An AI business case: how to calculate it

Give each step in the process three numbers: how long it takes, with how many people, and how often per month. Multiply that by twelve months and an hourly rate, and you know what that step costs per year.

The formula:

Cost per year = (minutes ÷ 60) × number of people × times per month × 12 × hourly rate

The example we often use: a step of 40 minutes, with 2 people, 80 times per month. That's 1,280 hours per year. At an hourly rate of € 45, that one step costs € 57,600 per year.

Three tips for getting those numbers right:

• Ask the people who do the work. Not their manager. They know that 'just entering an order' takes forty minutes when the customer sends a PDF instead of an Excel file.
• Calculate per step, not per process. 'Order processing takes a lot of time' is too broad. Reading the email, retyping the order, checking stock and sending a confirmation are four steps, and AI isn't equally good at all four.
• Use an honest hourly rate. By default we calculate with € 45 per hour, including employer costs. If it's mostly specialists doing the work, calculate higher. Adjust it, but back it up.

In doubt about the numbers? Then measure for a week. A tally sheet next to the keyboard tells you more than an estimate in a meeting.

Estimate a range for each step: what's certainly achievable (cautious) and what's achievable if everything goes your way (ambitious). The expected figure is the midpoint between the two. And never calculate with more than 90 percent per step: there's always checking, exception handling and management left.

Why a range and not a single number? Because nobody knows in advance exactly how well AI will do in your process. A single number suggests a certainty that isn't there. A range is more honest and also gives you a lower limit you can build on.

Why never 100 percent? Because something is left at every step:

• Checking. In most processes, a human approves the result before it goes out. That takes time, less than the work itself, but it isn't zero.
• Exceptions. The order in a handwritten fax, the customer who asks three questions in one email. Those go to a human.
• Management. Someone keeps an eye on whether it's still running properly.

How much is left depends on the role AI gets. In the APO method there are four: human (AI stays out), AI assists (the human does it, AI delivers a first version), human approves (AI does the work, a human checks) and fully AI (with spot checks). Most steps land on 'human approves'. A sales proposal that goes from four hours to one hour is a good example: 75 percent less time, and the account manager still decides.

If the data is missing, or there are no examples of good work, we set the confidence of an opportunity at 40 percent at most, however good it sounds otherwise. That isn't a rejection, but an assignment: first the foundation, then the build.

Most failed AI projects don't run aground on the model. They run aground on the data: it isn't there, it isn't digital, or there are no examples of good output to learn from. That's why we ask three gate questions for every opportunity:

1. 01Is the data digitally available? Emails, PDFs, system data. Not a filing cabinet.
2. 02Are there examples of good work? Preferably ten or more, available right away. 'Sitting next to Henk for a week' doesn't count.
3. 03Is the process clear? Can someone explain what goes in and what has to come out?

If the answer is no anywhere, the opportunity gets an honest warning in the business case. You then see straight away what you need to sort out first. Sometimes that's a matter of collecting examples for a month. Sometimes it turns out the data is in three places and nobody knows which one is right. Then the project starts there, and that's a gain too.

Calculate the saving in three scenarios: cautious, expected and ambitious. The investment stays the same in all three; only the return differs. Base your decision on the expected scenario, never on the best case.

The cautious scenario shows the lower limit: the minimum outcome. The ambitious scenario shows what's possible if everything goes your way. We show it, but we never sell it. Anyone who decides on the best case is disappointed after a year, even if the project went well.

Keeping the investment the same in all three is deliberate. The price is fixed; the return is the uncertain factor. That way you know exactly what you pay and you see how sensitive the outcome is to setbacks.

A worked example

An example, not a client case. A company has three opportunities, with labor costs of € 45 per hour:

• Order entry. 40 minutes, 2 people, 80 times per month: € 57,600 per year. AI takes over 50 to 80 percent, expected 65 percent: € 37,440.
• Preparing customer questions. 8 minutes, 1 employee, 2,000 questions per month: € 144,000 per year. AI takes over 40 to 70 percent, expected 55 percent: € 79,200.
• Weekly report. 4 hours, 1 person, 4 times per month: € 8,640 per year. With a good prompt and a fixed setup 60 to 90 percent, expected 75 percent: € 6,480. This is an opportunity the team picks up itself, without a build.

Together that delivers € 123,120 per year in the expected scenario, once everything is running. Cautious it's € 91,584, ambitious € 154,656. The table below this article puts it side by side.

Add up all the costs of the first year and divide them by the saving per month. That's your payback period. At NinA we only start when the expected saving in the first year is at least twice the investment.

Payback period in months = investment ÷ (saving per year ÷ 12)

Include everything in the investment: the analysis, the build, hosting and maintenance, licenses and the time of your own people. In the example:

• AI Partnership package S, including the AI Quick Scan: € 60,000 in the first year
• Bronze maintenance contract for what's live: € 245 per month, € 2,940 per year
• Together: € 62,940, excluding VAT

The expected saving is € 123,120 per year, so € 10,260 per month. The payback period is then 62,940 ÷ 10,260 = just over 6 months. In the cautious scenario just over 8 months, in the ambitious one almost 5.

The guarantee test

Our guarantee: at least twice the investment back in the first year, otherwise we don't start. With package S (€ 60,000 in year 1), the expected saving must therefore be at least € 120,000. With package M (€ 96,000) that's € 192,000, with package L (€ 150,000) € 300,000.

At € 123,120, the example comes in just above the threshold. Here we would start, but not blindly. The customer questions carry most of it. If the Quick Scan shows there are no examples of good answers, the 40 percent brake applies. Without that opportunity, € 43,920 is left. Then we don't advise a partnership, but a workshop first, so the team picks up the first gains itself and collects the examples.

From practice: in an anonymous project, € 77,000 was saved against € 28,000 in build costs. Paid back in about 4.5 months.

One caveat: processes go live one by one. In the first year, the saving therefore comes in gradually. In the roadmap we plan per quarter what's running, so you see when the money actually comes back.

The main mistakes: calculating with the best case, forgetting the checking time, only looking at license costs and no longer measuring after go-live. And skipping the difficult question: what happens to the time that's freed up?

• Calculating with the demo. In a demo everything goes well. In practice there are exceptions. Calculate with the expected scenario.
• Forgetting the checking time. If a human reviews everything, the saving isn't 100 percent. Include that time.
• Only counting licenses. A license of a few tens of euros a month looks cheap. But without training nobody uses it, and without an integration with your systems it stays copy and paste.
• No baseline. If you don't know how much time something takes now, you can't show afterwards what it delivered.
• No longer measuring after go-live. Measure every month: how many times did it run, what percentage went well, how much time does checking still take.
• Time saved isn't automatically money saved. If the time evaporates into more meetings, you won't see it back. Most organizations we talk to use the time to grow without hiring extra people. Make that choice consciously, and write it down.

For a first impression you can calculate it yourself, with the calculator below, with the formulas from this article or with our free spreadsheet in Excel. If you want a business case you can base a decision on, with the real numbers from your processes, the AI Quick Scan is the fastest route.

In the AI Quick Scan we spend a day with you: conversations with management and the team, interviews on the work floor and processes observed live. Within a week you have:

• a baseline of your AI maturity on seven dimensions
• a business case per opportunity, with the three numbers per step, three scenarios and the confidence
• an impact-effort matrix: what first, what later
• a roadmap for twelve months
• honest package advice, or the advice not to start (yet)

The Quick Scan costs € 6,000, excluding VAT. You can buy it on its own, and the outcome is yours. If you then choose the AI Partnership, it counts towards the first-year investment.

Want to know first what AI costs in general? Read what AI costs for your business.