threshold below which it is worse than not having it. Ask what that threshold is for your use case, how it will be measured, and what the plan is if the first build lands under it. A firm without an answer has not thought about it.
Who owns the code and the model at the end?
Get it in writing. A system you cannot modify or move is a recurring cost dressed as an asset, and the time to discover that is before signing rather than at renewal.
We build a working prototype on your real data in 72 hours, before any commitment. It is not generosity — the data above explains it. Cutting discovery short pushes $80K–$150K of rework into deployment, and the fastest way to find out whether a project is viable is to try it on the actual data rather than describe it in a proposal. Both sides learn more from 72 hours of building than from three weeks of meetings.
Where to Actually Reduce Cost
Frequently Asked Questions
Want a Number You Can Actually Budget Against?
We build a working prototype on your real data in 72 hours, then quote the full three-year cost — build, inference, monitoring and retraining. Book a free 30-minute call. NDA first, no obligation.