Faster replies to customers, with people where it matters

Customers get a correct answer sooner, and your team spends its time on the cases that need judgement instead of writing the same reply again. Many requests repeat: where is my order, how do I change my address, what does my contract cover, which spare part fits. The knowledge to answer them exists, but it sits in manuals, help pages, old tickets and the heads of experienced colleagues.

Customer contact is where German companies use AI most. In Bitkom's 2026 survey of companies with 20 or more employees, 72% of the companies that use AI use it in customer contact, the top area in the 2025 wave as well (88%). Yet a Capgemini Research Institute study published in 2025 found only 38% of German consumers satisfied with customer service, and more than 70% of consumers prefer a human. So the aim is faster, better answers, and customers can still reach a person whenever they want one.

• Shorter waits for routine requests, because the reply is drafted and the customer record is already open.
• Consistent answers across the team, drawn from one approved knowledge base instead of individual memory.
• New colleagues productive sooner, because each draft shows its source and the similar past case.
• Requests sorted and routed on arrival, so urgent and sensitive cases reach the right person first.
• A known cost per handled request, measured before and after on your own tickets.

The assistant prepares the routine part of each request in your text channels and leaves the decision to send with your team. For simple request types you have approved after testing, it can answer directly, with a clear notice that the customer is talking to an AI.

Phone is a separate case. Voice bots run on specialised telephony platforms such as Parloa or Cognigy, and Cloudsail does not build or sell one. What we supply sits behind a voice bot: the knowledge layer it answers from, the integrations into CRM and ticket systems, and the tests that show whether it answers correctly and hands over when it should. One public example is the Württembergische Versicherung: the voice platform vendor Parloa reported in 2025 that its agent takes calls in natural language, classifies and routes them, with waiting times for callers a third shorter.

• Email and shared mailboxes in Outlook: reads the request, recognises the customer and the topic, and drafts a reply with the sources it used.
• Web chat: answers questions about products, delivery and terms from approved content, and hands over to a person with the conversation summarised.
• Tickets in ServiceNow, Jira Service Management, Zendesk or Salesforce Service Cloud: classifies, sets priority, suggests the queue and finds similar solved tickets.
• Agent assist for your service staff: while a colleague works a case, the assistant suggests the next answer, the relevant contract clause or the manual page.
• CRM and ERP look-ups: order status, contract details or open invoices read from Salesforce, SAP or your CRM, and changes such as a new address prepared for a person to approve.
• Handover: a summary of the case, what was already said and what is missing, so the customer does not have to repeat themselves.

Your team decides what reaches the customer whenever money, a contract or a complaint is involved. The assistant prepares, and a person approves or takes over.

• Complaints, cancellations, goodwill payments and refunds: always handled by a person, with the assistant only summarising.
• Changes to customer data in the CRM: prepared by the assistant, confirmed by a colleague before they are written.
• Questions outside the approved knowledge: the assistant says it cannot answer and hands over, instead of guessing.
• Customers who ask for a person get one, in every channel.
• Which request types may be answered directly is your decision, type by type, after the test results are in.

The assistant answers only from content you have approved, reads live customer data through narrow, logged access, and runs where your customer data is allowed to be.

If your tickets and knowledge already live in one help desk, check its own AI features first. Vendors such as Salesforce, ServiceNow or HubSpot sell AI for their own tools, and if those answers pass a test on your tickets, that may be enough. A custom assistant pays off when knowledge is spread over several systems, when customer data must stay on your own servers, or when you need tested answers across email, chat and tickets at once. We are glad to run the same test on a standard product.

• Knowledge base: help centre articles, manuals, product data, terms and conditions and solved tickets from SharePoint, Confluence or your file shares, indexed with owner and version so outdated pages can be retired.
• Answers with sources: every draft shows the article or ticket it is based on, so your team can check it with one click.
• Live systems: CRM, ERP and ticket system connected through their APIs with a service account that may read only what the use case needs. Nothing is written back without approval.
• Hosting: in your own cloud account in an EU region. When customer data is sensitive, such as health, insurance or contract data, it runs on your own servers with an open-weight model, meaning a model whose weights you can run yourself.
• One LLM gateway: a single controlled access point to all models, with keys, budgets and logs. It shows the cost per request and every model call.
• AI notice and handover built into each channel: the customer sees that they are talking to an AI and can reach a person.

Before any customer sees an answer, the assistant runs an acceptance test on a set of your real past tickets and emails, with the correct answers agreed by your service team. You set the pass threshold before the test, and only the request types that pass go live.

Many teams then start in shadow mode, where the assistant drafts on live requests but nothing is sent, so the numbers come from real traffic without risk. In operation we run the same tests after every change to prompts, models or knowledge, and report the share of drafts sent unchanged, handover rates and cost per handled request. We measure handling time per request before and after, on your own tickets.

• Answer accuracy: the share of drafts your team would send unchanged, with one edit, or not at all.
• Source check: every statement backed by an approved source, with no invented terms, prices or promises.
• Handover rules: complaints, cancellations and requests for a person reach a human every time.
• Permission leaks: no customer sees another customer's data, and staff see only what their role allows.
• AI notice: shown in every channel where the customer talks to the assistant directly.

From 2 August 2026, Article 50 of the AI Act requires that people are told when they are interacting with an AI system, so a customer-facing chat or email assistant needs a clear notice. Many service teams have not prepared for this: in the Trend-Studie Contact Center 24/25 by VIER, RUF Beratung and infinit.cx, published in 2024, 72% of the customer service experts surveyed had not engaged, or only partly engaged, with the AI Act.

Data protection is the barrier companies name most: in Bitkom's 2026 survey, 66% of companies that use AI name data protection requirements. We give your data protection officer the technical facts: which customer data the assistant reads, where it is processed, what is stored in logs and for how long, and which model provider, if any, sees it. When customer data is sensitive, the whole system runs on your own servers and no data leaves the company. You decide, advised by your data protection officer and lawyers. This is not legal advice.

Because the system records which colleague accepted or changed which draft, it can be capable of monitoring performance, which involves the works council under § 87 (1) no. 6 BetrVG. We design the logs with pseudonymised IDs, agreed retention and no individual performance reports unless agreed, and document the design for the works agreement.

Start with one channel and the request types that repeat most and are fully answered by documented knowledge. Then both the test and the measure are clear.

• In a free 30-minute first call we look at your channels, volumes and systems. We need no access to systems or data for it.
• A sample of past tickets or emails, anonymised if you prefer, shows which request types your approved knowledge can answer and how often each comes up.
• We agree the acceptance threshold and the first request types, then test on your past cases.
• The assistant runs in shadow mode on live requests, then goes live type by type.
• Each step has a fixed price, agreed in writing after the first call.