Agentic AI · from assessment to operations
One prompt. Your whole company.
We build agentic systems that go into production, from frontier models, your own data and an orchestration layer that stands up to an audit. The first use case is taking shape right now in energy and district heating.
Why most AI pilots never reach production
It is almost never the model. It comes down to three things nobody has to solve in a prototype, and without which no production system is ever signed off.
The prototype knows nothing about permissions
In a pilot everyone sees everything. In production the answer has to depend on who is asking, otherwise the assistant becomes a data leak with a search box.
Nobody can explain where the answer came from
With no source and no log there is no sign-off: not from audit, not from data protection, not from the regulator.
Every use case starts from scratch
If the second costs as much as the first, scaling never adds up. The initiative then dies in the budget process, not in the technology.
What we actually build
The business case rests on four layers. Three of them are technology you can buy. The fourth decides whether the system is ever allowed to go live, and that one does not come off the shelf.
Audit logevery step traceable
Permission is checked first
searches only approved sources
Documents, ERP, databases, file stores
passes context
Action in the target system only after approval
The path of a question: permission is decided before retrieval, not after, and every step is written to the log. Those two properties are exactly what separates a production-ready design from a prototype.
Models
We integrate frontier models rather than training our own. Models age in months; the architecture around them does not. Swap the model and everything else stays in place.
Data retrieval on your data
Answers come from your documents and systems, retrieved at runtime rather than trained into the model. Permissions come from the source system and are not reinvented.
Orchestration
Several specialised agents share one task: research, check, summarise, act in the target system. An orchestrator decides who goes when, and when a human takes over.
Guardrails and operations
Permission model, logging, escalation on critical actions, monitoring, operational handover. The unglamorous part, and the one that decides whether you get sign-off.
Three phases, three exit points
01 Assessment: 2 weeks
We review your candidates, check the state of your data and quantify the benefit. Result: a prioritised list and a clear statement of which case will carry and which will not. You can stop there.
02 One case to production readiness: 6 to 12 weeks
One use case, real data, real permissions, a real audit log. No demo dataset. Result: a system that can be signed off, or a documented reason why not.
03 Operations and scaling
Handover to your team, or operations run by us. Every further use case builds on the same foundation, which is why it costs a fraction of the first.
The questions your auditors will ask
Who sees what?
Permissions are taken from the source system, not reassigned inside the AI system. What someone may not open in the original never appears in an answer either.
Where is the data held?
You decide. Whether it runs in your own environment, in Switzerland or in the EU is settled before we build, and follows your requirements.
Is our data used for training?
No. The model runs in your environment or in ours, not at an external provider. Your content does not leave that environment and does not flow into the weights of the model.
Can you trace how an answer was produced?
Every answer names its sources and every agent step is logged. Traceability is not an add-on; it is part of the design.
What happens before an agent changes anything?
Write and trigger actions go through an approval step. You define which operations a human has to confirm.
What if the model gets it wrong?
Without a verifiable source the system gives no answer; it says it did not find one.
Change the industry, the method stays
Regulation, data landscape and vocabulary differ from industry to industry. The architecture of data retrieval, orchestration and guardrails does not. Energy and district heating is the first vertical in which we are putting the method into practice, highlighted for exactly that reason.
The uncomfortable questions
What does a first use case cost?
It depends on the case. A number on a website would be a guess, so there isn't one here. We work at a fixed price or on a time and materials basis: you decide where the risk sits. The assessment phase is priced separately so you can walk away after two weeks with no commitment. We give you a firm number as soon as we have seen your case.
Do we need our own AI people for this?
Not to build it. To run it, yes, which is why we design the handover in from the start and can place specialists in your team on request.
Are we locking ourselves into one model provider?
No. Providers are moving at different speeds, so we deliberately stay ready to switch. The business case is built so that the model remains replaceable. Orchestration, data retrieval and guardrails do not depend on any one provider.
Our data is unstructured and partly out of date. Does this still work?
That is the norm, not the exception. The assessment phase settles exactly that, and it will also tell you when a case fails on the state of the data.
What if the models change fundamentally within a year?
Then you swap the model. That is precisely why the investment sits in the architecture and not in the model.
Bring two processes, we will tell you which one will carry
45 minutes, two of our people, no slide deck. You leave with an assessment of which of your use cases is production-ready and where the others would fail.
We reply within one working day.