AI for your team's workload: what changes and what stays with people

The same or more output with fewer people. KfW's chief economist put the pressure this way in 2026: "Alleine schon der demografische Wandel und der damit verbundene Schwund an Arbeitskräften macht es erforderlich, dass die Unternehmen sich mit den Möglichkeiten, die KI bietet, vertraut machen", demographic change and the shrinking workforce alone make it necessary for companies to get to know what AI can do. In the YouGov survey for Alibaba.com and bevh in 2026, about 40% of SME decision-makers in e-commerce and procurement named order processing and customer service as a major burden.

Most AI in companies already sits in operations. In Bitkom's 2026 survey of companies that use AI, 72% use it in customer contact, 54% in marketing and communication, 25% in controlling and accounting, 22% in production and 22% in internal knowledge management. What counts is shorter queues, faster answers and fewer errors in your own process. In each function that looks different:

• Customer service: incoming emails and tickets sorted, replies drafted from approved knowledge, the customer record looked up in Salesforce or your CRM. A person sends the reply.
• Sales and inside sales: quotes prepared from past offers, ERP prices and product data, with approval thresholds for discounts.
• Finance operations: incoming invoices read and matched against orders and goods receipts in SAP or DATEV, with exceptions passed to a person.
• Procurement: requisitions checked against rules, supplier documents compared, order confirmations matched with the order.
• Production and service: technicians find the fix in manuals, service reports and parts lists, and quality reports are pre-filled for review.
• HR: answers to staff questions from policies and agreements. CV screening and ranking of applicants is a high-risk use under the AI Act, with extra duties for the employer. We do not automate those decisions. People make them.

Management, the CFO and IT ask them, often in exactly these words:

• "Which processes are suitable for automation? Where are the time sinks?"
• "Do you know your three most expensive manual processes?"
• "Who checks critical results?"
• "Who is responsible when something goes wrong?"
• "Can AI use our data from ERP and machines without triggering something wrong?"
• "Can an assistant answer standard requests without giving customers wrong advice?"
• And the test an IT practitioner put plainly in a 2026 forum thread: "Wenn es mehr Arbeit macht als einspart, braucht man es nicht", if it makes more work than it saves, you don't need it.

Your team decides whether a system is used or worked around. In McKinsey's HR-Monitor 2026, 38% of German employees use generative AI regularly, up from 19%, but only 28% say their employer offers formal AI training, and their main concerns are "Fehler oder Ungenauigkeiten in KI-generierten Inhalten sowie Datenschutzbedenken", errors or inaccuracies in AI output and data privacy. In IW Köln's 2026 survey, 29.1% of employees who use AI at work use tools their employer did not provide. In Bitkom's 2025 survey, only 26% of companies gave staff access to generative AI.

• Errors in AI output: every answer shows its source, and the steps that matter go to a person to approve, so nobody has to trust a result they cannot check.
• Being monitored: we report by team, task and model and never score individual people, and what the system logs is agreed with the works council before go-live.
• Missing approved tools: an approved assistant on company data, behind one gateway, gives people a better option than private accounts.
• Their jobs: we do not make promises about headcount, which management decides. Involving the team in writing test cases and showing them the results does more than reassurance.
• Training: a short introduction to the tool they use, its limits and when to check a result.

• Which step comes first. Pick one with high volume and clear rules, such as sorting incoming mail or matching invoices, and measure hours and cases per week today.
• What the system may do alone and what needs approval. Reading, sorting and drafting can run on their own. Anything that changes an order, a booking or a customer answer goes to a person until the measured error rate justifies more.
• Which systems it reads and writes. ERP, CRM, ticket and document systems such as SAP, Salesforce, ServiceNow, SharePoint, d.velop or ELO, each with a separate account and only the rights the agent needs.
• Buy or build. If the AI features of your ticket, CRM or document system already cover the step, use them. A built agent pays off when the step crosses several systems or the data must stay on your own servers.
• How quality is judged. Test cases from your own work, with your experts deciding what counts as correct, agreed before anything is built.
• Who owns it after go-live. Someone in your team who reviews exceptions, and an operator in IT or at Cloudsail who keeps it running.

People accept a system when they keep the last word. In Inform's 2025 study of logistics and supply chain companies, reported by transport-online, 46% preferred a model in which AI makes suggestions and a person makes the final decision, 81% expected explanations and 61% wanted to be able to intervene.

Technically, the agent prepares the case in a review queue with the documents and the reasoning it used, the person approves, corrects or rejects it, and only then does a tool write to SAP, the CRM or the ticket system. Corrections become new test cases, so the system improves on the cases your team handles.

The fear is an agent that acts wrongly in live operations. In a 2026 administrator.de forum thread, an external IT provider described a car dealership's internal AI server: "System hat halluziniert, eigenmächtig Aufträge geändert, falsche Bestellungen losgetreten …", the system hallucinated, changed orders on its own and triggered wrong orders, followed by a great deal of work for staff to put it right and calm customers. Each risk below has a technical answer:

• Wrong actions in ERP or CRM: agents get tools that do one business action each, with the least rights needed. A person approves actions that change orders, bookings or customer data, and every call is logged.
• Wrong answers to customers: replies come only from approved knowledge and show their source, uncertain cases go to a person, and customers are told they are dealing with AI where the AI Act's Article 50 requires it, which has applied since 2 August 2026.
• The last few percent: an AI system rarely gets every case right. The acceptance test shows the error rate on your cases before go-live, and the review step is sized to it.
• Missing source data: if the information is not digital or cannot be reached, AI cannot help, as a VDMA expert put it in 2026: "Fehlen allerdings die digitalen Ausgangsdaten, dann kann auch eine KI-Lösung nicht weiterhelfen." A source inventory shows this before anything is built.
• Disruption to running work: the system first runs alongside the team on live cases, its results compared with theirs, before it takes over any step.
• Results that drift: regression tests from your own cases run after every change to the model, the prompts or the data, and errors and cost per case are monitored.

• A test on your own cases with the error rate per type of case, not a demo on prepared examples.
• The approval points in writing: which actions the system takes alone and which a person approves.
• Hours or cost per case before and after, measured on your process.
• What happens when the system is unsure, or when a connected system such as the ERP is down.
• How your team reviews exceptions, and how much review work that creates.
• What the system logs about your staff, agreed with the works council.

Bring one process and a rough count of cases per week to a free 30-minute call with an engineer. We look at which step to take first, which systems it touches and where your team approves. The call needs no access to your systems or data. We measure hours and cases per week before and after, on your own cases, and hand you the test set.

The use case pages below show what this looks like in customer service, sales, finance, procurement, HR, service and maintenance, the IT service desk, document processing and knowledge management. If you go ahead, each step gets a fixed price, agreed in writing after the first call.