AI project pricing varies wildly with scope. Here is how a realistic budget breaks down — integration, data, maintenance — and the questions to ask before you sign.
The question that stalls more AI projects than any technical obstacle is a simple one: what will this actually cost? Public answers tend to be either vague ("it depends") or misleading (a subscription price that hides months of integration work). Neither helps a small or medium business decide. So let us be concrete about how AI project budgets are actually built.
The price of an AI project is a scope question, not a technology question
There is no market rate for "an AI project" any more than there is for "a website." A chatbot answering questions from a fixed FAQ and a system that reads thousands of supplier invoices per month are both called AI, and their budgets differ by an order of magnitude.
What drives cost is almost always the same short list:
- Data readiness. Is your information already structured, accessible and clean, or does half the project consist of collecting and organising it?
- Integration surface. Does the AI need to plug into your CRM, ERP, email or internal database — and do those systems expose usable APIs?
- Ambition of accuracy. A tool that suggests answers to a human reviewer is far cheaper than one that must be right autonomously.
- Number of workflows. One well-defined use case costs far less than five half-defined ones.
A useful exercise before requesting any quote: write one paragraph describing what a person does today, step by step, and which of those steps the AI should take over. If you cannot write that paragraph, the project is not ready to be priced — and any consultant who quotes confidently without asking for it is guessing.
Where the budget actually goes
A realistic breakdown for a small or mid-sized business looks roughly like this.
Discovery and design. A short phase to map the workflow, choose the model, and define what "working" means in measurable terms. Skipping it is the most common cause of overruns later.
Integration. Connecting the model to your tools and data. In practice this is often the largest line item — frequently more than the AI component itself. If your systems have no APIs, expect this to grow.
Model and infrastructure. API usage from providers such as Claude, OpenAI or DeepSeek, or hosting an open-source model yourself. For most SMB use cases, API calls are a modest monthly cost; self-hosting only makes sense with high volume, strict data-residency needs, or heavy customisation.
Evaluation and testing. Building a set of real examples and checking the system against them. This is what separates a demo from something you can rely on.
Maintenance. Models are updated, prompts drift, your data changes, your business rules evolve. Budget for it from day one rather than treating it as a surprise.
The honest pattern: the visible "AI" part is usually the smaller share, and the surrounding engineering is the larger one.
Questions to ask before you sign
- Which use cases are included, and which are explicitly out of scope?
- What happens to the price if my data turns out to be messier than expected?
- Who owns the prompts, the code and the evaluation set?
- What does a typical month of maintenance cover, and what would be billed separately?
- How will we measure whether this is working — in numbers, not impressions?
- What is the exit plan if we stop?
A consultant who answers these clearly is worth more than one who quotes lowest. For a freelance consultant, as for an agency, price should reflect scope and risk, not a discount race.
Start smaller than you think
The most reliable way to control cost is to shrink the first version. Pick one workflow, one audience, one measurable outcome. Ship it, measure it, then decide whether to expand. A modest first project that works teaches you more — and costs less — than a grand platform that never quite finishes.
If you are weighing an AI project and want a realistic view of scope and budget before committing, get in touch and we can talk through what your situation actually requires.