AI for managing directors: from mandate to a process that pays off

You are asked to show that the company uses AI where it pays off, without putting data, money or the workforce at risk. That is harder than it sounds: in the DZ Bank study of more than 1000 owners and managing directors, reported by Börsen-Zeitung in 2026, 41% say the concrete benefit of AI is not yet clear, and only 16% plan notable AI spending.

The pressure comes from competitors and the labour market rather than from a board. In Bitkom's 2026 survey of companies with 20 or more employees, 61% agree that companies without AI have no future, and 33% of AI users say they use it mainly out of fear of losing touch. KfW's chief economist said in 2026 that demographic change and the shrinking workforce alone make it necessary for companies to get to know what AI can do.

Many companies have already bought licences and seen little change. Bitkom's president put it plainly in 2026: “Wer nur eine Lizenz kauft, hat KI eingekauft, aber noch nichts verändert”, whoever only buys a licence has bought AI but changed nothing yet. In the same survey only 3% of AI users say they exploit AI's potential rather strongly.

In practice, then, the mandate means naming the processes where hours, errors or waiting times cost money, showing a measured result on one of them, and building it so that it runs every day with an owner and a budget.

• Less routine work in a named process: fewer hours per quote, order, supplier invoice or service request, measured before and after on your own cases.
• Relief where people are missing: work that is not getting done today because the team is short of staff.
• Data that stays in the company: offers, contracts, customer and personnel files processed in your own cloud account or on your own servers.
• Costs you can see: a cost per handled case and a budget per team instead of surprise bills.
• A workforce that comes along: the people doing the work and the works council are involved early, before go-live.

Once AI is on the agenda, the same questions come from every side. Have short answers ready, and expect most of them to come out of the first measured process rather than out of a workshop.

• From yourself and the shareholders: “Where do we start, and what does it concretely bring us?”
• From finance: “Does the quote give separate figures for build, operation and consumption?” and “What does the result look like if only half the expected benefit arrives?”
• From IT: “Where is our data processed, and who runs the system after go-live?”
• From the data protection officer: “Which data may go into the model at all?”
• From the works council: “What does the system log, and can performance or behaviour be evaluated from it?”
• From staff: “Will AI replace jobs here?” and “Why may I not use the tool I use at home?”

Decide in this order, because each answer narrows the next. A strategy is useful once it rests on one measured result: KPMG found in 2026 that 98% of companies have an initial AI strategy, yet only 39% say top management actively steers it.

• First, the goal in business terms: which measure counts (hours per case, cases per week, waiting time, error rate) and which result would justify the next step.
• Second, the first process: frequent, rule-heavy work on documents or requests where today's effort can be counted, such as incoming orders, quotes, supplier invoices or service tickets. KfW calls finding the first practical application “eine zentrale Hürde”, a central hurdle (2026). Scoring a few candidates on volume, effort and risk settles it quickly.
• Third, where the data may go: your own cloud account in an EU region, your own servers, or a provider's service. Decide it with IT and the data protection officer before anything is built.
• Fourth, build, buy or both. Where a standard product such as Microsoft 365 Copilot or the AI built into your ticket or invoice software already covers the job, buy it. Build where the process runs across your own systems, such as SAP, DATEV and your document management, or where data must stay in the company.
• Fifth, who owns it after go-live: a named person in the business, a budget for running costs, and who operates it, your IT, Cloudsail or both.
• Sixth, stop criteria: the measured result that ends the project instead of extending it.

Data protection and legal uncertainty are the obstacles named most often: 43% of owners and managing directors in the DZ Bank study (Börsen-Zeitung, 2026) call them the biggest one. Most of these risks are handled in how the system is built. A policy document only adds to that.

• Data leaving the company. Models run in your own cloud account in an EU region or on your own servers. An LLM gateway, one controlled access point to all models with keys, budgets and logs, decides which model may see which data, and the logs stay in your systems.
• Shadow AI and personal liability. When staff paste customer data into private tools, the company stays responsible, and a labour lawyer points to managing directors' duty of care under § 43 GmbHG and § 93 AktG in such cases (Pöppel Rechtsanwälte, 2026). In practice, a ban only works when there is an approved tool that people want to use.
• Wrong actions in live operations. An IT service provider described an internal AI system at a car dealership that had “halluziniert, eigenmächtig Aufträge geändert, falsche Bestellungen losgetreten …”, hallucinated, changed orders on its own and triggered wrong orders (administrator.de forum, 2026). In what we build, the agent prepares the case and a person approves anything that changes an order, a payment or a customer record, and every step is logged.
• Costs nobody planned. 54% of AI users in Bitkom's 2026 survey name hard-to-calculate costs as a barrier. We work with a cost per handled case, budgets per team in the gateway and separate figures for build, operation and model usage.
• AI Act duties. 89% of companies that are affected or still checking expect high implementation effort (Bitkom, 2026). Most process automation is not high-risk, but some uses are: screening and ranking job applicants and credit scoring of individuals are high-risk uses under the AI Act. Under Annex III, from 2 December 2027, you as the deployer then owe human oversight, use according to the instructions, monitoring and logs, and information for the people affected. We do not automate such decisions without people.
• Losing the workforce. Lack of employee acceptance is a barrier for 42% of AI users (Bitkom, 2026). Brief the works council before go-live, keep reporting at team level, and never score individual people.

A consultant described in 2026 what owners want: “Ein Prototyp, den man anfassen kann, sagt mehr als 90 Folien”, a prototype you can touch says more than 90 slides (Business Punk, 2026). Ask every provider, including us, for the following.

• A working prototype on your own documents or cases, not a demo on sample data.
• A baseline: how many hours and cases the process takes today, measured before the project starts.
• An acceptance test: agreed test cases from your own work, a pass threshold, and the result in writing before go-live.
• A description of the data flow: where data is processed and stored, for how long, and who can access it.
• Separate figures for build, operation and model usage, and a cap on usage costs.
• An exit path: you receive the code, configuration, tests and runbooks, and the system can work with another model or provider. The same consultant warned: “Wer da nur ein Abo verkauft, verkauft eine Abhängigkeit, kein Werkzeug”, whoever only sells a subscription sells a dependency, not a tool.

Start with a free 30-minute call with one of our engineers. Bring one process and a rough idea of its volume. We need no access to your systems or data, and we bring no slides. We do not show invented case figures. We show the method and measure it on your own cases.

If you already pay for ChatGPT, Copilot or similar licences, the free AI usage review is a useful second step: one page on what the licences cost and how they are used, built from exports you choose to send. After the call you receive a fixed price for each step, agreed in writing.