How We Price Machine Learning Projects
Last updated:
The model is the cheapest part
Business owners often expect the price of a machine learning project to depend on how sophisticated the model is. It rarely does. Training a gradient boosting model on clean, well-understood data is a small part of the work. The money goes on getting the data into that state, putting the prediction where people will use it, and making sure it keeps working when the world changes.
That is why two projects with the same model type can differ in price several times over. One has a single tidy table and a prediction that lands in an existing report. The other has four source systems, a decade of changing definitions and a real-time integration into an order screen.
The range, and what moves a project within it
Our machine learning service page gives the range we see: roughly £8,000 to £60,000 and above, with most builds taking four to twelve weeks. Where a given project lands depends on a handful of factors.
| Factor | Pushes the price down when | Pushes the price up when |
|---|---|---|
| Data sources | One system, one export | Several systems that disagree with each other |
| Data quality | Outcomes recorded consistently | Definitions changed, gaps, free text where codes should be |
| Labels | The outcome is already in the data | Someone has to label examples by hand |
| Delivery | Nightly batch scores into a table | Real-time predictions inside a live application |
| Interface | Predictions appear in a tool you already use | A new dashboard or review screen is needed |
| Risk and governance | Low-stakes internal decision | Decisions about people, needing explanation and audit |
Notice that model choice is not in the table. Deep learning for images or audio does add cost, mostly through labelling and compute, but for the tabular problems that make up most business ML it is rarely the deciding factor.
What you pay for at each stage
- Discovery and written scope: free for most enquiries. Where the process needs serious mapping first, we propose a paid discovery and say so before starting.
- Proof of value: a short, separately priced piece of work on your real data, fixed before it starts, ending in a go or stop recommendation.
- Production build: a fixed price against a written scope, with payment tied to milestones you can verify, typically 25% at kick-off and the rest against deliverables.
- Support after launch: a 90-day warranty on defects, then an optional monthly retainer for monitoring and retraining.
Staging the payments this way means the largest cheque is only written once the evidence says the model is worth building. It also means you can stop at any stage with something useful in hand.
Fixed price where we can, and honesty where we cannot
We fix-price the production build because by that point the proof of value has removed the biggest unknown, which is whether the data supports the prediction at all. What remains is engineering, and engineering can be estimated.
The part that genuinely resists fixed pricing is open-ended research: trying several fundamentally different approaches to see which one works. We avoid selling that as a fixed build. As we explain in fixed price or time and materials, a fixed price carries a premium because we carry the estimate risk, so it only makes sense for work that can be estimated honestly.
If someone quotes a fixed price for a machine learning build before looking at your data, they are pricing their optimism, not your project.
Running costs people forget to budget
A model in production has costs that a normal feature does not. None of them are large for a typical small or mid-sized business, but all of them are real.
- Hosting and inference: usually modest for batch scoring, higher for real-time predictions or GPU-backed deep learning models
- Monitoring: watching data drift, prediction quality and latency so decay is caught early
- Retraining: refreshing the model as new outcomes arrive, and checking the new version before it replaces the old one
- Language model usage, where a project includes one, billed by the provider directly to your account
As a planning figure, allow roughly 15–20% of the build cost a year for evaluation and maintenance. That is the same guidance we give for other AI work in what an AI project actually costs, and it is the line most often left out of the business case.
How to make it cheaper without making it worse
The reliable savings all come from narrowing scope. Predict for your top product lines before all of them. Deliver nightly scores into a spreadsheet or CRM field before building a dashboard. Keep a human reviewing the model's recommendations in version one, so edge cases do not need to be engineered away before launch.
What does not save money is skipping the evaluation set, skipping monitoring or training on data nobody has audited. Each takes a line off the quote and adds a larger one to the first year. At SpiderHunts we will tell you which of your requirements are driving the price, so you can decide which ones are worth it.
Frequently asked questions
How much does a machine learning project cost with SpiderHunts?
Do you charge for the initial scoping?
Is there a licence fee for the model?
Can you quote in dollars or euros?
What does the ongoing cost look like?
Want a price you can take to the board?
Tell us the decision and the data behind it. Scoping is free, the written scope is yours either way, and the price for each stage is fixed before that stage starts.