From our work
Retyping invoices into a spreadsheet, working through hundreds of reviews, sorting the inbox. The agent does it on its own, the result lands where you already work, and when it is not sure, it asks a person.
prototype 490 € · then from 1 490 € price breakdown →
I want an AI agent →If you are looking for something to answer visitors on your website, what you need is an AI chatbot, and that one costs from 299 €. This page is about something else: the work someone in your company does by hand today, and how to take it off their plate.
Where you are
We say it straight, because at this price it is the question that comes up most. A custom agent only makes sense once an off-the-shelf tool hits its ceiling. Here are three situations. Find yours before you write to us.
When you can manage it by hand
from 20 € / mo
When it is always the same
from 290 €
When every case is different
from 1 490 €
If you found yourself in the first column, we will tell you so on the consultation and we will not build you anything. You can come back when you hit the ceiling.
Three sample briefs from the third situation, the one where every case is different. These are not our results, they are the situations we hear about most often. On the left, what a company has and cannot keep up with. On the right, what comes out of it once the agent is running.
Document processing
What you have
Forty invoices a month sitting in attachments, every supplier in a different format. Someone opens them one by one and retypes them into a spreadsheet.
What you get
An overview where supplier, amount, due date and category are already filled in, the file is stored, and when an amount looks unusual the system asks a person before writing it down.
Data analysis
What you have
Six hundred reviews and survey answers over a year. Nobody has read them all the way through, so decisions get made on the few that stuck in someone's memory.
What you get
A list of the topics that keep coming back, with how often each one appears and verbatim quotes for every one of them. You also see what got worse year over year.
Sorting and recording
What you have
Inquiries, orders and invoices all land in the same inbox. Someone goes through them in the morning and retypes the important details into the system.
What you get
Messages are sorted and the details from them written where they belong. A person still writes the replies, the agent only prepares the groundwork.
Others we handle the same way
The difference from an off-the-shelf tool is not the model. It is what the model can see and what happens to the output next.
What they have in common is that this is repetitive work with text or documents where no rule can be written in advance, because every input looks a little different. Where a rule can be written, the cheaper automation is enough. What a language model actually is, we explain in the glossary.
What it looks like when it is running
This is the question that comes up most once we agree on the brief: where will I actually see it? The answer is almost nowhere. An agent is not another tool you have to open. It runs in the background and the result lands where you already work.
01 Trigger
Something arrives
An email with an attachment, a file in a folder, a record in a system, or just a certain time of day.
02 Work
Reads and sorts
It pulls out the details and compares them with what you already have. It takes seconds.
03 Output
It lands where it belongs
A row in a spreadsheet, a record in a system, or a message in a channel you already follow.
04 The uncertainty branch
When it is not sure, it asks
A detail does not add up or an amount looks unusual? The agent writes nothing and sends it where you already read your messages: by email, or into a Slack or Teams channel if you use them. The message says what it found, what does not add up and what it suggests. You confirm it or correct it and the agent carries on.
The rest of the batch keeps running in the meantime. Nothing waits for your reply, only that one item stops.
For the first few weeks the agent runs alongside your current process, so the outputs can be compared. Only once they match does it go live. Even after that you can see what it processed and where it asked for confirmation.
This is not a list of limitations. It is the reason it can be put into live operation at all. A tool that is allowed to do anything cannot be checked, and what cannot be checked cannot be let near your accounting or your customers. That clear boundaries and human oversight belong to a trustworthy AI deployment is also set out in the NIST AI Risk Management Framework.
If you need something that decides on its own and without oversight, we will tell you that we do not build that. Not because it could not be done, but because then there is nobody left to answer for the result.
What it costs
At this kind of money, asking whether it will work at all is fair. So we do not start by building, we start with a prototype, and the prototype has its own price. If it turns out that it does not solve the problem, we stop there and you have spent 490 € instead of the full amount.
Phase 01 · 490 €
We build the smallest working version and run it on a sample of your data. What comes out is a number for how accurate it is and an answer on whether it can be relied on.
Scope: one task, a sample of up to a hundred documents or records, results within ten business days.
Decision point
Here we look at the result together. If the prototype is not convincing, there is no point continuing and we will say so.
Phase 02 · from 1 490 €
Connecting to the system you use, checks on the output, handover to a person when the agent is unsure, and training for the team.
−490 €
The prototype is deducted from the price of the solution
What sets the price: the volume and format of the data, how many systems have to be connected, and how strictly the output has to be verifiable. With sensitive data, the way it is processed comes into it as well.
What running it costs: the agent pays the model for the volume it processes. At the volumes a small business usually has, that comes to a few euros a month. The account is yours and you can see into it, we take no margin on it.
If a configured automation or an off-the-shelf tool is enough for you, we will tell you. And if it turns out that what is missing is somewhere for that data to live, that belongs to a custom business system. The difference between the three situations is explained higher up on this page.
You get the exact figure once we have gone through your data and your goal. An overview of all our services is on the pricing page.
An AI agent from us means a program that does the repetitive work with documents and data for you: it reads incoming files and messages, pulls the details out of them, sorts them and writes them into the system you already use. When it is unsure, it asks a person instead of guessing. We start with a prototype on your real data, which verifies whether it really does solve the problem. The prototype costs 490 € and is deducted from the price of the solution, which starts from 1 490 €. Prices are final, we are not VAT registered.
If you can handle it yourself, do it. We say the same thing to clients and it is one of the reasons that scale is up there.
The difference starts to matter in three cases: when data must not leave your company, when the output has to travel somewhere automatically, and when you need to know where an answer came from. An off-the-shelf tool does none of those.
The model is often the very same one in both cases. What you pay for is everything around it.
When a model has nothing to go on, it can fill in an answer and sound convincing while doing it. That is called a hallucination. Which is exactly why the solution is not built around letting the model work from memory. It gets your source material and builds the output from it, with a source for every value.
Numbers go through a check that verifies the format and the range. When a value looks unusual, the system asks a person instead of confirming it.
We use the same principle with chatbots: when the answer is not in the source material, the bot says so and hands the conversation to a person. Where exactly that line should sit is something we went through in the article AI chatbot vs. a live operator.
It depends on what you are comfortable with, and we go through it right at the start, not once we are deploying.
In ordinary cases the request goes through the model provider's API, which does not use it for further training. When data must not leave your infrastructure at all, we build on a model that runs on your side, even though that means lower performance.
So the first question is not which model is best, but what is allowed to happen to your data. What companies have to do when they deploy AI is also being set out step by step by European AI regulation, which is why we settle it right at the start.
From our work
Related services
Is someone at your company doing this by hand?
