Looking for a 2019 letter that nobody can find in any of the boxes? Digital archiving starts with a method: name, code and give every item a single shelf. The tool comes afterwards, and it never decides on its own which shelf an item goes on.

In many institutions, paper piles up in boxes while scanned files pile up in a shared folder where everyone names things their own way. Scanning alone is not enough: a badly named scan gets lost just as easily as a sheet of paper. Here is the method we apply with e-Archives, our digital archiving platform, and what the machine does – and does not do – at each step.

1. Start with the name and the guide

It all starts with a naming rule. Every scanned item receives a standard name: a category code, the year, the month, the registration number, then a short label. The pattern: CODE-YYYY-MM-NNN_label.

Three rules prevent most mistakes:

• The label states the subject, not the sender. ‘Maintenance contract renewal’ tells you something; ‘Letter from the ministry’ tells you nothing.
• No dates or numbers in the label: they are already in the prefix. Digits that are part of a name or a rank (‘2nd session’) are fine.
• Number unknown? Write 000, never 001. 001 is a real number, since the register starts again at 1 every January.

These rules fit into an archiving guide: the list of codes, what they mean and the naming rule. At Lya & Co, we wrote this guide for the institution. When a difficult case is settled, it is recorded in two places: in the guide for staff, and in the instructions the machine follows.

2. Scan your paper archives and let the platform do the filing

The institution’s staff scan the documents and upload them by drag and drop in the browser. From the code at the start of the name, the platform sets the document type and category; the year and month in the name then place the item in a category / year / month folder. During our tests in September 2026, a batch of 15 uploaded items was filed in 69 seconds, and every item was searchable straight away.

Scans go through character recognition, in French and English, and are then converted to PDF/A, the archival format.

A typing mistake does not lose the item. A name whose code is not recognised is flagged ‘Name to correct’ (‘Nom à corriger’ on screen; interface labels are translated from French throughout this article), with a note explaining why; a reference that has already been used is flagged ‘Number already used’.

Every evening, an automatic pass goes over the day’s uploads. It reads the exact day from the document itself (the name only gives the year and month), along with the addressee and the country concerned, then suggests a shelf for items that have none. It never replaces a field that is already filled in. Only an overly vague label (empty, generic, very short) is rewritten, and the reference code stays intact.

Here is where the line is drawn. The foundation – type, category, folder, reference, character recognition, full-text search – runs on fixed rules and works without artificial intelligence. AI sits on top of it: exact day, addressee, country concerned, rewriting vague labels, suggesting shelves and correspondents, and the assistant.

During testing the AI once proposed a date in 2076; since then it only writes specific fields, subject to checks (same year as the name, no future dates). The reference code never depends on the AI: the program realigns it with the recorded reference, whatever the model suggests. On a real collection, the September 2026 campaign rewrote generic, overly short or truncated labels, and a full before-and-after comparison of every record confirmed that no field was lost.

3. The archivist confirms the shelf

Each shelf in the platform matches a real shelf in the archive room. A sheet of paper sits in one box only, so it carries one shelf, never two.

The most common trap is confusing the sender with the subject. A letter sent by the ministry does not belong on the ‘ministry’ shelf for that reason alone.

Why not let the machine decide? Because AI on its own picks the right shelf only about half the time (measured in September 2026). So the platform also looks at the shelves of the closest documents that are already filed, with no suggestion pending. When the two agree, the suggestion is marked ‘neighbours agree’. In the measurement of 8 September 2026, this level covered about a third of suggestions; it was right in 86 to 90% of cases. The others are marked ‘to be confirmed’.

Either way, for every new upload the archivist decides: no suggestion is ever confirmed automatically. The task stays short: check or correct the suggested shelf, clear the alert, and never leave two shelves. A shelf the archivist has removed is never suggested again.

One last principle: nothing is deleted, everything is versioned. A file strictly identical to one already archived is refused on upload. When an existing item is scanned again, the new scan is attached to its record as a new version, and the version attached last is the one displayed.

For near-duplicates, similar text is never enough: the same note sent to two addressees makes two originals. Someone has to read both, then keep the most complete scan as the reference – an extra page, the arrival stamp, the signature.

4. Finding things: search, filter, ask

Full-text search covers the title, the recognised text of every page, the correspondent, the type and the shelf. It handles the quirks of French: ‘lettre d’invitation’ returns the same results as ‘lettre invitation’, and a word in the singular also finds the plural. Results are ranked by relevance: look at the first lines, not the total displayed.

So what about that 2019 letter? For one specific item, search is enough. For every 2019 item on a given shelf, that is an inventory: use the filters, not the search bar.

Then there is the ‘Ask the archive’ assistant (‘Interroger l’archive’ on screen), for questions in everyday language (‘what do we know about…’). It cites the documents it relies on, each with a link to its record, and that record shows the shelf where the original is kept. It answers within seconds (measured in September 2026) and says so when it cannot find anything.

Access is through personal accounts, with three permission profiles: document management, editing and read-only. The interface is entirely in French, in the institution’s own colours.

Are your archives sitting in boxes or in a disorganised shared folder? Let us start with your collection and your naming habits: get in touch for an initial conversation, or find out more about e-Archives on our Products page.