Fifteen answers, none of them evasive.
Including the ones that cost us work: what we turn down, what we are not certified against, and what an agent still cannot do.
Working together.
How long before a first agent is in production?
Two weeks of scoping, then four to six weeks of build. The first agent covers a deliberately narrow scope so it stays measurable and quick to correct. Teams that try to ship three agents at once usually ship none.
Do we need a data or AI team already?
No. You need someone who owns the process and can answer questions about how it really works, and someone who can grant access to the systems involved. We bring the rest, and we train your team during handover so the agent does not depend on us.
Do you work with smaller companies?
Yes, if the process has enough volume to be worth measuring — roughly from 200 cases a month. Below that the effort rarely pays back, and we will say so rather than sell you a build.
What do you turn down?
Processes two teams still disagree about, briefs to "add AI" with no target process, work that changes fundamentally every quarter, and data nobody has permission to use yet. Each one produces a system that cannot be tested.
Can you take over an agent somebody else built?
Sometimes. We start with an audit against the same criteria we apply to our own work: is there a test set, are permissions scoped, is there an approval path, can a run be reconstructed. If none of those exist, rebuilding is usually cheaper than repairing.
How it is built.
Which models do you use?
Whichever passes your test set at an acceptable cost and latency, usually a frontier model for reasoning and a smaller one for classification. Models are a dependency, not the product: the agent is built so one can be swapped and re-scored without a rewrite.
How do you measure that an agent works?
With a business test set built from your own history during scoping — the ordinary cases, the edge cases and the ones that went wrong. Every change to prompt, model or retrieval is scored against it before it ships. We report what the set does not cover rather than quoting one accuracy number.
What happens when the agent gets it wrong?
It should already have escalated. Anything crossing a business threshold is drafted and queued for a person instead of executed. When something does go wrong in production, the run id links the input, the retrieved sources, every tool call and the approver, so the cause is found and the writes are reverted rather than reconstructed by hand.
Does the agent train on our data?
No. Your data is used for retrieval at query time, not to fine-tune a model. Nothing you give us is used to improve a model for anyone else, and that is in the contract rather than only on this page.
What does it cost to run once it is live?
Model inference and hosting, paid directly to your own provider so the cost stays visible and portable. For the agents we run today that is between 40 and 600 EUR a month depending on volume. We estimate it during scoping.
Data and governance.
Where is our data hosted?
In the EU. Model providers, vector storage and logs all run in European regions, and subprocessors are listed by name in the DPA. If a component cannot meet that, we say so before you sign rather than after.
Who can see what the agent did?
You can, from your own side. Traces are queryable and carry the input, the retrieved passages, the tool calls, the gate decisions and the identity that approved the output. Retention is set by you and enforced by deletion, not by hiding rows from a view.
Are you certified against ISO 27001 or SOC 2?
No, and we say so plainly. We complete security questionnaires, sign DPAs and take part in vendor review, but we will not imply a certification we do not hold. What we do have is published on the security page.
How do you handle the EU AI Act?
We classify the system at scoping and document it. Most business agents we build land outside the high-risk categories, but the classification is written down with its reasoning, so your compliance team can disagree with it on the evidence.
Who owns what you build?
You do. Prompts, test sets, retrieval configuration and infrastructure code are assigned to you outright. There is no runtime licence and nothing stops working if you stop working with us.
Still unanswered?
Write to us with the question. If it is a good one it usually ends up on this page.