Agents that take over a step of the work
An agent does not only answer. It reads the document, checks it against your master data, creates the record in your system — and hands the case to a person the moment it is not confident enough. That threshold is the difference between a demo and something allowed to run in daily operations.
When is an agent worth it?
Suitability is decided by the process, not by the tool. Four traits separate the cases that hold up from the ones that stall in a pilot. Miss one and the case is not unsuitable — it needs more safeguarding, and the arithmetic changes. Miss two and it is the wrong first use case, however interesting it sounds.
Volume with a stable sequence
Twenty cases a month do not justify an integration; two thousand do. The sequence matters more than the count: a process that changes shape every week cannot be handed over, no matter how often it runs.
A result you can call right or wrong
If nobody can say afterwards whether the outcome was correct, there is nothing to accept and nothing to measure. Judgement calls stay with people — the agent prepares them.
Limited damage from a mistake
Either the error is cheap, or there is a checkpoint that catches it before it takes effect. Without one of the two, automation moves risk rather than work.
An interface to the target system
Where only a user interface exists, driving it costs more than the manual work it replaces. We say this before the project, not during it.
What we actually build
Three parts that every productive automation here is made of. They sound unspectacular and they are exactly the difference between a demo that convinces a meeting and something allowed to touch production. An agent without a handover threshold and without a log is not a faster system — it is a risk nobody notices until it gets expensive.
Handover threshold
Below a defined confidence level nothing goes through; the case lands with a person. The share of handovers is measured and part of the acceptance criteria — not a flaw to hide.
- Measured, not assumed
- Part of sign-off
- Tunable per process
Narrow permissions
The agent gets exactly the rights of the step it takes over. No shared account, no administrator access “to get started”, no standing credentials that outlive the task.
- Scoped to one step
- No shared accounts
- Revocable per action
Complete log
Every case with input, decision, reason and timestamp. Without it there is no debugging and nothing to show an auditor — which is the same requirement arriving from two directions.
- Input and reasoning
- Timestamped
- Exportable for audit
How does a project run?
In three stages, and the first one answers the only question that matters at the start: does this process carry an agent at all? That is a question about the process, not about the model — and the answer decides whether the rest is worth doing.
Process mapping
1–2 weeksWe follow the step as it is actually performed, not as the process documentation describes it. The gap between the two is where most automation projects quietly fail.
Pilot on real cases
3–4 weeksThe agent runs alongside the existing process on real cases. Nothing is written to your systems yet; we measure the share it handles and the share it hands back.
Production with a threshold
4–6 weeksWrite access, narrow permissions, full logging, and a named owner for the cases that come back. Handover goes to a person who has time for it — otherwise the queue becomes the new bottleneck.
What we will not build an agent for
Four cases we decline. Saying it at the start is cheaper than discovering it halfway through, and it is the part of the conversation most vendors leave out.
FAQ
What is the difference between an agent and a chatbot?+
A chatbot answers; an agent acts. It reads a document, checks it against your master data, creates or updates a record in your system, and hands the case to a person when it is not confident enough. The handover threshold is what makes it deployable rather than impressive.
What happens when the agent is unsure?+
The case stops and goes to a person, with the input, the partial result and the reason it stopped. The share of handovers is measured during the pilot and written into the acceptance criteria — a system that never hands anything back is either trivial or hiding its mistakes.
What access does the agent need to our systems?+
Exactly the permissions of the step it takes over, and no more. No shared service account, no administrator rights for convenience, no credentials that outlive the task. If a process can only be automated with broad access, we say so before the project rather than during it.
How do you prove what the agent did?+
Every case is logged with input, decision, reason and timestamp, exportable for audit. This is the same requirement arriving from two directions: you need it to debug, and your auditor needs it to sign off.
What does it cost?+
Discovery with process mapping starts at €5,000 and ends with a measured share the agent can handle plus an effort estimate. An implementation project on existing infrastructure starts at €25,000. Inference runs on our own hardware, so the cost does not scale with the number of employees.
How long before it runs in production?+
Process mapping one to two weeks, pilot on real cases three to four, production another four to six depending on how deep the integration goes. The pilot writes nothing to your systems — it runs alongside the existing process so the two can be compared.
Where is the data processed?+
On our own infrastructure in the EEA, not passed to a US provider. Your data is not used to train models, the processing locations are disclosed, and a data processing agreement is part of the contract.
Can an agent decide cases with legal consequences?+
No, and we will not build one for that. Credit, employment and benefit decisions stay with a person: the agent assembles the file and shows what it is based on. This is both a legal position and an engineering one — such decisions need an accountable signature, not a confidence score.
Related
The cheapest test
Describe one step someone in your company performs dozens of times a day. We will tell you whether it carries an agent — including when the answer is no, which is the answer roughly half the time.
Describe the process