AI for specific jobs in your company — not for slide decks

Who it’s for

A company receiving invoices as PDFs by email, where someone types them in one by one. And then into the spreadsheet.

A company with a lot of accumulated paperwork, where finding a clause, an old quote or an agreed condition takes half a morning.

A company receiving hundreds of emails that have to be read, sorted and passed on before anyone can do anything with them.

What’s included

• Reading and filing invoices
• Generating invoices
• Reading and analysing legal documents
• Sorting incoming email
• Transcription
• Searching your own documentation
• Assistants and chatbots, when there’s a specific job to do

On that last point, something that also explains our own site: we build a chatbot when there’s a defined job for it. There isn’t one here, so we haven’t put one in. An assistant that doesn’t solve anything specific is expensive decoration.

What isn’t included

This is the part worth reading in full.

• We don’t build algorithms that support clinical decisions.
• We don’t build software for activities without legal recognition.
• We don’t hand over systems without human oversight in sensitive sectors.
• We don’t train models on data whose origin isn’t transparent.
• We don’t sell an AI project when the problem is your data. If the information is scattered, duplicated or unreliable, no AI fixes that — it repeats it faster. In that case we say so and propose something else, or nothing.

That last one costs us projects. We keep it because a system built on bad data fails anyway, just three months later and after we’ve been paid.

What happens to your data

Your data, documents and conversations are never used to train public models. Not by us and not by the providers. The information stays yours.

Where each thing sits. Your data is stored on servers inside the European Union. To process it with AI, documents pass through the model provider: we use commercial plans where data doesn’t feed training and, where the provider allows it, with retention switched off and processing kept in Europe.

We tell you which provider handles what, in writing, before we start. Not afterwards.

Documents with sensitive data. These can be anonymised before processing. Anonymising reduces the risk, it doesn’t remove it entirely: for the most delicate cases we say so and decide together whether it’s worth it.

Encryption TLS 1.3 in transit and AES-256 at rest.

Accuracy and how we work

We don’t promise percentages before seeing your data. Before starting we measure accuracy on real cases of yours and give you the figure. Whatever the system can’t resolve confidently goes to a person, and that share falls every month.

If the measured figure doesn’t reach the threshold we agreed, there’s no project.

Model costs. These are recurring and we don’t hide them inside the quote: you pay them directly from your own account and see the real spend. If there’s an agreed monthly fee, they can sit inside it.

Common questions

Is this going to replace my people?

No. It replaces tasks, not people. Whoever types invoices moves to checking the doubtful ones. If your goal were a smaller payroll, we’d tell you.

Will my data train an AI?

No. We use commercial plans where that doesn’t happen, and we document which provider is involved at each step before we start.

What happens when it gets something wrong?

It does get things wrong, which is why a person always checks whatever the system flags as doubtful. A system that never fails doesn’t exist; one that says when it isn’t sure does.

What does it cost?

The project is quoted at a fixed price. Model costs sit outside it and you pay them directly, with the invoice in front of you.

Why doesn’t your own site have a chatbot?

Because there’s no specific job for it here. We build them when there’s defined work to do, not to show that we can.