An AI consultant has no catalogue price, and anyone who quotes a figure before seeing the problem is guessing. The price depends on five things: how clear the scope is, the state of your data, how many systems need connecting, what a mistake costs, and who runs the system after handover. An honest start is a free first conversation, then a short audit with scope and price fixed upfront.
I have been building AI systems since before ChatGPT went mainstream, and I founded Sapio AI in 2021. What follows is how I think about cost when someone asks me "what would this cost?", without invented rates and without ranges so wide they stop meaning anything.
Why you can't find a clear price online
Search for what an AI consultant costs and you will find hourly and project ranges wide enough to hold both a one-day workshop and a system that runs in production for a year. They are accurate and useless at the same time. A range that wide only tells you the question has been framed wrong.
It is a bit like asking how much it costs to renovate a flat. A serious builder does not answer over the phone. They come round, knock on the walls, look at the wiring, and only then give you a number. And the expensive part is rarely the part you can see. It is what sits behind the plaster. In AI projects, what sits behind the plaster is the data and the systems the solution has to talk to.
What you are actually buying when you buy "AI consulting"
Different things are sold under the same name, and each has its own cost logic.
| What you buy | What you get | What moves the price |
|---|---|---|
| A diagnostic conversation | An opinion: does AI belong in your problem or not | Almost nothing; it should be free |
| An audit and a roadmap | Processes mapped, data checked, what to build, what to buy, what to leave alone | How many processes and data sources are in scope |
| Implementation | A system running in production, integrated, with people trained to use it | Data, integrations, the accuracy required |
| Operation | The system kept alive: data updated, running costs, adjustments | Usage volume and how often the data changes |
Most price confusion comes from comparing different rows of this table: a consultant's quote for a roadmap set next to a firm's quote for a whole system. They are not the same thing, so the prices cannot be either.
The five factors that move the price
1. How clear the scope is
"We want something with AI" and "we want supplier invoices read and entered into the ERP automatically" are projects with completely different costs, even if the technology underneath is the same. The vaguer the scope, the more of the budget goes on finding it. That is why a short audit, paid separately, is usually the cheapest way to buy clarity: you pay a little to find out what the rest costs.
2. The state of your data
This is where most budgets are won or lost. Digital, structured data in one place is one project. The same information scattered across scanned PDFs, spreadsheets in different formats and email threads is another project, a much larger one, in which the model is the small part. A consultant who does not ask about your data before giving you a price is, in effect, pricing someone else's project.
3. How many systems need connecting
An assistant that sits alone on a web page is simple. The same assistant reading from the CRM, writing to the ERP, sending WhatsApp messages and respecting each employee's access rights is a different conversation. Every integration brings authentication, formats, failure cases and testing. In retail, that usually means the catalogue, the orders and the returns policy (AI consulting for retail). It is not only a small-company problem: in a 2025 BCG study (The Widening AI Value Gap, 1,250 executives), 72% named integrating AI with existing systems as a blocker.
4. What a mistake costs
A system that suggests an answer a person then checks can afford to be wrong now and then. One that decides on its own about money, health or someone's rights cannot. The distance between "works in most cases" and "works almost always, and says so when it doesn't know" is a serious share of the cost: evaluation on real cases, safety rules, people in the loop. The same goes for confidentiality. A solution that must run on your own servers, with no internet access, costs differently from one in the cloud. For a law firm, confidentiality is usually the first filter (AI consulting for law firms).
5. Who runs it after handover
An AI system does not end at launch. Data changes, new cases appear, and every question sent to a language model has a running cost. You choose between paying someone to keep it alive and preparing your own team to do it. The second option costs more at the start and less afterwards, provided the people are genuinely prepared (AI workshops for teams). Who owns the system inside the company is an organisational question as much as a technical one (how to structure your company for AI). Always ask for running costs under three usage scenarios (low, medium, high), never a single figure.
A real example: where the work went on ai-aflat.ro
ai-aflat.ro is the free AI assistant I built for Romanian legislation. From the outside it looks simple: you describe a situation in your own words and get the relevant laws back, with the official source cited and linked. It has more than 15,000 users and grew without any marketing budget.
If I had to point at where the work sat, I would not point at the model. I would point at the data: more than 220,000 legislative acts, updated daily, structured so that every answer can send you to the exact source on legislatie.just.ro. And the part that keeps costing after launch lives there too: the daily update, and checking that what the system finds is what it should find. The language model is the piece you can swap most easily.
The lesson applies to any project. When you receive an estimate, ask where the data work is and who does it after launch. If the answer is vague, so is the price.
Hourly, fixed price or retainer?
Each payment model rewards something different, and it is worth knowing what before you choose.
- Hourly. Rewards time spent, not results. It makes sense for short sessions, when you do not yet know exactly what you are asking. On a whole project, the overrun risk stays with you.
- Fixed price. Rewards a clear scope. It works well after an audit, once everyone knows what is being built. Before an audit, a fixed price on a vague project is either inflated to cover the risk or a promise that gets renegotiated along the way.
- Monthly retainer. Rewards availability. It makes sense for operation and continuous improvement once the system runs, not as a way to start.
This is how I work: a fixed-scope audit, then delivery in stages, each with acceptance criteria written in advance, and payment on the accepted stage. If the project does not justify itself after the first stage, you stop there, with something usable in hand.
Why the cheapest option often ends up the most expensive
The visible cost of an AI project is the invoice. The real cost is the project nobody ends up using. An IBM Institute for Business Value study (2,000 CEOs in 33 countries, 2025) found that only 25% of AI initiatives delivered the expected ROI, and only 16% scaled across the business. Most of the reasons are about the process, not the model; I went through them in why AI is not delivering results.
There is another detail that says a lot about where the brake actually is. Among European companies that looked at AI and did not go ahead, 70.3% say the reason was a lack of expertise. Cost comes sixth, at 38.4% (Eurostat, statistical report KS-01-26-009, Table 7, reference year 2025). For most of them, the problem is not what it costs. It is not knowing whom to believe.
A process that eats a day a week of three people's time already has a price. It just never shows up on an invoice.
How to ask for a quote you can compare
Send everyone the same information and ask everyone for the same things back. Otherwise you are comparing apples with pears.
What you send:
- The problem, described by the people who live with it every day.
- The volume: how many times a day or a month it happens.
- Where the data lives and in what format.
- The systems the solution has to talk to.
- The constraints: confidential data, running on your own servers, deadlines.
What you ask for back:
- What is included and, just as important, what is not.
- The stages, with acceptance criteria for each.
- Running costs under three usage scenarios.
- What you own at the end (code, documentation, access).
- The name of the person who will actually do the work.
If you want the full list of questions for choosing the person, not just the price, I wrote it separately: how to choose an AI consultant.
The question before the price
In my view, the conversation about cost changes the moment someone puts on the table what the problem costs them today. A process that eats a day a week of three people's time already has a price. It just never shows up on an invoice, so nobody negotiates it.
So before asking what an AI consultant costs, it is worth asking the other question: how much am I already paying, month after month, to change nothing?
If you want to work through it on your own case: vladtudor.com/consulting.