How we check
Rules first, AI second, and a qualified person decides. This page explains what we check, how we know it works, and where it stops.
What goes in
A controlled export of your current documents with their version numbers and effective dates, taken from your document management system. Before any check starts, we compare the export with the master list from your document management system. Documents that are referenced but missing, and material left outside the check, are reported first.
Bring the export on an external drive with a Thunderbolt or USB-C connector, in a form a computer can read (for example PDF or Word). We review documents written in English or Estonian. Scanned PDFs are fine as long as the text is machine-readable.
Two kinds of checks
Anything that can be checked with a fixed rule is checked with a rule. AI is used only where the meaning of the text has to be compared.
| Checked by | What |
|---|---|
| Rules | Document codes and titles, versions, effective date against approval date, signatures, references to forms or annexes that do not exist |
| Local AI | Limits, deadlines, roles and equipment that differ between documents; procedures compared with the EU GMP clauses we give it as text, never from the model's memory |
| Client QA reviewer | Every finding: confirmed or rejected, with the reason recorded; the follow-up is decided in your own quality system |
What you receive
A finding report. Each line names the document, version and clause, describes the possible inconsistency, quotes the text it is based on and leaves room for the reviewer's decision. It can be filed with your self-inspection or change control records. See a sample finding → · Full sample report →
How we know it works
We test the system on a synthetic document set with a known number of planted errors, and we report three numbers: errors found, errors missed and false findings. A result applies only to that task and that test. The test set and the answer key can be downloaded, so anyone can inspect the planted errors and run their own comparison: Demo Pharma test set (ZIP, documents in Estonian). Results of test 1 →
When something changes
A new model, new settings or a new document structure can change the results. After any such change the test set is run again before the system is used on your documents.
Where it runs, and your data
For a review we bring our own machine, a Mac Studio. Its disk is encrypted, and its network connections (Wi-Fi and Bluetooth) are switched off while your documents are on it. We never need an account in your systems. The work is covered by a confidentiality and quality agreement, signed before any document is opened.
| Stage | What happens |
|---|---|
| In | You hand us the export on your own external drive. It is connected by cable; nothing travels over a network. |
| During the work | The documents are read and checked on our machine at your site. Nothing is sent to a cloud service. |
| Out | The finding report, the list of findings, a record of what was checked and a plan for putting the system right are saved to your drive. |
| What leaves with us | By default only the signed handover record and a work log that contains no text from your documents. We keep a copy of the report only if you agree in writing, and then for an agreed period. |
| What we delete | Everything else: the working copies of your documents, the search index built from them and, unless you have agreed otherwise, our copy of the report. This is done before the machine leaves your site, and both sides sign a short handover and deletion record. |
| If the machine were lost | The disk is encrypted, so nothing on it can be read without the key. We would tell you without delay. |
For continuous use the system is installed on a Mac Studio that you own and we configure. It stays at your site and works without an internet connection. Once a quarter we check that everything works, at your site or, if you prefer, by remote connection. A remote connection is opened only for maintenance and under your IT rules. The local AI is updated only with your QA's approval: the test is repeated before the new version is used, and the previous version is kept so that it can be restored.
Your documents are not sent to any cloud service. Running locally tells you where the data is processed; it does not by itself prove security or GMP suitability. Access control, logging and fitness for the intended use are shown separately.
Where it stops
- AI does not release batches, approve documents or replace the Qualified Person.
- A finding is a suggestion until a qualified person has reviewed it.
- Before use we assess with your QA team how the findings will be used and what could happen if the system misses an issue. The consultation draft of EU GMP Annex 22 (2025) says that generative AI should not be used in critical GMP applications.
Using it well
Before your team relies on the results, it should be able to answer four questions:
- Which task is the tool suited for, and which is it not?
- How do we check an output before acting on it?
- When do we not use the result?
- How do we record the check and the decision?
We help your team answer them as part of the introduction, and in our training courses.