Does AI pay off? How to build an AI business case

Does AI pay off for our company?

It pays off for a process, not for a company as a whole. The useful question is whether one named process, such as incoming invoices, customer enquiries or quotes, costs less per handled case, runs faster or makes fewer errors with AI than it does today, once every cost of the AI system is counted. If you cannot yet name the process, the question cannot be answered yet, and that is the first finding of your business case.

Many companies are stuck at exactly this point. In Bitkom's survey published on 14 September 2026, 54% of companies that use AI and 63% of those that do not named hard-to-calculate costs as a barrier. In a DZ Bank study reported by Börsen-Zeitung on 24 September 2026, 41% of owners and managing directors said the concrete benefit was not yet clear to them. And according to Forrester's predictions for 2026, as reported by CIO.de, fewer than a third of AI leaders can clearly show AI's value in financial terms.

Budgets are watched more closely as well. Lünendonk's study of the German IT services market, reported by IT-Finanzmagazin on 6 August 2026, describes AI investments as "stärker auf ihre Wirtschaftlichkeit geprüft und enger gesteuert", more closely checked for economic viability and more tightly steered. A business case for one process stands up to that scrutiny because it rests on your own numbers.

What does the process cost today?

Measure the current process before anyone builds anything, because a saving can only be shown against a starting point. The energy association BDEW's 2026 study asks, for every AI use case, what value is plausible and how it will be measured. Answer the second half first. Take these figures from your systems and from the people who do the work, measured on real cases:

• Cases per month, counted in the ERP, the ticket system or the mailbox.
• Minutes per case, timed on a sample of ordinary and difficult cases, including searching, copying between systems and asking colleagues.
• The loaded hourly cost of the people doing the work, agreed with controlling.
• The error and rework rate: how many cases come back, and what a correction costs further down the process.
• Turnaround time and backlog, where speed matters to customers or to cash flow.
• What the process already costs in software and outside services.

Write down how each figure was measured. If the AI version is measured differently later, the comparison will not hold, and finance will notice.

What does AI cost? Build, run and use

Sort every cost into one of three groups, because they behave differently. Build costs arise once, before go-live. Run costs arise every month, whether the system handles few cases or many. Use costs grow with every handled case. Bitkom's 2026 survey shows where the money goes in practice: companies that use AI named infrastructure such as cloud and computing capacity (51%), data preparation (50%), integration (41%), licences (21%), external consulting (21%) and training (19%) as large cost items. Only 8% named tokens, the units in which model providers bill usage.

An LLM gateway is one controlled access point to all models, with keys, budgets and logs, which makes the model cost per case visible. The Bitkom figures only say how often an item was named as large. They say nothing about its size in your case. In a process where people approve the results, review minutes and escalations can outweigh the model bill, which is why they sit in the use column next to model usage. The guide on the full cost of an AI workflow works through an example.

How good must the results be, and who checks them?

Set the quality threshold before the build starts, because it decides both the benefit and the review effort. Write it as failure types with a severity: a wrong amount booked in the ERP is not the same as a clumsy sentence in a draft reply. Some failures must never pass without a person, and for the rest you agree the share of your own test cases that must be handled correctly. Poor result quality is a barrier for 26% of companies that use AI, according to Bitkom's 2026 survey, so management will ask about this line first.

Review effort follows from the threshold. Multiply the share of cases a person checks by the minutes per check and the loaded hourly cost. Plan for full review at the start, lower the share only when measured results allow it, and keep a random sample for good. A system that is often right but needs every case checked in full may save little. A system that shows its sources, so that a check is quick, may save more.

Count reviewer time as a cost even if nobody's role changes. The guide on budgeting for human review shows how to plan those hours.

What does one handled case cost, compared with today?

Bring the three parts together into one figure, the cost per handled case. It makes the AI process comparable with today's process and with other options, such as a bought product or a simple rules-based workflow.

1. 1

   Add up the use costs per case

   Model usage and review time, plus escalated and corrected cases at today's manual cost.

2. 2

   Add a share of the run costs

   Divide the monthly run costs by the cases handled in that month.

3. 3

   Add a share of the build costs

   Spread the build over the period in which you expect to run the system without a major rebuild, and say so openly. Show the result with and without this share, because the figure without it is what the running process costs.

4. 4

   Divide by cases completed

   Count only the cases the system finished or handed over. Cases it started and never finished do not count.

5. 5

   Compare with the baseline

   Set it against today's cost per case, measured the same way. Report speed, backlog and error rates alongside the figure, but keep them out of the calculation.

Payback follows from the same numbers. The monthly saving is today's cost per case minus the new cost per case without the build share, multiplied by the cases per month. Divide the build cost by that saving and you have the number of months until the build has paid for itself. If the saving is close to zero, there is no payback to discuss. Before you sign, add one more test: what happens if only half the expected benefit arrives? Run the calculation again with half the saving.

How do we measure the ROI of AI when the value is diffuse?

Value is diffuse when AI is spread across everyone, such as chat assistant licences for the whole office: each person saves a little, nobody can say how much, and no process changes visibly. Companies still want to measure it. In KPMG's survey of 480 decision-makers in Germany, published on 9 June 2026, productivity and efficiency gains (65%) and revenue or growth contribution (48%) are what counts as success. Yet in Flexera's report of 5 November 2025, only 19% of respondents planned to measure AI success in 2026.

Three steps make diffuse value measurable:

• Measure use and spend first: who uses which tool, how often, and at what cost per team. Licences nobody uses are the easiest saving there is.
• Pick the few recurring tasks where usage clusters, such as drafting replies or summarising documents, and time them with and without AI on the same cases.
• Turn the most frequent of those tasks into a process step connected to your systems, then build a business case for it like any other process.

Our free AI usage review covers the first step: one page on AI spend and quality, built from exports you choose to send, with no access to your systems.

When do we stop, and who decides?

Write the stop criteria into the business case before the first euro is spent, because once a pilot has a team and a demo, nobody wants to end it. An expert of the engineering association VDMA put it plainly to produktion.de on 1 June 2026: "Nicht jedes KI-Pilotprojekt ist zum Erfolg verurteilt", not every AI pilot is bound to succeed. Typical criteria are:

• Quality on your own test cases stays below the agreed threshold at two review points in a row.
• The cost per handled case without the build share is not below today's.
• The share of cases that need review does not come down as planned, so the saving stays on paper.
• People do not use the system for the cases it was built for.
• A simpler option, such as a fixed rule or a better form, delivers most of the result for less.

Name who decides at each review point, usually the process owner together with finance, and what each outcome funds: continue, change the design, or stop and retire the system. The guide on when to stop an AI project covers how to take that decision without sunk-cost pressure.

Is there an AI business case template we can fill in?

Yes, the one below. Copy this table into a spreadsheet and add one column for today and one for the AI version. Leave a line empty rather than guessing, and mark which numbers are measured and which are still assumptions. Every assumption becomes something the pilot has to measure.

If you want a second pair of eyes on the numbers, talk to an engineer. The first call is free, takes 30 minutes and needs no access to your systems or data. Any work after that has a fixed price for each step, agreed in writing after the first call.

Free AI usage reviewAI strategy and rolloutFor CFOs and finance