How much does AI really cost a business? What to look for in the quote

An AI licence is only part of what it costs to bring AI into your business. If you want it to work with your data, the quote also needs to cover preparing that data, connecting the systems the task depends on and the time your people will spend on the change. You need to know what will still require work after handover, too. The balance depends on the job and the condition of your data. Among German businesses already using AI, one in five (21%) calls subscriptions and licences a major cost; one in two (50%) says the same about data preparation, according to a Bitkom survey published on 14 September 2026.

There is no single euro figure that fits every project. You can, however, get to a useful estimate. Take ten real cases, do the task you hope to give AI by hand, and note every point where you get stuck. Then ask for a quote that answers five questions: what work is needed, who will do it and in how many days, what continues after handover, how people’s jobs change, and what assumptions sit behind the estimate.

What goes into the cost of AI

Bitkom, Germany’s digital industry association, interviewed 603 companies with at least 20 employees by phone in July and August 2026. It asked those already using AI which items were a major cost (“großer Kostenpunkt”):

• infrastructure (cloud and computing power): 51%;
• data preparation: 50%;
• integration with existing systems: 41%;
• subscriptions and licences for AI software: 21%;
• external consultancy: 21%;
• staff training: 19%;
• token consumption (tokens are the units a model uses to process the text it reads and writes): 8%, with another 11% calling it a minor cost.

These are percentages of businesses, not shares of their budgets. The 50% figure means that half the businesses using AI consider data preparation a major cost. It does not mean they spend half their AI budget on it. The press release gives no amounts in euros.

What counts as infrastructure depends on how you use AI. Buy it as a service and the provider runs the systems behind it; you pay for them through that service. Run models on your own systems and you need to account for the infrastructure separately. Either way, someone has to prepare the business information AI will use. If AI must also read from or write to your existing software, someone has to make those connections. Ask what the service covers and what falls to you.

What to look for in an AI quote

Once you have an offer in front of you, start with two checks. They reveal what is included and what you will need to arrange yourself. The supplier does not have to present the quote in a particular format.

A line for the data work. Alongside licences, setup and training, look for an item that says which data needs attention, who will handle it — the supplier, your team or both — and how many days it will take. If you cannot find one, ask why. Perhaps the work is unnecessary, already done or included elsewhere. Find out before you sign, while you can still make the scope and responsibilities clear. If the model will read documents containing customer or personal information, ask where those documents will be processed.

All seven categories. Go through Bitkom’s list, one item at a time. What is included? What will the supplier do, and what will you do? Then distinguish work done once from the activities and spending that will continue.

We have an interest to declare: preparing data is part of our work, too. The advice applies whoever does it. You can check an item that names the tasks and their owners; a broad heading leaves you guessing.

Do you actually need AI for this?

Define the job before you price a model. If a fixed rule applied to data you already have will answer the question — adding amounts or comparing dates, for instance — a report may be enough.

“Which customers pay late, and by how many days?” is one such question. You need invoice due dates, payment dates, each payment matched to its invoice and a rule for part-payments. If you have those records, your business software has what it needs to answer. If the matches are missing, completing them comes first. A model cannot reliably supply a link you never recorded.

AI may help when information arrives in forms your software cannot read on its own: an email, a PDF or a document laid out differently by every customer. A model can identify who is ordering, which products they want and how many, then produce a draft for you to check. That can save retyping. You still need to know how to match what the model reads to the records in your system.

What “preparing the data” means in practice

Imagine orders arriving by email as PDFs, each on the customer’s own form. This is an illustrative scenario, not a particular client project. At present, someone reads each order and enters it into your business management system, or ERP. You would like AI to prepare a filled-in draft for a person to check and confirm.

The model has to recognise the customer and what they are ordering. That turns a vague question about data quality into practical ones:

• product codes: the customer uses their own code, or a description, while your system uses yours. You need a table linking them and someone to keep it current;
• customer records: one customer may appear twice under slightly different names. Someone in accounts must decide which record to keep, or an order could end up against the wrong one;
• units of measure: one customer orders by the box, another by the piece. Define the conversion rather than asking the model to guess;
• exceptions: an unmatched code or an unusual quantity needs a review. Decide who handles it and where the order waits in the meantime.

The draft then has to reach the right place in your ERP. That might be through a file import or a direct connection; decide with whoever looks after the system. Data preparation and integration meet at this point. Product codes, customer records and units of measure help the model understand an order and allow the system to record it correctly.

Bitkom’s president, Ralf Wintergerst, describes what agents — AI programs that carry out tasks independently — require: “Ein Agent, der eigenständig handelt, braucht verlässliche Daten, verbundene Systeme und Regeln, was er darf und was er zu lassen hat.” An agent acting on its own needs reliable data, connected systems and rules about what it may and may not do.

How to estimate the data work

To estimate the work, try doing it first. Take ten or so real cases from last month, perhaps ten orders, and work through the task you hope to give AI. One pass will do if you write down every interruption: a product code you cannot find, a colleague you need to ask, a spreadsheet you have to open. Think of it as a manual trial run.

It will not give you a figure on the spot. It will give you the list of work that figure must cover. Use your notes to ask for a quote that sets out:

1. what needs doing to the data and the connections: matching codes, removing duplicates, defining conversions and deciding how the draft gets into your ERP;
2. who will do it and how many days it will take the supplier and your team, including whoever looks after your systems. Internal time is a cost even when it does not appear on an invoice;
3. what continues after handover: who updates the code table when a new product arrives, who checks exceptions each day and who maintains the connection when your ERP is updated;
4. how the job changes for the person who enters orders today, and who will explain their new tasks;
5. what the estimate assumes about monthly volumes, document formats and exclusions.

Pause at point 4. In an article published on 20 January 2026, BCG recommends devoting 10% of the effort to algorithms, 20% to technology and data, and the remaining 70% to people and processes. That is advice for allocating effort in broader transformations, not a measure of what your project should cost. It is a useful prompt to ask where the quote allows for changes to people’s work.

Do you need to clean up all your data first?

No. In the same article, BCG says that, in its experience, waiting for perfect data “often leads to unnecessary delays”.

Focus on the data your chosen task needs. For incoming orders, that means product codes, customer records and units of measure. Other data can wait until a project needs it. This gives the work a manageable boundary: you know what to prepare now and what to revisit if the task changes.

What you keep when you pay for the data work

This is our reading of the situation, rather than a survey finding. You pay a subscription for as long as you use a service. You can change model or provider, but the switch brings work of its own: moving connections, repeating tests and reviewing the contract.

Work on your own data may last longer. A table linking product codes, customer records without duplicates and one trusted record for each customer remain useful if someone keeps them up to date. You can reuse them in a project without AI, and you will need them again if you change AI provider.

Where you keep the information matters as well. If orders, customer records and products already sit in one ERP system, such as Odoo, you may have fewer connections to build. The software’s name alone cannot tell you how much work that saves. The answer depends on the state of your data and what you want the system to do.

Italy and Germany: the same obstacle

We have not found a recent public survey of AI costs among Italian businesses already using it that is comparable to Bitkom’s. There are, however, official figures on barriers to adoption. Istat, Italy’s national statistics office, and Destatis, Germany’s Federal Statistical Office, collect them through the same European questionnaire harmonised by Eurostat.

According to Istat, 15.7% of Italian small and medium-sized enterprises with 10 to 249 employees used at least one AI technology in 2025. Among businesses with at least 10 employees that do not use AI but have considered it, 45.2% cite “the unavailability or poor quality of the data needed” as an obstacle. For the same question and year in Germany, Destatis reports 44%.

There is a clear limit to the comparison. These surveys cover different businesses from Bitkom’s and do not track the same companies over time. They cannot show that a business which overcomes the data obstacle will then face the costs Bitkom records. They do show data appearing in both sets of answers: as a barrier for businesses that considered AI without adopting it, and as a major cost for some already using it.

Where to start

Choose a task and check whether it needs AI or whether a report can answer the question. If a model would help, work through ten real cases by hand. Write down where you get stuck, then ask for every piece of data work in the quote to have a description, an owner and an estimate of the days required, including your team’s time.

Our AI integration page explains how we approach these projects. If the trial run leaves you with a question, get in touch and tell us which step held you up. Before deciding, clarify that step and check that the estimate includes the work it requires.

Gitogi Editorial Team