Off-the-Shelf Platform or Custom AI Pipeline? A Decision Guide for Archives
A new generation of SaaS platforms promises archives AI cataloging out of the box: upload your scans, get back searchable text, tags, and catalog records. At the other pole are custom pipelines — models selected and fine-tuned for one institution's specific material. We build the latter, so discount our bias accordingly; but the honest answer to "which should we choose?" is it depends on your material — and the dependence is predictable enough to write down.
What platforms are genuinely good at
Credit first. A well-built archive platform gives a small institution things it could never build alone: a hosted repository, user management, web access for researchers, standard workflows, and baseline OCR — all for a subscription instead of a project budget. For collections of modern, printed, single-language documents in reasonable condition, the AI features work roughly as advertised, because that is the material the underlying models were trained on.
If that describes your collection, and your pain is workflow — intake, description queues, researcher access — a platform is probably the right call, and a custom pipeline would be over-engineering.
Where the promise breaks
Platforms sell a universal pipeline. Archives, almost by definition, hold non-universal material. The gap shows up in predictable places:
Historical scripts and typefaces. Generic OCR degrades sharply on pre-modern print and handwriting. On historical Hebrew scripts — Rashi, rabbinical cursive, Yiddish typefaces — off-the-shelf models are often unusable, and no platform toggle fixes that; it takes training on your kind of pages.
Mixed-language pages. Letters that switch between Hebrew, Russian, and German mid-page; documents with Latin and Cyrillic stamps on a Hebrew form. Universal pipelines pick one language and quietly corrupt the rest.
The long tail of damage. Bleed-through, fading, water damage, tight bindings. Each defect is solvable with targeted preprocessing — but "targeted" is exactly what a universal pipeline is not.
Accuracy accountability. Platforms report that processing completed, rarely how well. For scholarly use the difference between 85% and 97% character accuracy is the difference between a search that finds things and one that silently misses them. A serious workflow needs calibrated confidence reporting, which most platforms do not expose.
The questions that decide it
Four questions sort most cases:
1. Would a generic model read your material? Run a feasibility sample — a representative set of difficult pages through a trial account, measured against hand-made ground truth. Above roughly 90% character accuracy on the sample, a platform is viable; far below it, no amount of workflow polish will compensate.
2. Who owns the enrichment? Images you can always get back. The question is the text, metadata, and AI-generated enrichment: can you export all of it in standard formats — Dublin Core, MODS, EAD finding aids, ALTO for page text? If the enrichment cannot leave, the subscription is not a service, it is custody of your catalog. This single contract question outweighs every feature comparison.
3. Where must the data live? Restricted collections — personal records, testimony, legally protected material — may not be allowed on shared cloud infrastructure at all. Custom pipelines can run on-premises or in a private cloud; few platforms can.
4. Is search the goal, or research? Platforms deliver search. If the goal is a research instrument — entity linking across the collection, question-answering over the full text, corpus-wide analysis — you are describing a pipeline shaped like your research questions, which is precisely what cannot be bought off a shelf.
The hybrid route most institutions actually take
The choice is rarely all-or-nothing, because collections are rarely uniform. The pattern we see work:
- The easy majority — modern, printed, single-language — goes through a platform (or the institution's existing collection-management system) with baseline OCR. Cheap, fast, good enough.
- The hard fraction — historical scripts, mixed languages, damaged or high-value items — goes through a custom pass: fine-tuned recognition, targeted preprocessing, confidence-scored output.
- The results converge in the institution's own catalog, in standard formats, with machine-generated fields flagged. The platform serves access; the custom work serves accuracy; neither owns the data.
Economically this usually beats either pure strategy: custom processing is a one-time project cost applied only where it pays, while the subscription covers the bulk at commodity rates. It also sequences well — a scoped pilot on the hard fraction produces the accuracy numbers you need to plan the rest, before any long-term commitment is signed.
The uncomfortable summary
Platforms are the right answer for standard material and workflow pain. Custom pipelines are the right answer when accuracy on difficult material is the bottleneck. The expensive mistakes are the crossovers: paying project rates for material a subscription would have handled — or trusting a universal pipeline with the irreplaceable material it was never trained to read, and discovering the silent misses years later, one failed search at a time.
Measure first. An afternoon with a feasibility sample settles more than a month of vendor meetings.
We run feasibility samples as a standalone engagement — your pages, measured honestly, with a recommendation that may well be "use a platform." Send us a difficult page.
Frequently Asked Questions
When is an off-the-shelf archive platform the right choice?
When your material is what the platform was trained for — typically modern, printed, single-language documents in good condition — and your main pain is workflow and access rather than recognition accuracy. For standard material, a good platform delivers most of the value at a fraction of the setup effort.
When does a custom pipeline become necessary?
When accuracy on your specific material is the bottleneck: historical scripts, mixed languages on one page, handwriting, damaged documents, or specialized terminology. Generic models plateau on such material, and only fine-tuning on your own documents moves the number.
What is the most important contract clause when choosing a SaaS platform?
Export. Before signing, verify you can extract everything — images, full text, metadata, and any AI-generated enrichment — in standard formats (Dublin Core, MODS, EAD, ALTO). If enrichment data cannot leave the platform, the platform owns your catalog, whatever the pricing page says.
Is the hybrid approach more expensive than picking one side?
Usually the opposite. Custom processing is a one-time project cost applied only to the hard fraction of the collection; platform subscriptions are ongoing but cover the easy majority cheaply. Paying custom rates for easy material, or forcing hard material through a generic platform and correcting it manually, are both more expensive than splitting.
How do I test a platform's claims before committing?
The same way you would test a custom pipeline: a feasibility sample. Send a genuinely representative set of your difficult pages through a trial account and measure the results against a hand-made ground truth. An hour of measurement beats any demo.