Artificial intelligence consulting: what you really buy, what it costs and how to choose who to trust

I see the scene at least once a month. A business owner comes back from a trade fair with three quotes to "put artificial intelligence in the company". The first offers an assistant that answers customers, the second a 40,000 euro platform, the third a course for all employees. None of the three asked how much time the order office loses today.

Serious artificial intelligence consulting starts from a measurable problem and not from a technology: in two to four weeks it finds one or two processes where AI saves hours or errors, in six to eight weeks it proves it with a pilot on your data, and for an SME it usually costs between 3,000 and 25,000 euro depending on how far you go. This guide explains what the consultant does, what it costs, how to choose one and when to skip it.

What does an AI consultant really do?

An AI consultant does three jobs in one: picks the right problem, proves on your data that AI solves it, and designs how the result enters daily work without depending on the consultant. A good one also tells you when AI is not needed: in about one case out of three, a rule in twenty lines of code or a changed procedure works better. To understand the tool first, read how language models really work.

Which problems does AI solve today?

Projects that work in SMEs are almost always of four kinds: reading documents, searching company information, preparing drafts and routing requests.

Technical background: building a RAG pipeline, hybrid search, AI hallucinations and RAG, AI agents in .NET.

How do you choose the first project?

The first project is chosen with two numbers: what it is worth per year if it works and how bad it is when it fails. Start from high value and low damage. Example: 1,200 email orders a month at 6 minutes each are 120 hours, 43,200 euro a year at 30 euro an hour. If the system removes 60%, that is about 26,000 euro a year, with a person confirming every proposal.

The four phases and the first 90 days

Well-run consulting has four phases with an exit after each: assessment, pilot, production and handover. The assessment takes one to two weeks and ends with a short document: candidate processes, state of the data, proposed technique, success criteria as numbers. The pilot takes four to eight weeks on a narrow scope. Production takes three to six weeks and is 70% of the real work: integration, permissions, logs, cost controls. The handover makes the consultant replaceable.

How this fits with the rest of company software is described in software consulting for companies.

What does it cost?

For an SME an assessment usually costs 2,000 to 5,000 euro, a pilot 8,000 to 25,000 euro and production 10,000 to 40,000 euro, plus a running cost of 50 to 800 euro a month. These are my estimates, based on daily rates between 600 and 1,200 euro. In the order example the project pays for itself in about eighteen months. If a vendor promises payback in three months, ask for the spreadsheet.

How do you measure and how do you choose the consultant?

You measure with a test bench: at least one hundred real cases already solved by people, held out, on which the system is queried and compared with the right answer. A vendor shows five chosen examples. A consultant asks for two hundred random cases and brings back a percentage.

Ask at the first meeting: a project where AI was advised against, a system still in use today, where your data will go, who will own the code, what it will cost per month afterwards, and at which point you can stop without wasting what was spent. Warning signs: starting from technology, demos on their own data only, savings promised before seeing the process, training a model from scratch for a document-reading problem.

Data, privacy and what must stay

You need less data than you think and more order than you hope. Privacy is governed by three choices: where the model runs, what you send to it and what the contract says. The EU AI Act (Regulation 2024/1689) has been in force since 1 August 2024 and applies in stages. Most uses in this guide carry low regulatory risk. Have your lawyer check current dates.

At the end six things must stay with you: source code, model instructions, test bench, accounts in your company name, a dashboard with cost and quality, and an internal owner. Related reading: refactoring legacy code with AI and adopting AI in a development team.

Your step for Monday

Write down the three most repetitive jobs in your office that run through text or documents. For each one note four numbers: times per month, minutes each, cost of a mistake, whether past cases exist. If one job is worth more than 15,000 euro a year and a mistake can be stopped by a check, you have a candidate. If none is, you have saved the cost of a consultancy. To talk it through, start from software consulting for companies or contact me.