The AI business case: how to calculate it

Almost every quote for AI promises time savings. Hardly any show how that number was arrived at. Yet you can do the sum yourself, in five steps, with figures you already have. Below is the model we use, including the places where a business case usually gets made to look too good.

Why don't most AI business cases add up?

Three things go wrong. First, time savings are estimated instead of measured: someone thinks it saves two hours a week. Second, only the build is counted, not the running costs. Third, time freed up is booked as a saving, while nobody leaves.

That last one matters most. Freed-up time only becomes money if you do something with it: less use of contractors, no extra hire, more revenue with the same people, or a shorter lead time that customers notice. Write down which of those four applies to you. If you can't, your business case is a feeling.

Step 1: measure how much time the process takes today

Take one process. For two weeks, measure how often it happens and how many minutes it takes, including interruptions and the searching around it. Let the people who do the work keep the tally themselves. A manager's estimate is almost always off, usually on the low side.

Also note how often something goes wrong and what fixing it costs. A wrongly booked invoice doesn't cost the two minutes of the booking, but the half hour of figuring it out. How to set up that baseline is explained under what manual work costs.

Step 2: convert that time into money

Use the fully loaded cost per hour, not gross salary. So salary plus employer costs plus workspace plus overheads. For contractors, use the rate you actually pay.

An example of the format, with made-up numbers. Say: 600 invoices a month, 6 minutes per invoice, that's 60 hours a month. At €45 fully loaded cost per hour, that's €2,700 a month. If you automate 70 per cent of the work, you're left with €1,890 a month. Fill in your own numbers; the outcome means nothing until you have.

Step 3: include the cost of errors and delays

Beyond hours, there's value in what goes wrong today. Think of early payment discounts you miss because invoices are approved too late, payment reminders you receive, duplicate payments, or quotes that go out so late the customer has already signed elsewhere.

Only include items you can back up from your own records. If you can't show an item, leave it out. A business case that just falls short without invented numbers is a business case you shouldn't pursue.

Step 4: set the full costs against it

A business case that only counts the build never adds up. Include these items.

• Build: design, integrations, testing and go-live. This is a one-off cost.
• Licences: model or usage costs, plus any software licences per user.
• Running costs: monitoring, spot checks, corrections and updating integrations when a vendor changes something.
• Internal time: hours from your own people for design, testing and getting up to speed.
• Preparation: cleaning up data and arranging access. This is consistently underestimated.

Indicative prices per type of project are under what an AI solution costs and what an AI agent costs. If it's mainly about integrations between systems, see what an integration costs.

Step 5: work out the payback period

Divide the one-off costs by the net monthly return, so the return minus the monthly costs. That's your payback period in months.

Use a simple threshold: under twelve months it's interesting, between twelve and twenty-four months there needs to be a second reason such as risk or lead time, and above twenty-four months you don't do it unless it's strategically necessary.

Also work out what happens if the automation only does half of what you hope. Does it still pay off? Then you have a robust case.

What should you watch for in the assumptions?

Four assumptions make or break the sum. The percentage that really runs automatically is lower at the start than the demo suggests; allow for a start-up period in which a person checks the work. Take the volume over a full year, not your busiest month.

The running costs continue for as long as the solution runs; at the very least, expect to spend time every year on changes in connected systems. And you need to be able to turn the time saved into something: less use of contractors, more work with the same team, or a shorter lead time.

When won't the business case add up?

With low volumes, with processes that run differently every time, with systems without an API that you have to work around, and with processes where a mistake can't be undone and you therefore always keep a person alongside. In those cases the answer isn't always no, but it is: do something else first.

There's often a cheaper solution. Cutting a step from the process, a setting in existing software, or a template. It's less exciting and it often pays off better. There's more on this in our article on when AI is not a good idea.

How do you get from an estimate to a number that's right?

The hard part of this sum isn't the arithmetic, it's gathering the figures. For that, you need to be with the people who do the work, look inside the systems and measure the real lead times.

That's what the AI Opportunity Audit does: in five days, establish where manual work is holding back your growth, and express that in one number with three opportunities underneath. The format and price are on the AI Opportunity Audit page. If a plan comes out of it, the next step is under AI implementation.