How much does it cost to implement AI in a company (and how a CFO calculates the return)

When a CFO asks "how much does it cost to implement AI", the honest answer is: it depends on what you want to solve. But there is a predictable cost structure, and there is a right way to measure whether the investment pays off. And there is a second question that usually gets left out: how much does it cost you to keep doing the work by hand?

This article brings the two together: where the cost of an AI project lies, which costs are usually forgotten and how to work out, with a simple sum, how long the investment takes to pay back.

The hidden cost of manual work

Before talking about the price of AI, it is worth looking at the price of what you already pay and don't account for.

Picture someone who spends four hours a week copying order data from an email into the invoicing system. That is 16 hours a month. Multiply it by the real cost of that hour — salary, employer costs and the fact that this person isn't doing higher-value work — and a "small" task ends up costing hundreds of euros a month, every month.

Now multiply that by the similar tasks spread across the company. That is the hidden cost. An AI project doesn't create an expense out of nothing: it replaces a recurring expense, which grows with volume, with an investment that needs maintenance.

Where the real cost lives

The cost of a software licence is usually the smallest part. To budget realistically, split the cost into three layers.

1. Software and infrastructure. The AI tools, the services the solution runs on and the use of AI models, which is paid for in tokens. For most SMEs it is the most predictable layer. Each provider's per-token prices are in the AI API pricing comparison.

2. Implementation. Connecting AI to the systems you already have (CRM, invoicing, email) and preparing the data it needs, because a model is only as good as the information it receives. This is where most of the effort lies, and where projects go off the rails when this layer is underestimated.

3. Ongoing operation. Maintaining, adjusting and improving. In the first months any solution needs fine-tuning; then it stabilises, but still needs attention when processes change. Budget a monthly share from the start. The project has an end; the system still needs attention and continuous improvement.

The costs that get left out

For the sum to be honest, add these three as well:

Data clean-up. If the information is messy, part of the project is tidying it up. It is real work and it has a cost, but it serves far more than this automation.

Team training. People need to know how to work with the new system and to trust it. Without that, the team works around the system and the automation ends up abandoned.

Change management. Every automation changes habits. Allow time for the team to adapt and to adjust the solution to what its users tell you.

They don't have to be large, but ignoring them means the return promised on paper doesn't show up in practice.

How to calculate the return without illusions

The most common mistake is to measure the investment and forget the cost of doing nothing. The right sum compares the cost of implementation with the value of the hours freed and the errors avoided. It takes five steps:

Step 1: measure the hours spent today. How many hours a month the task takes, adding up everyone involved.

Step 2: put a value on the hour. Use the total cost of the hour to the company, not just take-home pay.

Step 3: calculate the current monthly cost. Hours times hourly value. That is what you pay today, every month.

Step 4: estimate the saving. Automation rarely eliminates all the work. Decide, with the people who know the process, what share of the hours will no longer be needed, and work out the monthly saving.

Step 5: divide the project cost by the net monthly saving (the saving minus the monthly maintenance cost). The result is the number of months until the project pays for itself.

An illustrative example

The figures below are an example to show the sum, not a market average:

• Task: manual handling of orders and invoicing.
• Hours per month: 16.
• Cost of the hour to the company: 25 euros.
• Current monthly cost: 400 euros.
• Share automated, in the example: 85%, i.e. 340 euros a month saved.
• Project cost: 3,000 euros, plus 100 euros a month for maintenance.
• Net monthly saving: 240 euros.
• Time to pay back: just over 12 months. From then on, the net saving continues for as long as the process does.

The sum doesn't include the value of that person being free to sell, serve customers better or handle what requires judgement.

To run this sum with your own numbers, use the automation ROI calculator, which starts from this same example, or the manual process cost calculator. The AI costs guide brings the rest of the topic together in one place.

What's really worth automating

Not everything should be automated, even if it is technically possible. The best candidates share four traits:

• High volume: it happens many times a week or a month.
• Clear rules: the process is predictable and follows logic you can describe.
• Little human judgement: it doesn't depend on sensitivity, negotiation or delicate context.
• Contained cost of error: an occasional failure is recoverable, not catastrophic.

Tasks that meet all four give the best return with the lowest risk: repeated communication, moving data between systems, generating documents, qualifying leads, answering frequently asked questions. And before automating, simplify: automating a badly designed process only makes it go wrong faster. See the list of use cases by business area.

The first-vector rule

The approach with the lowest financial risk is not to invest in everything at once. It is to pick the process with the heaviest manual load and the highest likelihood of error, automate that one, measure the return against the initial sum, stabilise it, and only then move on to the next.

Automating several things at once looks faster, but it multiplies the risk and scatters attention. One thing at a time means each project funds the next, in savings and in confidence. To measure each stage, see how to measure AI ROI.

The number that matters

Don't just ask how much it costs. Ask how long it takes to pay back and how much it costs to keep doing it by hand. When those two numbers are clear, the decision stops being a matter of faith and becomes a matter of management.

Frequently asked questions

How much does it cost to implement AI in a company?

It splits into three layers: software and infrastructure (including token usage), implementation (integration with systems and data preparation) and ongoing operation. The largest initial share is usually implementation. The right figure is judged against the monthly saving the project generates, not in absolute terms.

How do you calculate the ROI of an AI automation?

Multiply the task's monthly hours by the real cost of the hour to get the current cost. Estimate, with the people who know the process, what share of those hours will no longer be needed. Divide the project cost by the net monthly saving and you get the months until it pays for itself.

What is worth automating first?

High-volume tasks with clear rules, little human judgement and a contained cost of error. Repeated communication, moving data between systems and lead qualification are typical examples.

How long does an automation take to pay for itself?

It depends on the task volume, the cost of the hour and the project cost. In this article's illustrative example, just over a year. The ROI calculator runs the sum with your numbers.

Which costs usually get left out of the budget?

Data clean-up, team training and change management. They don't have to be large, but without them the return on paper doesn't show up in practice.

Is AI automation reliable?

Yes, as long as someone is responsible for following it in the first months and there is a clear metric. Failures come more from implementations with no owner and messy data than from the technology.