How much does it cost to automate a process with AI? Cost items and payback

There is no price list that fits everyone. The cost of AI automation depends on how complex the process is, how many systems need connecting, how much human review is needed and how much it will be used, because AI models are usually billed by usage. There is, however, a right way to do the maths: add up first-year costs, one-off and recurring, and compare them with the value of the time the automation frees.

The cost items at a glance

Item	When you pay	What it depends on
Process analysis	At the start	Steps, exceptions, people involved
Development and configuration	At the start	Rules and cases the system has to handle
Integrations	At the start	Systems to connect and the quality of their APIs
Preparing data and content	At the start	The state of your documents and source information
Testing and security review	At the start and at every major change	Data handled and actions the system can take
AI model usage	Monthly, by usage	Requests, length of texts, minutes of conversation
Infrastructure and tools	Monthly	Servers, databases, monitoring, any licences
Maintenance and evolution	Monthly or as a package	How often things change and the response times required
Human review	Always	Share of cases a person has to check

One-off costs

• Process analysis. Describing who does what, with which data and where exceptions come from. It is the item that most affects all the others: a misunderstood process is paid for twice.
• Development and configuration. Rules, instructions for the model, error handling, screens for the people who review.
• Integrations. Connecting business software, email, CRM, phone. A system with documented APIs is cheap to connect; one without may need custom work.
• Data and content. Price lists, procedures and documents need tidying before they are given to the system. If they are scattered or contradictory, this item grows.
• Testing and security. Test cases with expected answers, checks on permissions and allowed actions. We cover this in the guide to prompt injection.

Recurring costs, the ones quotes forget

• AI model usage. Text is billed by the amount processed, voice by the minute. Ask for an estimate based on realistic volumes and a spending cap: unbounded consumption is one of the ten risks OWASP lists for applications built on language models.
• Infrastructure and tools. Servers, databases, activity logs, any licences for automation platforms.
• Maintenance and evolution. Connected systems change, models are updated, the process evolves. Automation without maintenance stops working silently.
• Human review. The time of the people who check uncertain cases is a real cost and belongs in the calculation.

What pushes the cost up, and what brings it down

It goes up with:

• many exceptions and unwritten rules;
• systems without APIs or with poor-quality data;
• irreversible actions, such as payments or messages to customers, which need extra checks;
• personal or confidential data, which need dedicated protection measures.

It comes down with:

• a process that is already described and stable;
• a single system to connect at the start;
• a limited goal, extended after measuring;
• accepting that part of the work stays with people.

How to calculate the return

Three formulas are enough for a first estimate:

1. Hours freed in a year = people × hours per week × working weeks × automatable share
2. Yearly value = hours freed × cost of one hour of work to the company
3. Payback time in months = investment ÷ (yearly value ÷ 12)

An example with numbers chosen only to show the calculation, not a market average: three people each spend five hours a week on the same task, for 46 weeks a year. That is 3 × 5 × 46 = 690 hours a year. If the software can take over 60% of it, the hours freed are 414. At a cost of €30 an hour to the company, the value is €12,420 a year. An investment up to that amount pays back within twelve months; one of half that amount pays back in about six.

You can redo the calculation with your own numbers in the ROI calculator: everything stays in your browser.

What the calculation does not see

• Recurring costs. Subtract them from the yearly value before working out payback.
• How the freed time is used. The value exists if those hours go to useful work: sales, support, the backlog.
• Indirect benefits. Fewer transcription errors, faster answers, service outside office hours. They are real, but harder to measure.
• Uncertainty about the share. The automatable share is an estimate: you verify it in the first weeks, by measuring the cases the system closes on its own.

Eight questions for reading a quote

1. What is included and what is excluded?
2. Which volumes is the usage cost based on, and what happens if they double?
3. Are the AI model and service accounts in my company’s name?
4. Is there a monthly spending cap, with an alert when we get close?
5. How much does maintenance cost and what does it cover?
6. How do we measure the result, and after how long?
7. Which security checks are included?
8. Do the code, documentation and access credentials stay with me?

Start with a single process

The least expensive way to find out what automation is worth is to start small: one process, one measurable goal, one measurement before and one after. To choose which, see the guide to processes to automate with AI.

If you want to run the numbers on one of your processes, write to us: after a first call we send you a written proposal with phases, timings and costs.