Reading documents
Invoices, delivery notes and orders arrive as PDFs or scans and land in your system as structured data instead of being retyped.
AI automation for business
Work someone does by hand today: reading documents, moving data, triaging requests. With models that never leave your building, where that matters.
Automation of a named task, not a project
Most AI initiatives fail not on the technology but because nobody can name which work is supposed to disappear. So we start at the other end: you show us a task a person repeats by hand today, and we work out how many hours are in it.
Only then do we build. Usually that is not a large system but a small, unremarkable automation that sits at one point in your process and saves hours there every day. If it does not pay for itself we tell you — that is what the assessment is for.
What can realistically be automated
Moving data between systems
When the same information is entered by hand in three places today, that is the cheapest automation available.
Triaging requests
Sort incoming email by topic, urgency and ownership, and put a draft reply up for approval.
Producing text and reports
Recurring summaries, minutes and status reports from data you already hold, for a person to check.
Build it yourself?
ChatGPT in a browser tab is not automation yet
Plenty of staff already help themselves: open the document, copy the content, paste it into an AI tool, copy the answer back. That is resourceful and it does save time. We would not talk anyone out of it while it is one case a week.
At two hundred documents a month the arithmetic flips. The copying is the work, not the thinking — and it happens outside your systems, with no record, with company data in somebody else's browser tab. Automation means the task runs without a person in the middle, and the person only approves.
- Connected, not copied
- Straight from the mailbox, folder and target system — no trip through the clipboard.
- Traceable
- Every case logged: what arrived, what was extracted, who approved it.
- Data protection settled
- Local models on request, so no documents go to outside providers.
How we deliver automation
- 01
Measure the task first
We look at the process and estimate the hours per month inside it. That number decides whether we start at all.
- 02
A person keeps the approval
Automation prepares; it does not decide alone. Where correctness matters, your staff stay in the loop with an approval step.
- 03
Models in-house when needed
Where your data cannot leave the company, we run models on your own hardware.
Common questions
Does our data leave the company?
Only if you want it to. We can run models locally on your own hardware so no content goes to external providers. It takes more setup, but for sensitive documents it is often the only sensible option.
Which tasks are actually suitable?
Anything frequent, rule-bound and time-consuming today: capturing documents, moving data, sorting, summarising. One-off cases full of exceptions rarely pay off.
How do you measure whether it paid off?
Before we start we record how long the task takes by hand and how often it occurs. After launch we compare that with actual effort. Without that number, any claim about benefit is guesswork.
Do we need large amounts of data?
For most applications, no. Modern models already understand language and documents; we add your domain knowledge and rules. Custom training data is only needed in special cases.
Other services
Not sure what you need?
In 30 minutes we work out together which tasks are worth automating and what a sensible first step would be.
Already know what you want?
Send us the scope and the timeline. The assessment comes from one of the founders, not from a sales desk.