Documents, emails and data someone still copies by hand. We read them, check them and write them into the systems you already use.
Three things we have already built and tested on real data. All of them plug into the systems you already use: management software, CRM, email, Microsoft 365.
Bank statements, foreign and non-SDI invoices, delivery notes, contracts, certified emails with attachments. A system reads them, checks every field and writes the data into your management software. Every value carries its source; doubtful cases go to a queue where a person approves them in ten seconds. We measure accuracy on your documents before you sign.
It reads emails, documents and web pages, sorts what comes in, drafts the entry or the reply. Every action waits for a confirmation and stays in the log. Nothing runs on its own. It is the missing piece between ChatGPT and your management software.
Log · 14:02 approved by M.R. · 13:58 sorted · 13:57 received
A chat that answers across documents, emails and CRM together, cites its source and says "I don't know" when it doesn't. It inherits permissions: nobody reads what they couldn't open. It works where Copilot doesn't reach: sources outside Microsoft, data that cannot leave.
We test on your real documents, with a test corpus and reference cases. You see the number before the contract, not after.
Page, line, confidence level. A reviewer can check, you can trust.
TeamSystem, CBI and the others, verified byte by byte. Not a CSV to fix by hand.
Review queue, a log of every action. No automatic write without a trace.
Scope and acceptance threshold written down first. If the system doesn't reach the threshold, we fix it.
Also on request: AI Assessment and Claude setup for people who don't write code.
Four systems built in 2026 and tested on real data. The status of each is written below, as it is.
An engine that reads bank statements and invoices, checks every transaction and writes the TeamSystem import file. The check is deterministic: if the numbers don't add up, the file doesn't go out.
3,875 tests · 255 real test documents · import file validated byte by byte
A chat across Microsoft 365, Slack and the CRM of a group with more than twenty sites. It cites the source, says "I don't know", and shows nobody what they couldn't open.
353 security tests · 3 connected sources · inherited permissions
Agents that read the pages of 255 e-commerce sites, classify the changes and flag only the ones that matter. Tested on real data for a logistics operator.
255 sites · 400 pages re-checked · AI cost under 1% of operating costs
The system we use ourselves: agents with quality gates, a firewall on actions and a log of every run. Running since July 2026.
since July 2026 · every run logged · no action without a gate
For every project we start by understanding what the person using the software does today, and what they could do tomorrow with an extra hour. Technology comes second. The first days with a client are about listening, not proposing.
We build software that goes to production and stays there for years. No proof-of-concepts for their own sake, no slide-deck architectures. Every project has a named Tech Lead from day one to the last.
The people who work with us aren't a number on a contract. We choose them with care, train them, and grow together, project after project: that's how you build a relationship that lasts.
A fixed-price project with an acceptance test, or a monthly plan with a Tech Lead and releases every two weeks. In both cases the contract is a works contract: we commit to the result, not to the time.