Until now, the finding aid of the Central Registry of the Church Council of the Evangelical Church of the Palatinate consisted of two series of index cards – a keyword card catalog and a card catalog organized according to the registry plan. All metadata existed in analogue form, and further work in the Central Registry remained largely analogue.
To enable contemporary workflows and research, and to reuse the existing metadata for archival purposes, a digital shift in the Central Registry’s processes was both indispensable and long overdue.
As a fundamental first step in overhauling those workflows, a digital working foundation had to be established. Traditionally, this would have begun with manually transferring the metadata from the index cards into the database. In the case of the Central Registry’s card index, however, the Central Archive in Speyer decided instead to launch a small-scale AI-supported pilot project.
The reasons for this decision are manifold; here are just a few aspects:
Initially, the archive scanned the index cards in-house according to the service provider’s specifications. A small, easy-to-use document flatbed scanner with an automatic feeder was sufficient to complete this task quickly. More time was required to locate “challenging special cases” among the cards, which deviated significantly from the standard template and served, among other things, as test files.
The file, comprising a total of 3,096 scans of 1,548 DIN A5 index cards, was intended to serve as the basis for the AI-assisted conversion of machine-printed index cards with handwritten annotations into an Excel file.
Archipanion preprocessed the scans by pairing front and back images, performed field extraction using a large language model (LLM), and standardized the formats. After a few test runs, this succeeded in many cases where the cards deviated little from the basic structure.
For cards whose structure deviated significantly, the AI understandably reached its limits. Semantic interpretation of handwritten annotations and the assignment of notes remained manual tasks. To capture as much information from the analogue cards as possible, a very thorough quality control process was chosen, in which each scan was compared to the metadata generated in the spreadsheet.
To make this as user-friendly as possible, Archipanion provided a web front end for direct comparison of scans and generated entries, into which corrections could be entered directly. This enabled the resource-intensive quality control to be performed at a very high level, in a modern way, and within a manageable timeframe. Missing or misassigned information could be quickly and easily added and simultaneously checked for archival plausibility. The quality control was completed by the registry clerk within six weeks.
| Metric | Value |
|---|---|
| Cards processed | 1,548 |
| Perfectly recognized (no errors) | 481 cards (31 %) |
| Minimal corrections (≤ 5 % CER) | 926 cards (60 %) |
| Average error rate (CER) | 7.27 % |
| Fields recognized perfectly | 89 % |
| Time for QC | approx. 6 weeks |
Note on the error rate: the measurement is conservative. During quality control, additional errors in the original finding aid were corrected or updated (e.g., spellings, classifications). These improvements are not calculated separately and therefore appear in the CER statistic, meaning the actual data quality is higher.
The pilot clearly shows: AI excels where it takes on rote tasks and creates more time for professional work. From 1,548 index cards, searchable datasets were produced in a matter of weeks – without aiming to replace human expertise, but rather to strengthen it. For the project manager, this represents a step toward a modern, digitally operating archive that makes the possibilities of data processing usable for archival purposes.
Your 30-minute “Status & Goals Briefing” brings clarity:
Outcome: A clear picture of your situation and a realistic initial assessment – what is possible? Which next steps, e.g. an in-depth workshop or a pilot, are worthwhile?