11/08/2026

The Backlog Sprint: Clearing an Archival Digitization Backlog with AI

Every archivist we talk to has the same number in the back of their mind: the years of accumulated, undescribed, unsearchable material in storage. Boxes accessioned but never processed. Collections scanned a decade ago into TIFFs nobody can search. Finding aids that stop at the folder level, or at 1987.

The standard response has been resignation — the backlog is a permanent condition, chipped at when an intern appears. This post argues for a different model: the backlog sprint, a scoped, time-boxed, AI-assisted push that takes one defined slice of the backlog to finished — searchable, described, in the catalog — in 8–12 weeks.

Why "sprint" and not "program"

Archives have been burned by big digitization programs: multi-year plans, infrastructure purchases, and dashboards that describe a pipeline rather than deliver a product. The sprint model inverts the priorities:

  • One slice, done completely, beats the whole backlog advanced by 4%.
  • A fixed end date forces every decision toward shipping: if a step doesn't get material into the catalog, it's out of scope.
  • A visible result — this collection is now searchable, here is the before and after — is what convinces boards and funders to pay for the next slice.

Archival theory quietly agrees. The "More Product, Less Process" school has argued for minimal viable description for two decades. AI changes the economics of that argument: minimal human effort no longer means minimal description, because machines now draft the description and humans review it.

What AI changed, specifically

Three capabilities matured enough in the last few years to make the sprint model work on real archival material:

Reading at scale. Modern OCR and handwritten-text recognition can produce usable text layers from printed and much handwritten material — including historical Hebrew scripts, where off-the-shelf tools long failed. "Usable" is doing real work in that sentence: the point is not perfection, it is a searchable layer with known, honestly measured confidence.

Drafting descriptions. Given the text layer, language models draft item- and folder-level descriptions: dates, places, persons, document types, summaries. Drafts — every one flagged as machine-generated until a human approves it. The practical shift for staff is from writing descriptions to reviewing them, which is several times faster.

Connecting records. Entity extraction links the same person, family, or organization across documents — the beginning of a knowledge-graph layer over the collection, and the feature researchers notice first.

The anatomy of a sprint

The shape we recommend — and run — has five phases:

1. Choose the slice (week 1). One collection or series with a clear boundary, real user demand, and manageable risk. Not the hardest material; the material where success will be visible.

2. Feasibility sample (weeks 1–2). Take a genuinely representative sample of pages through the whole pipeline before committing to the whole slice. This is where you discover the second script, the bleed-through, the misfiled folders — while the discovery is cheap. The sample sets the accuracy expectations and the review budget for everything that follows; skipping it is the single most common cause of failed digitization projects.

3. Machine pass (weeks 2–5). Scan audit and batch processing: image cleanup, OCR/HTR, entity extraction, draft records. This phase is compute-bound and runs mostly unattended.

4. Human review (weeks 4–10, overlapping). Subject experts review the drafts — confirming, correcting, rejecting. Review is prioritized by confidence score and by user value: low-confidence records and high-demand material first. This phase is the real bottleneck, and the sprint is scoped around it from the start.

5. Import and handover (weeks 10–12). Approved records go into the institution's own catalog in standard formats (Dublin Core, MODS, EAD — whatever the preservation setup expects), machine-generated fields flagged as such. The handover includes the audit trail: which model versions, which confidence thresholds, who reviewed what.

Notice what is not in the list: buying a platform, hiring engineers, migrating systems. A sprint leaves your infrastructure alone and your catalog richer.

Where the effort actually goes

The compute is the cheap part — for a typical sprint slice, OCR and language-model processing are a minor line item. The real budget goes to three places: preparation of problem material, the human review hours, and project management of the review loop. This is also the honest argument for the sprint's fixed scope: since expert review is the scarce resource, a plan that processes ten times more material than your experts can review has not actually described anything — it has produced a new, digital backlog.

A useful rule of thumb when scoping: start from the review capacity you actually have (hours per week of subject-expert time), work backwards to the number of records that capacity can approve, and let that determine the slice — then size the machine pass to match.

After the sprint

The deliverable is the finished slice, but the durable value is what the institution learns: real accuracy numbers on its own material, a calibrated sense of review speed, a documented pipeline, and — not least — a concrete result to put in front of a funder. Institutions rarely stop at one sprint; the first one exists to make the second one fundable.

The backlog stopped being a permanent condition the moment description became cheap to draft. What remains is choosing the first slice.

We scope and run backlog sprints for archives, museums, and research institutions — including the feasibility sample as a standalone first step. Talk to us about your first slice.

Frequently Asked Questions

What is a digitization backlog sprint?

A time-boxed project — typically 8 to 12 weeks — that takes one well-defined slice of an archive's undescribed or unsearchable material through a fixed pipeline: scan audit, OCR/HTR, draft description, human review, and import into the catalog. The goal is a finished, quality-controlled slice, not a partially processed everything.

How much material can a sprint realistically cover?

It depends on material condition and language, which is why the sprint starts with a feasibility sample. As an order of magnitude, machine processing is never the bottleneck — human review is. Whatever your reviewers can verify in the review weeks defines the scope; the pipeline is sized around them, not the other way around.

Does AI-assisted description meet archival standards?

The output format can be fully standards-compliant — Dublin Core, MODS, or EAD finding aids. The professional standard that matters is disclosure: machine-generated fields should be flagged as such, with confidence levels, so future archivists know which records were auto-drafted and human-approved.

Is a sprint worth it for a very small archive with no technical staff?

A sprint is specifically designed for that case: the technical pipeline is brought in from outside for a fixed period, and what stays behind is searchable text, draft records in your catalog, and a documented process. You need subject experts for review, not engineers.

What should NOT go into a first sprint?

The hardest material: badly damaged items, rare scripts with no trained models, legally restricted collections. A first sprint should take material where success is likely and visible — the credibility it builds is what funds sprint number two, which is when the hard material's turn comes.