AI for CFOs: a business case that holds up in the numbers

Your job is to back AI where it lowers cost and to stop the spending where it does not. In Deloitte's CFO Survey of spring 2026, 68% of German CFOs focus on cost reduction to stay competitive, and 64% focus investment on integrating AI. Both goals land on the same desk.

Proving the return is the hard part. Forrester, as reported by CIO.de in 2025, found that “Weniger als ein Drittel kann den Wertbeitrag von KI klar finanziell darlegen”: fewer than a third of AI leaders can clearly show AI's value in financial terms. In Flexera's 2025 survey of IT leaders, only 19% planned to measure AI success in 2026.

• A measured cost per case for the process today, before any project starts.
• A total cost for the whole running period, split into build, operation and model usage. The project price is only part of it.
• A scenario in which only half of the expected benefit arrives.
• Stop criteria agreed before the project starts.
• Spend on existing licences that is visible per team and justified by use.

These are the questions finance teams and their advisers ask about AI in Germany today. Each one should have an answer in writing before money is approved.

• “What does AI cost us today, where does the money go, and how much of it is wasted?”
• “Are you actually measuring what your AI tools deliver, or trusting the vendor's slide?”
• “Does the quote give separate figures for build, operation and consumption?”
• “Is there an agreed cost cap for model consumption?”
• “What does the result look like if only half the expected benefit arrives?”
• “Are the saved hours actually cut, or just reallocated?”
• “Can I switch provider later?”

Most of the money does not go to the model. Among AI users in Bitkom's 2026 survey, the large cost items are infrastructure such as cloud and compute (51%), data preparation (50%), integration into existing systems (41%), licences (21%), external consulting (21%) and tokens (8%). Tokens are the units in which model providers bill each request.

Usage costs still deserve attention, because they grow with volume. Lünendonk's 2026 study of IT leaders reports that token costs and consumption billing are becoming a significant cost factor that narrows IT budgets. Telekom MMS's 2026 ROI study, reported by Silicon Saxony, says decisions should weigh benefit against full costs including data, integration, operations, security, governance and change.

A cost per handled case brings these together: build cost spread over the cases you expect, plus running cost and model usage per case, plus the time people spend reviewing each case, compared with what the same case costs today. We build that calculation with your own figures, without list prices.

• Build: connecting systems such as SAP, DATEV, the document management system or the ticket system, preparing data, building and testing.
• Operation: servers or cloud resources, monitoring, updates, tests after model changes, and support.
• Usage: model calls per case, driven by document length, the number of steps and the choice of model.
• Human review: the time people spend checking and correcting AI output is a real cost and belongs in the case.
• Licences you already pay for: per-seat tools that are partly unused, or cheaper as pay-per-use behind your own front end.

Decide the measures before the money. Once the threshold and the stop criteria are in writing, the project can be judged on its result rather than on the effort already spent.

• First, the process and its cost per case today: hours, corrections and waiting time, measured on real cases.
• Second, the threshold: which cost per case, which quality level and which volume justify going live.
• Third, the half-benefit test: does the case still hold if only half of the expected benefit arrives?
• Fourth, what happens to saved hours: cut, moved to work that is left undone today, or absorbed by growth. Decide it openly, because the works council will ask.
• Fifth, the running budget: who pays for operation and usage after go-live, and from which cost centre.
• Sixth, stop criteria: the measured result that ends the project instead of extending it.

Hard-to-calculate costs are a barrier for 54% of AI users and 63% of non-users in Bitkom's 2026 survey. Each of the risks below has a concrete control.

• Licences nobody uses. The free AI usage review shows seats and usage per team from your exports, so unused licences become visible before renewal.
• Usage bills that keep growing. All model calls run through one gateway with a budget per team, alerts and a hard cap, and the cost per case is reported regularly.
• Savings that never materialise. The same measure is taken before the project and after go-live, on the same kind of cases, and compared openly.
• Poor data. 36% of finance leaders name data quality as the biggest challenge for AI in finance (KPMG Global AI in Finance report, 2026). We check the data for the chosen process before any build is priced.
• AI deciding in the numbers. In a University of St.Gallen study from 2025, over 90% of finance professionals trust AI advice with human oversight, but only 49% trust purely autonomous AI advice. So the AI prepares postings, matches and payment proposals, and people approve them.
• High-risk uses. Credit scoring of individuals is a high-risk use under Annex III of the AI Act, with deployer duties from 2 December 2027 such as human oversight, monitoring, logs and informing the people affected. We do not automate such decisions without people.

Figures from other companies are at best a hint. Ask for the method and for results on your own process. We don't publish invented case figures. We measure on your cases and put the result in writing.

• A baseline measured before the project, on your own cases.
• Separate figures for build, operation and usage, with a usage cap.
• An acceptance test with a pass threshold, and its result in writing before go-live.
• The cost per handled case after go-live, reported in the same way as the baseline.
• A half-benefit scenario and stop criteria in the proposal.
• An exit plan: what you receive, and what it takes to run the system without the provider.

When purchasing runs the contract, it looks at more than the price. The agenda of the BME practice forum on buying AI in 2026 named “Beschaffung als Haftungsarchitektur”, covering governance, documentation and risk allocation, and “Lock-in früh erkennen und aktiv begrenzen”, with purchasing, IT, legal and the business units deciding together. These are the points to settle in writing. How they are judged legally is for your lawyers, and this is not legal advice.

• Contract type. A defined result with an acceptance test is closer to a Werkvertrag, a contract for work; engineering time is closer to a Dienstvertrag, a contract for services. Our written offer states what each step delivers, at a fixed price, and how it is accepted.
• Who owns code and tests. Ask that code, configuration, test sets, runbooks and measurements are handed over and can be used without the provider. That is how we work.
• Lock-in. Ask whether the system runs in your own cloud account or on your own servers, whether models can be switched behind one gateway, and what an exit takes in effort.
• Liability. “Who is liable when the AI agent makes a mistake?” The contract allocates liability. Technically, approvals for consequential actions and a full log make every step traceable, which is what any liability question will need.
• Model changes. Ask who tests and who pays when a model provider changes or retires a model, and ask for the acceptance tests to be re-run after every change.
• Data and security terms. A data processing agreement, the list of sub-processors, single sign-on, and the certifications the provider holds and for which scope.

Start with the free AI usage review. You choose which exports to send, such as licence lists, usage reports from ChatGPT, Copilot, Claude or Cursor, and cloud invoices from Azure or AWS. You receive one page on what AI costs today, where the money goes, what to change first and what to measure next. It needs no access to your systems.

If you have a new process in mind, bring it to a free 30-minute call with an engineer. After the call you receive a fixed price for each step, agreed in writing, with the measures and stop criteria in the offer.