Questions we hear in almost every first call

A founder-led, senior-only consultancy in Berlin for AI strategy, agentic workflows and AI agents, founded in 2025. We are vendor-neutral: we hold no reseller agreements and recommend models and platforms based on your data, your systems and your risk appetite. Our focus is production over pilots, governance built in, humans in the loop and a business case you can measure. Everyone who works on your engagement has built and run AI systems themselves. More on the about page.

Remote-first, on site when it helps. Discovery workshops, leadership sessions and the start of a pilot are often better in the same room, and we travel to you in Berlin, across Germany and, where it makes sense, in Austria and Switzerland. Day-to-day build work, reviews and check-ins run remotely in your tools. We agree the rhythm with you at the start so that nobody spends time on trains for a status update.

English and German, natively in both. Workshops, documentation, roadmaps and the agents themselves can be delivered in either language or both, which matters when a German Mittelstand company has an English-speaking engineering team or an international group needs German-language agents for its local operations. Multilingual agents for further languages are part of the build, not a special request.

Startups from seed stage, Mittelstand companies (owner-managed German SMEs, typically between fifty and a few thousand employees) and business units of corporates. The work differs: startups need product and sequencing decisions, the Mittelstand needs a plan a lean IT team can execute, and corporates usually need governance and a reference architecture to get pilots into production. See startups, Mittelstand and enterprise.

With a 30-minute intro call. We listen, ask about your processes and tell you honestly where agents would pay off and where they would not. If it makes sense, the usual next step is a readiness assessment of two to three weeks or a scoped pilot of six to eight weeks. After the call you receive a written proposal with scope, deliverables, timeline and price. No slide decks, no obligation.

Fixed-scope packages for defined formats such as the readiness assessment, the strategy and roadmap or an agent pilot, and monthly retainers for managed AI operations, fractional Chief AI Officer roles and longer scale-out programmes. Every proposal is quoted after the intro call, once we understand the scope, and states what is included and what would be extra. We do not bill for the intro call, and we do not publish rate cards, because scope varies too much for a single number to be meaningful.

Software that uses a language model to work toward a goal in several steps: it reads inputs, decides what to do next, uses tools such as your CRM, ERP, email or a database, checks the result and continues until the task is done or a person needs to decide. A chatbot answers a question; an agent processes the order, drafts the reply, updates the system and flags the exception. Our article What is agentic AI? goes into detail.

A chatbot converses: it answers questions in a dialogue and stops there. RPA (robotic process automation) replays fixed click sequences on structured screens and breaks when a form changes. An agent understands unstructured input, decides between steps, uses systems through their interfaces and can hand over to a person when unsure. In practice they combine: an agent may use RPA-style connectors as tools and offer a chat interface. See agents vs. RPA vs. chatbots.

Processes with high volume, unstructured input such as emails, documents and tickets, clear rules for the normal case and a measurable outcome: order intake, claims and invoice processing, customer service, onboarding, contract review, internal knowledge questions, reporting drafts. Poorly suited: rare, high-stakes decisions with little precedent, processes nobody can describe, and anything for which the data does not exist yet. Our guide to choosing the first use case includes a checklist.

Typically six to eight weeks from kickoff to a production-grade pilot running on your real data and systems with the team that owns the process. The first two weeks go into scoping, access and the baseline, the rest into building, testing with real cases and iterating. Longer timelines are usually caused by access to systems and data, not by the AI, which is why we settle access in week one.

Buy where a vendor's agent already covers your process well and the data may go there; build where the process is specific to you, the data is sensitive or the agent becomes part of your product. Many companies do both: a platform such as Microsoft Copilot Studio, n8n or a CRM vendor's agents for standard cases, custom agents for the workflows that set them apart. We are vendor-neutral and help you decide per use case. See build vs. buy.

Whatever fits the use case, the data and your existing landscape: models from OpenAI, Anthropic, Google, Mistral and open-weight models on EU infrastructure; platforms such as Microsoft Copilot Studio, Salesforce Agentforce, n8n, LangGraph or custom code; and the Model Context Protocol (MCP) to connect agents to systems. We have no partnerships or reseller agreements and no preference beyond what works for you and stays operable by your team.

This will happen, so we design for it. The model sits behind an abstraction layer, prompts and evaluations are versioned, and every agent has an evaluation suite so that a replacement model can be tested against real cases within days. Where data protection or contracts require it, we keep an EU-hosted or open-weight alternative ready. Your agents belong to you, not to a model.

As a design constraint from day one, not as a review at the end. Each use case starts with a look at which personal data is involved, whether it is needed and where it may be processed. We minimise the data sent to models, prefer EU processing, support your data-protection officer with data-flow documentation and, where required, provide the input for a data protection impact assessment. This is not legal advice; it is engineering that makes your legal review straightforward.

Yes. Options include EU-region endpoints of the large providers, European providers, private cloud deployments and open-weight models on your own hardware. The trade-off is between capability, cost and control, and it differs per use case: a knowledge assistant for internal documents may run entirely on your own infrastructure while a customer-facing agent uses an EU-hosted frontier model. Our article on sovereign AI and EU hosting explains the options.

Yes, in a manageable way. The AI literacy duty applies to any organisation using AI in its operations; transparency duties apply when people interact with AI or read AI-generated content; and a small number of use cases, for example in HR, credit or critical infrastructure, count as high-risk with extensive obligations. Most business agents are outside the high-risk list. We classify each use case, document it and recommend confirming with counsel, since guidance and timelines are still evolving. Start with our EU AI Act guide.

In Germany, usually yes, whenever a system could be used to monitor performance or behaviour, which describes most agents that log who did what. Involving the works council early, with a clear description of what the agent does, what it logs and what it is not used for, is far quicker than a late objection. We bring templates for that description and experience with the questions works councils ask.

Only what the task requires, and only through permissions you control. Agents access systems with their own scoped credentials, retrieval is limited to the documents the requesting person may see, and sensitive fields are masked or excluded where they are not needed. Everything the agent reads and writes is logged. We define this scope with you in writing before the build starts.

You do. Everything we build for you, including code, prompts, evaluation suites, documentation and configurations, belongs to your company. We may reuse general know-how and patterns, never your data or your specifics. You are free to run, change or extend the system with your own team or another partner. Our proposals say this explicitly.

With a baseline before we build: hours per case, cycle time, error rate, response time, backlog, revenue effect, whatever the process actually turns on. The pilot measures the same numbers on real cases, and the business case uses your volumes and your cost rates, with the assumptions written down. We do not promise a figure in advance; we agree the metric, measure it and let the numbers decide whether to scale.

Then you find out in weeks, not years, and with a written explanation. Pilots are scoped so that failure is cheap: a bounded process, a fixed timeframe, real cases, a clear success criterion. If the agent does not meet it, the report says why, whether a smaller scope would work and what would need to change first. Sometimes the honest answer is that the process is not ready, and we say so.

In most companies the first agents absorb backlog and growth rather than replace people, because the routine work they take over was crowding out the work that needs judgement. Change still needs managing: roles shift toward reviewing, deciding and handling exceptions. We involve the people who run the process from the first workshop, and we offer training so that teams learn to work with agents instead of around them.

Language models can produce confident nonsense, so we never rely on the model alone. Agents are grounded in your documents and systems, cite their sources, work within guardrails and hand anything below a confidence threshold to a person. Evaluation suites measure the error rate before go-live and monitoring watches it afterwards. The goal is not zero errors, which no process achieves; it is an error rate you know, that is lower than today's and that is caught before it reaches a customer. See guardrails for agents.

Agents need care like any production system: models change, documents change, new edge cases appear. After go-live you get monitoring, an escalation path and a handover to your team. If you would rather not operate it yourself, managed AI operations covers monitoring, evaluation, updates and improvements on a monthly retainer. Either way, the system is documented so that the choice remains yours.