Production over pilots. Everything else follows from that.

Not every engagement runs through all five, but each one sits somewhere on this path, and you always know which phase you are in.

1. 1

   Assess

   We establish where you stand: data, systems, skills, governance, and the two or three use cases that are obviously worth it. Interviews with leadership and the people who run the processes, a look at your systems, and a short written diagnosis. You get a readiness report, a first list of use cases and a clear next step.

   Typically 2–3 weeks
2. 2

   Design

   We turn the candidates into a plan: a scored use-case portfolio, a target architecture that fits your stack, a governance model and a business case with its assumptions visible. For a single use case the same happens in miniature. You get a roadmap, an architecture, a governance model and a business case your board can act on.

   Typically 4–6 weeks
3. 3

   Pilot

   We build the first agent or agentic workflow to production grade: real data, real integrations, an evaluation suite, human checkpoints and a security review. The pilot runs inside your process with a small group and is measured against the baseline from the design phase. You get a running system, evaluation results and a go/no-go decision based on numbers.

   Typically 6–8 weeks
4. 4

   Scale

   What worked is rolled out to more users, processes and teams. Shared components are hardened, internal owners are trained, and the governance framework is applied to each new use case without starting from zero. You get several use cases in production, trained owners and reusable building blocks.

   Typically 3–6 months
5. 5

   Operate

   Agents are monitored, evaluated after every change, upgraded when models change and reviewed monthly with the business owner. Either your team runs this with our playbooks, or we do it as a managed service. You get stable quality, controlled costs and documentation that stands up to an audit.

   Ongoing

• Start with the workflow, not the model

  The question is never which model is best. It is which step in which process costs you the most, and what an agent would need in order to take it over. Models are chosen last, and they are replaceable.

• Humans in the loop by design

  Every agent has defined points where a person reviews, approves or takes over, sized to the risk of the decision. Oversight is part of the workflow, not a sentence in a policy.

• Evaluate before you automate

  A test set of real cases with expected outcomes exists before an agent goes live, and every later change has to pass it. Without evaluation, automation is a guess.

• Small, senior teams

  Two or three experienced people who understand your process outperform a large team that needs managing. It also keeps your own people's time investment low.

• EU data protection by design

  Data flows, legal bases, retention and hosting location are decided during design, not discovered during review. Where EU hosting or open-weight models are the right call, we say so. See sovereign AI and EU hosting.

• Transfer knowledge, do not create dependency

  Documentation, training and internal owners are deliverables, not extras. The measure of a good engagement is that you could carry on without us and choose to keep us anyway.

Each format maps to a phase. Durations are typical and are confirmed after the intro call.

• A sponsor

  Someone in leadership who wants the outcome, can take decisions within days rather than weeks and will remove obstacles when they appear. Without this, even well-designed projects stall.

• A process owner

  The person who knows how the work is actually done, can judge whether an agent's output is right and will own the system after go-live. Typically a team lead, not a manager two levels up.

• Access to systems and people

  Read access to the relevant systems and data early on, test environments where needed, and a few hours a week from the people who run the process. Security reviews and data processing agreements are handled properly, on your terms.

• Willingness to measure

  A baseline before we start and honesty about the number afterwards, including when the result is that a use case did not pay off. That is information, and it belongs in the roadmap.