Faster customer service and less routine billing work for Stadtwerke, with your data kept in the company

For a Stadtwerk, AI changes the back office and customer service first: customer emails and letters are sorted and answered faster, meter readings and forms reach billing without retyping, and staff find the right passage in an operations manual without calling the one colleague who knows. Studies give the grid side more attention, but these are the processes where people spend their days on routine work, and they look much the same at a utility with a few hundred staff as at a large group.

The intent is there. In the BDEW Digital@EVU study 2026 with Kearney | IMP³ROVE, one third of energy companies have already implemented their own AI strategy and a further 58% plan one, but only 29% have a personnel strategy for building future skills. Horst Dringenberg of Kearney | IMP³ROVE put it plainly in BDEW's 2026 release: many suppliers now have AI strategies on paper. The BDEW and Strategy& study "KI unter Strom" (2026), based on 26 interviews, found that about half of the companies report concrete financial benefit, but most use cases remain stand-alone pilots, and firms reported shortages of cloud computing capacity.

The reasons are rarely the model. In a 2025 guest article for ZfK, Fabian Dömer of Arthur D. Little names a missing integrated data estate, missing interfaces, poor data quality and often no mandate to anchor AI strategically. He counts around 1000 Stadtwerke and sees many of them waiting. BDEW's 2026 release adds that smaller utilities find the known hurdles especially hard to overcome. We work on exactly these gaps: one process connected to billing, SAP and your documents, running in your own environment, with a measured result.

In processes where requests and documents arrive in volume and follow rules, so a language model can prepare the work and a person decides. The utilities named below were reported in the press or by vendors. Cloudsail did not build any of them. Each use case has its own page, linked below.

• Customer emails and letters. The assistant sorts incoming emails and documents by topic, such as a move, a meter reading, an advance payment or a complaint, prioritises them and drafts a reply from your tariffs, terms and the customer record, which staff check before sending. Stadtwerke Halle reported in the press in 2026 (ZfK) that about 800 of roughly 3100 employees have access to its internal generative AI platform, and that in customer service the AI classifies emails and documents, prioritises requests and drafts replies that staff review before sending.
• Meter readings and forms into billing. Scanned meter-reading cards, SEPA mandates and bank detail changes are read, checked against the customer and contract, and prepared for the billing system, with doubtful values passed to a person. Stadtwerke Pforzheim reported in the press in 2026 (ZfK) that it reads scanned meter-reading cards, checks them and transfers them to billing, with SEPA mandates and bank detail changes planned.
• Operations manuals and knowledge for field staff. Fitters, network technicians and service staff ask in their own words and get the answer from the operations manual of the right division, with the page it came from. Stadtwerke Pforzheim reported in the press in 2026 (ZfK) that it uses operations-manual agents for electricity, district heating, gas, telecoms and water that route each question to the right area.
• Purchase requests in SAP. An assistant asks staff what they need, finds the material or service and the cost centre, and prepares the purchase requisition in SAP for the usual release. Stadtwerke Pforzheim reported in the press in 2026 (ZfK) that its bot "Banfred" helps staff create purchase requests in SAP.
• Phone service. Voice agents for routine calls are a platform product. Rhenag reported in the press in 2026 (ZfK) a pilot with Parloa and Logen.ai for routine calls on meter readings and changes to advance payments, topics behind about 100000 requests a year, with non-standard cases passed to a person. We build what such a platform needs behind it: approved answers, the lookup in the customer record and the step into billing.
• Business customers and market communication. Arvato Systems reported, in an undated vendor case study about work from 2022, that for Stadtwerke Leipzig it clustered business customers in SAP CRM and built a bot for market communication messages, with more than 1000 records processed.

Part of this you buy off the shelf. A general chat and agent platform for writing and summarising is a product: Stadtwerke Halle uses Langdock, and Stadtwerke Pforzheim reported about 190 personal Copilot agents next to about ten centrally approved agents that run workflows (both reported by ZfK in 2026). Voice agents for the hotline come from platforms such as Parloa. Fichtner IT Consulting offers LUCRA, a cloud AI service for utilities and municipalities on data-centre infrastructure from SWU TeleNet. Its 2026 vendor document says each customer's data lies isolated in its own tenant. The savings it shows are projections, not measured results.

Building your own makes sense where a platform stops: when the work has to reach billing or SAP, when answers must respect who may see grid or customer data, when the system must run on your own servers, and when you need test results to show an auditor. Ralph Bogesch of Stadtwerke Pforzheim said in ZfK in 2026 that just bringing ChatGPT or Copilot into the company does not achieve the goal. Pforzheim also reported that its test environment is not enough for broader production use, that it is talking with external providers and that it is evaluating local open-source models for offline transcription. We fill that gap with infrastructure, integration and tests, next to the platform you already have.

KRITIS and NIS2 change what an AI system has to prove. The German NIS2 implementation act applies directly from 6 December 2025 with no transition period, the BSI portal for registration and incident reporting opened on 6 January 2026, and the act also amends the EnWG (Heuking, 2025). In the EY and BDEW Stadtwerkestudie 2026, based on phone interviews with the management of 100 Stadtwerke and regional suppliers, 46% expect large or very large investments in protecting critical infrastructure. For AI this means private hosting, documented data flows, controlled access, logs and test records you can show. It also means we work as a supplier inside your security requirements, with security clauses and audit rights in the contract where you need them.

The AI Act. Annex III point 2 lists as high-risk AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or the supply of water, gas, heating or electricity, with obligations applying from 2 December 2027 after the Digital Omnibus. Anything that works as a safety component in grid or plant operation is therefore high-risk, with duties such as human oversight for the company using it, and we leave it out: we do not build grid control, plant control or protection functions. Assistants for customer service, billing documents, procurement and manuals do not fall under this point.

Customers on the phone and in chat. From 2 August 2026 people must be told when they are dealing with an AI system. Melanie Florenz, head of customer service at Rhenag Energie, said in ZfK in 2026 that customers do not want to talk to robots about individual questions. We therefore design every customer-facing step with a clear hand-over to a person, and replies to individual cases stay drafts that staff send.

Unbundling and customer data. The BDEW and Strategy& study (2026) names unbundling rules for data exchange among the barriers to scaling AI. Grid operator and supplier data stay separated in the index and in every tool, following the access rules you already apply. Meter readings, consumption and bank details are personal data: they stay in your systems, and only staff who may open a customer record get answers from it.

Works council. An assistant that logs who asked what can in principle monitor performance, which can make it subject to co-determination under § 87 (1) no. 6 BetrVG, and the BDEW and Strategy& study lists works council acceptance among the barriers. We design logs with pseudonymised IDs, agreed retention and clear access rules, ready for the works agreement. This is the technical side. You decide, advised by your data protection officer and lawyers.

Daniel Krauss, head of digitalisation and IT at Stadtwerke Pforzheim, named three conditions in ZfK in 2026: data sovereignty, governance and a manageable infrastructure. We build to those three. On your own servers, an open-weight model, a model whose weights you run yourself, answers without data leaving your data centre. In your own cloud account, the models run in an EU region under your contracts. In front of every model sits an LLM gateway, one controlled access point to all models with keys, budgets and logs, so token use is visible per team and per agent. Stadtwerke Halle reported in ZfK in 2026 that one agent interaction consumed 8.6 million tokens, which is why it is discussing token controlling.

• Billing and customer data: customer, contract and meter records read live from your billing system through approved tools, and entries prepared there only through a tool that does one business action, with approval and a log entry.
• SAP: materials, services, suppliers and cost centres from SAP, and purchase requisitions created as drafts that go through your existing release workflow.
• Documents: operations manuals, work instructions, tariffs and terms from SharePoint, file shares, d.velop or ELO, indexed per division with the permissions they already have.
• Mail and CRM: Outlook mailboxes and shared customer service inboxes, and the customer history in SAP CRM or your CRM.
• Agent register: every agent that more than one person uses is listed with its owner, sources, tools and test results, and is shared only after approval. Stadtwerke Halle reported in ZfK in 2026 that it keeps 25 use cases in a register and that sharing an agent beyond its creator will need an approval covering quality, compliance, business value and resource use.

Bertil Kilian of Stadtwerke Pforzheim recalled in ZfK in 2026 that at the beginning the AI sometimes invented companies that did not exist. That is why nothing goes live untested. Before go-live the system is measured on your own past cases: customer emails with the reply your team sent at the time, scanned meter-reading cards compared field by field with the values in billing, manual questions with the right answer and the page it comes from, and purchase requests with the entries your buyers made.

You set the acceptance threshold with us, for example the share of meter readings that must be read correctly before they enter the approval queue, and the acceptance test shows whether it is met. We also test what must not happen: a sales employee receiving grid data, a reply quoting a tariff that no longer applies, an answer taken from the wrong division's manual.

Before it takes over any step it runs in shadow mode, preparing replies or entries alongside your team without sending anything. We measure hours per case before and after on your own cases, so management sees the effect in its own numbers. Test results, model versions and changes are recorded so you can show them in an audit, and the same tests run after every change to models, prompts or sources.

Start with one process in the back office or customer service that costs your staff hours every week and has clear past examples: the customer service inbox, meter-reading cards and SEPA mandates, purchase requests or questions about one division's operations manual. Describe it to an engineer in a free 30-minute first call, which needs no access to your systems or data.

Then we settle the first step and its acceptance threshold. Each step has a fixed price, agreed in writing after the first call. If your staff already use Copilot or a chat platform, the free AI usage review shows on one page what that costs and where it falls short, built from exports you choose to send.