# AI Use Cases by Department: Where Agents Fit in Mid-Sized Companies > Agents fit where people repeatedly read text, gather facts from several systems and write drafts. That usually means sales, quoting, service, purchasing, finance and knowledge work. For each department, check volume, data access and who signs off before you pick a process. Canonical URL: https://www.ruebstudio.de/en/guides/finding-ai-use-cases/ai-use-cases-by-department Author: Dr.-Ing. Marcus Rüb, Founder Published: 2026-10-12 Lists of AI use cases tend to run to twenty items and leave you guessing which ones suit your own business. This guide sorts the options by department and shows how to tell, department by department, whether an agent makes sense. It describes typical processes, not success stories. Status: October 2026. ## Which departments use AI today? A September 2026 survey by the German industry association Bitkom found that 57 percent of companies with 20 or more employees use AI. Customer contact is the most common area. The ranking tells you where AI is widespread, not where it would pay off in your company. According to the [Bitkom press release of 14 September 2026](https://www.bitkom.org/Presse/Presseinformation/Erstmals-nutzt-Mehrheit-Unternehmen-KI) (603 German companies with 20 or more employees, surveyed by phone), companies named these areas: - Customer contact: 72 percent - Marketing and communication: 54 percent - Controlling and accounting: 25 percent - Internal knowledge management: 22 percent - Production processes: 22 percent - Research and development: 16 percent - Sales: 14 percent - IT department: 9 percent - Legal or tax department: 8 percent - Management: 7 percent The release does not state the base for these shares, so treat them as orders of magnitude. The year-on-year jumps are striking: knowledge management rose from 11 to 22 percent, sales from 5 to 14 percent, and IT from 2 to 9 percent. [Eurostat](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises) gives a wider European picture. In 2025, 19.95 percent of EU enterprises with ten or more employees used at least one AI technology. Of those, 34.70 percent named marketing or sales as a purpose, 31.05 percent the organisation of administrative processes and 6.08 percent logistics. The two surveys differ in definition, company size and method, so do not compare them directly. The practical reading is modest: text-heavy areas such as customer contact, marketing and administration were opened up first. Whether that is the right starting point for you depends on your processes, not on a league table. ## How do you recognise agent potential in a department? A process suits an agent when it is frequent, based on text or documents, needs facts from more than one source, and a person can check the result before it takes effect. If one of these is missing, ordinary automation, or no new tooling at all, is usually the better answer. These signals help with a first assessment. The more of them apply, the more a closer look is worth. **Repeated reading and summarising.** People open emails, PDFs, specifications, delivery notes or inspection reports every day to copy three to five facts out of them. Language models handle this well because the inputs are unstructured and no fixed interface exists. **Searching across systems.** An answer needs the ERP system, the document archive, the ticket system and a mailbox. An agent can query these in turn and bring the findings together, provided it has access. **Draft character.** The output is a quote draft, a reply, an order proposal or a report that someone reviews anyway. An occasional mistake costs little because the review catches it. **High volume with loose input.** Ten cases a day with fixed rules are often better served by a form or a rule in the ERP. An agent earns its place when inputs vary but the underlying rules are known. **Knowledge in individual heads.** If only one colleague knows how to handle a special case, an agent can make that knowledge usable once it is written down. The writing down is often the larger part of the work. Counter-check: if a process is rare, differs every time, or involves a legal or safety decision that nobody can verify, leave it out for now. For a definition of the term, see the glossary entry [What is an AI agent?](https://www.ruebstudio.de/en/glossary/ai-agent). ## Which use cases exist in each department? Every department has typical processes in which agents do the preparation. The sections below describe them without customer stories and without result figures. The selection is not exhaustive; it covers the departments mentioned most often in mid-sized companies, meaning firms of roughly 50 to 500 employees. ### Sales and quoting In sales an agent prepares work: it reads incoming enquiries, extracts the requirements, finds similar past quotes and drafts a response. Price, discount and dispatch stay with a person. This matters especially for contract manufacturers, where enquiries often arrive as PDFs with drawings. The full description is in [AI quoting agent: handling RFQs, specifications and costing](https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-quoting-agent-rfq). For sales in the narrower sense, meaning lead research, call preparation and follow-up, see the overview page [AI agents for sales](https://www.ruebstudio.de/en/ai-agents/sales). Typical signal: salespeople spend a large share of the day researching and writing instead of talking to buyers. ### Customer service and technical service In service an agent sorts incoming tickets, proposes replies and searches manuals, fault reports and earlier cases for a matching solution. Customer contact is the most frequently named area in the Bitkom survey. For machine builders, technical service adds a further task: engineers gather spare parts, serial numbers and maintenance history. The agent usually prepares a reply rather than sending it. If it talks to end customers directly, you need clear limits: which topics it may answer, when it hands over and how conversations are logged. More on the page [AI agents for technical service](https://www.ruebstudio.de/en/ai-agents/technical-service). ### Purchasing and procurement In purchasing an agent reads supplier quotes, compares line items with framework agreements, checks lead times and prepares order proposals. Requests to several suppliers can be drafted and the replies evaluated. The signal: many supplier quotes in different formats, compared by hand in a spreadsheet. Limit: negotiation and award decisions stay human. The agent provides the basis for comparison, not a commitment. ### Accounting, controlling and finance In accounting and controlling an agent reads incoming invoices, matches them to purchase orders, proposes account assignments and flags deviations. In controlling it can comment on monthly reports and explain anomalies. In the Bitkom survey 25 percent of companies named controlling and accounting, up from 17 percent a year earlier. Traceability requirements are strictest here. Bookings must stay auditable, and any figure the agent reports has to come from a source you can show. The agent should not do the arithmetic itself; it should call a tool that does. ### HR In HR, agents suit pre-sorting applications against fixed criteria, drafting job adverts, answering recurring staff questions about leave, expenses or policies, and assembling onboarding material. Be most careful with anything that assesses people. Decisions on hiring, promotion or dismissal do not belong with an agent. Involve the works council, where one exists, and the data protection officer before personal data is processed. ### Knowledge management and IT A knowledge agent answers questions from manuals, project documentation, contracts and wikis and cites where it found the answer. In the Bitkom survey, internal knowledge management doubled from 11 to 22 percent, and the IT department rose from 2 to 9 percent. Typical signal: new staff need months to learn where things are documented. In IT an agent can take first-line helpdesk requests, sort reports or help with documentation. A starting point is the page [AI agents for knowledge management](https://www.ruebstudio.de/en/ai-agents/knowledge-management). It requires a document store the agent may read and a permission model it respects. ### Production, quality and maintenance In production, agents rarely stand at the machine. They stand beside it: they read inspection reports, evaluate deviation notices, suggest causes from earlier cases or prepare shift handovers. According to Bitkom, 22 percent of companies name production processes as an area of use. Controlling plant equipment itself has different requirements for real-time behaviour and safety. A language model does not belong in the control loop. It is useful for analysis and documentation, where a person reads the result. ### Marketing and communication Marketing is the second most common area in the Bitkom survey at 54 percent, and in the Eurostat data 34.70 percent of AI-using enterprises name marketing or sales as a purpose. Typical processes are drafts of product texts, newsletters, data sheet summaries and the preparation of technical articles. The risk is quality: texts that nobody checks technically will contain errors. Whoever publishes statements about their own products is responsible for them, not the model. ### Management Only 7 percent of companies in the Bitkom survey name management as an area of use. An assistant can prepare briefing material for meetings, pull key figures from several systems and draft decision papers. It does not replace the decision. As a first pilot this area is a poor choice, because access to sensitive data would be broad and the process rarely repeats. ## What does a department need before an agent can start? An agent needs three things: access to the relevant documents and systems, a person who knows the process and can judge results, and a clear point where a human signs off. If one is missing, a pilot rarely fails on technology. It fails on organisation. **Data and access.** Find out where the documents live: a file server, the ERP, the ticket system, individual mailboxes. Clarify whether an interface exists and who may release it. Poor document quality, such as skewed scans of drawings, lowers results but does not rule out a pilot. **A specialist with time.** Someone from the department must supply 10 to 20 real cases, explain the exceptions and rate the results. Without that person you judge the agent by gut feeling. **An approval process.** Decide who checks what the agent produces and what happens when it is wrong. The same applies to permissions: an agent should only read and write what the specialist may read and write in that process. **Legal frame.** Personal data requires a legal basis and clarity about where models run. Operation in an environment you control is possible; see the page on [sovereign AI](https://www.ruebstudio.de/en/sovereign-ai). This section is orientation, not legal advice. ## How do you proceed step by step? Proceed in five steps: collect processes, score them against the signals, choose one pilot, test with real documents, and only then roll out. The effort lies in choosing, not building. Starting with five candidates and picking one saves you from pilots that go nowhere. 1. **Collect processes.** In each department ask: which task repeats weekly, starts with reading and ends with a draft? Note frequency, systems and people involved. A workshop of two to three hours per department is often enough. 2. **Score against the signals.** Give each candidate points for volume, readability of the documents, data access, checkability of the result and damage if it goes wrong. You set the weighting yourself; what matters is that it is the same for all candidates. 3. **Choose one pilot.** Take a process with high volume, a cheap failure case and willing users. A pilot in a department that resists change teaches you little about the technology. 4. **Test with real documents.** Let the agent handle 10 to 20 completed cases and compare its output with what actually happened. Count errors by type, not only by number. 5. **Define operations.** Decide who looks after the agent, how errors are reported and when you adjust it. An agent without an owner goes stale. Methods for finding candidates are covered in the methods pillar article once it is published. How sectors differ is the subject of a separate article on industries. ## What are the limits and risks? Agents make mistakes, invent details and can be thrown off by manipulated input. They suit preparation work with human review, not unchecked decisions. The larger the potential damage, the narrower the permissions and the stricter the sign-off should be. **False confidence.** A draft reads convincingly even when it is wrong. If reviewers come to trust the agent blindly, errors slip into the process. Spot checks and an error log belong in operations. **Too many permissions.** An agent with write access to the ERP and a mailbox can do real damage, for example a wrong order or an email to the wrong recipient. Begin with read access and drafts. **Missing data foundation.** Where knowledge is documented nowhere, an agent cannot use it. Expect preparing the documents to take more time than the agent itself. **Overestimated effect.** Wide adoption says nothing about the benefit in your own company. Measure in the pilot what actually changes, such as handling time per case before and after. **Legal and data protection questions.** Personal data, trade secrets and customer data must not flow uncontrolled to external services. Clarify which models run where before a pilot starts. These points are orientation, not legal advice. ## How do you get started? Begin with a list of five to ten tasks from two or three departments, score them against the signals above and choose one process for a pilot. Plan for one meeting per department and one test run on real cases, not for weeks of preparation. These questions help in the first conversation with a department: - Which task would you most like to hand off, and why? - Which documents do you read for it, and where are they? - How do you notice that a result is wrong? - Who approves the result, and what happens after a mistake? - How often does the process occur, and how different are the cases? If the answers point to a clear, frequent process with a checkable result, a pilot is worth it. If they are evasive, the process is usually not yet clear, and documentation will help more than technology. For an overview of the functions that have dedicated agents, see the page [AI agents for mid-sized companies](https://www.ruebstudio.de/en/ai-agents). ## Frequently asked questions ### Which department should start with AI agents? Start where a process is frequent, built on readable documents and cheap to correct. Quote requests, service tickets and invoice checks often qualify. Management decisions and legally sensitive calls rarely do. ### Do you need clean data to begin? No, but you need documents that are findable and an agent that may access them. A pilot can work with imperfect master data as long as you know where the exceptions are. ### Will an agent replace people in the department? In mid-sized companies an agent mostly does preparation: reading, sorting and drafting. Decisions, approvals and customer contact stay with people, who remain accountable for the outcome. ### How long does it take to assess one use case? A first assessment needs one conversation with the people doing the work and a look at about ten real cases. Only a small test on real documents shows whether it is feasible. ## Sources - [Bitkom: Erstmals nutzt die Mehrheit der Unternehmen KI (press release, 14 September 2026, in German)](https://www.bitkom.org/Presse/Presseinformation/Erstmals-nutzt-Mehrheit-Unternehmen-KI) - [Eurostat: Use of artificial intelligence in enterprises (2025 data)](https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Use_of_artificial_intelligence_in_enterprises)