Every year, a company’s board of directors presents a report to its shareholders, and minutes are taken at every meeting. Over time, a large company accumulates hundreds of these documents, and if it also has subsidiaries, each one produces its own. Janium Collect takes the list of those documents that the archive already has and turns it into a catalog that can be searched by institution, by type of document and by date.
The example is a bank’s historical archive, which holds more than a century of annual reports and bylaws from the bank itself, from its financial group and from its subsidiaries, as well as reports from other institutions that reached the archive by other routes.
The starting point: a list
The archive had an inventory in a spreadsheet, with one row per archival file. Each row said which document it was, from what dates, how many leaves it had and where it was kept. With that list a document could be located if you knew exactly what to look for, but you could not ask, for example, which reports from a subsidiary exist between 1960 and 1980, or see at a glance everything each institution produced.
First you decide how to group
In an archive, documents are ordered by who produced them, not by subject. For that reason, before using Collect, the archive decided on the groups: one for the bank’s documents, another for the financial group’s and another for the subsidiaries’, with a section for each subsidiary. Within each group, reports are kept separate from bylaws. Reports from institutions outside the bank and its financial group went into a separate collection, so they would not be mixed with the bank’s even though they are the same type of document.
Then a column was added to the inventory with the group each row belongs to. Collect uses that column to place each document in its group.
From each row, a record
Collect reads the inventory row by row and produces one record per file in ISAD-G, the international standard for describing archival documents. What was already in the row passes into the record unchanged: the title, the dates, the number of leaves and the location. The language comes from the comments column: the record says English when the comments say so and Spanish otherwise. How the inventory’s columns are prepared is explained in Cataloging from lists and spreadsheets.
Each record also says who produced the document and gives a brief history of that institution. The producer is the institution of the group the file belongs to, and the history of each institution was written once, when the collection was set up in Collect, so all the reports from the same subsidiary carry exactly the same text.
What the AI writes
The list does not say what each report contains. To make the record useful in a search, the language model writes a description based on the type of document and its year: an annual board report usually presents the year’s results, the financial statements, the auditors’ opinion and the dividend proposal, and the description says so and places it in its period.
Today that description is marked in the record’s provenance as generated text, in a category separate from the fields that came from the row, so whoever reviews the record can tell which part the model wrote. The mark matters here because the description states what a report of that type usually contains, not what that particular report says. How that mark works is explained in Collect marks AI-made metadata.
What you gain: browsing the archive by groups
With each record in its group, Collect also builds the records of the groups themselves. The record for the subsidiaries group, for example, says how many files it contains and which years the oldest and the most recent are from, information that no row of the list gives because it describes the set as a whole. This way you can enter the catalog through the bank, go down to a subsidiary, see its reports ordered by year and open the one you need. How these values for the whole set are calculated is explained in From loose documents to Fonds and Series: generating the ISAD-G hierarchy automatically.
In the same archive, the minutes are organized in the same way, by institution and with one group per body: the board, each committee and the shareholders’ meeting. Knowing which body met is enough to reach its minute books.
Where to review
The first thing to review is the group column. If a row is in the wrong group, the record ends up with the producer and the history of another institution, and the group’s years are calculated incorrectly. Next comes reading the description the AI wrote, especially for the years when something happened that a typical report does not cover, such as a merger or a nationalization. How Collect flags the records that are worth looking at first is explained in Why trust a catalog made with AI.
Limits
The archive decides the groups. Collect places each document in the group the list gives. Without that column, it produces the records for the documents and leaves those of the groups unbuilt; if what is cataloged is the documents themselves and they are already in folders organized by group, each document’s folder serves that purpose.
The description starts from the list. The model writes from what the row says and from the type of document, so the description does not include what was decided at a specific meeting. To find a resolution you have to go to the minutes, and the record helps you get there. If you want to extract from each set of minutes who attended, what was resolved and the amounts involved, Collect has to read the document, as explained in The data in a contract, in fields.
The record inherits the errors of the list. A date entered wrongly in the inventory, or a report in English that nobody marked as such, passes into the record until someone corrects the row.
To continue the conversation
If your organization keeps board minutes and reports that today can only be found by opening folders or searching a list, write to us at info@janium.com. We would like to see how that list is built and which institutions or bodies produced the documents.