When your AI assistant can't find your documents, the problem is almost never the AI.
It's that your documents aren't readable to it. A scanned PDF is an image: to an AI system it contains not a single word. I measure how much of your archive is in that situation and tell you with numbers, not impressions.
- 0internet calls during the analysis, enforced by an automated test
- 1,000pages per sample, in Spanish
- 24-48 hoursfrom receiving the sample to the report
- 160automated tests passing on the tool
What gets measured
Five things, on a sample of your real documents. Each answers a question you currently answer by gut feeling.
Documents that can't be read
A corrupt PDF, a .docx that won't open. To an AI assistant, a
document that won't open simply doesn't exist.
Extraction coverage
The share of pages with text that can actually be extracted. A scanned PDF with no OCR is invisible to a retrieval system: it is the most common cause of "the AI can't find what I ask for".
Personal data exposure
Counts by type, separating identifiers with a verifiable format (national ID numbers, IBAN, card numbers, social security numbers) from mentions that only the language model proposes.
Fragmentation
How many chunks the text splits into, how many are too short and how many are duplicates. It predicts poor retrieval before you build anything.
Token variation after anonymisation
How much text the anonymisation moved. It goes in an annex, and the report states explicitly that it is not a cost saving: it can come out negative.
One real corpus, one number
In one test corpus of 631 real scanned pages (academic notes), 319 pages (50.6%) had no extractable text. It is a single corpus, not a benchmark and not an industry average: it shows the problem is real, not how large it is in yours. Yours is what the scan measures.
And after the scan, what?
The scan is the first of three steps. Each is proposed only when the previous one has shown, with numbers, that it is needed.
The scan
€950up to 1,000 pages · 24-48 hoursMeasures the real state of your archive. It is the diagnosis, and the only thing you need to buy to know whether the rest is worth it.
The setup
from €6,500from €5,550 after deducting the scanFixes what the scan found and ends with a second report that measures the before and the after.
The monitoring
€350/monthno lock-inCorpora rot. A monthly pass over what is new, and an alert when a metric crosses a threshold.
If the scan comes out clean, I tell you and that's the end of it. A healthy archive doesn't need a setup, and selling you one would be the opposite of everything this page says.
No document leaves your machine
It is why this analysis can run on files you would never upload to a cloud.
The whole analysis runs locally: there are no internet calls and no API keys at any point in the process. An automated test backs it: it intercepts any attempt to open a connection during a full analysis and fails if one happens.
It can be run in front of whoever looks after your systems, before a single file is touched.
What this analysis does not do
Putting it up front is part of the method. A report that doesn't state its limits isn't a report, it's advertising.
It doesn't estimate how often your AI is wrong
It measures the state of the documents, not the behaviour of the model.
It doesn't do OCR and it doesn't fix anything
It detects the pages without extractable text; it doesn't convert them. The report diagnoses: cleaning and structuring the archive is separate work.
Does your case fit?
Write and I'll tell you, with no commitment, whether it does. If it doesn't, I'll say so: it costs me a client and earns the only credibility I have.