Typical duration: 3–8 weeks
Budget band: EUR 3,000+

AI agents

An agent takes over repetitive work and records every step it takes. Without the second the first is unusable: unseen actions cannot be trusted, however right.

What is built

The boundary is written first

What the agent may not do is written down as carefully as what it may. Which operations are irreversible, which spend money, which send something outward are each marked separately. That list is the scope itself, not a security footnote appended to it.

What it connects to, and with what permission

An agent connects to email, a calendar, a database or a panel. For each connection the narrowest workable permission is chosen: if reading is enough, writing is not granted. The worst thing an agent can do stays bounded by what it was handed.

Every step is recorded

Which input produced which decision, which tool was called and what came back is kept in readable form. When a result is questioned, it can be traced. Without a record, an agent's mistake is only discovered through a wrong outcome, and by then it is late.

The points needing human approval are marked

Irreversible operations wait for confirmation. Where it stops is your decision and it is written at the start: the difference between preparing a draft and waiting, and sending directly, is one of the most consequential lines in the scope.

How the process works

01The work gets written down
However a person does that job today, it is written out step by step. It is common at this stage to find that part of the work simplifies without needing automation at all, and that is a good outcome: work that will not be done gets no agent built for it.
02A prototype on one narrow slice
Not the whole flow, but the single step that repeats most. The point is an early answer: can the model do this, and on which inputs does it stumble. If it cannot, that is learned in the first week rather than at the end of the month.
03Accuracy is measured
A set built from real examples is prepared and the agent's output is measured against it. You set the acceptable threshold; an agent put into service without that measurement is an agent whose error rate nobody knows.
04Deployment and monitoring
After it starts running, where and how often it stops keeps being watched. In the first weeks the number of points requiring human approval is deliberately kept high, and narrowed as confidence is earned.

What sets the price

Every system the agent touches adds a line to the price, and the state of the data inside those systems decides the rest. On most jobs, putting messy data in order takes longer than building the agent itself.

how many steps the flow has and how much it branches
how many systems it connects to, and whether they offer an API
the accuracy threshold being aimed at
how many points require human approval
monthly volume, which is what sets the model usage cost

Common questions

Which model is used, and what happens when models change?

The model is chosen for the job, and the setup is built so it is not locked to a single provider. That is a durability decision rather than a preference: models change every six months or so, and when they do, what needs rewriting is the call layer rather than the agent's logic.

What happens if it gets something wrong?

Irreversible operations already wait for approval, so the worst case is a wrong draft. On reversible ones, the record shows which step broke. Because accuracy is measured before deployment, the error rate is a known number rather than a guess.

Does my data go into model training?

On enterprise API access, providers do not use submitted data for training and commit to that contractually. Which provider offers which terms is shared in writing before setup. If the data must not leave your infrastructure at all, models running on your own hardware are an option, with cost and accuracy discussed separately.

Who pays the model usage fees?

The provider account is opened in your name and usage bills there directly rather than passing through here. An estimate based on expected monthly volume is given during setup; because volume is measurable, that estimate becomes a real figure after the first month.