Scanning is a manufacturing process. It has inputs of variable quality, a production line, a defect rate, and a finished good that either meets specification or does not. Agencies get into trouble when they treat it as a purchase rather than a process.
This is the operational checklist.
Preparation is most of the labor
Vendors quote per image, which makes the scanning itself look like the main cost. It is not. Preparation typically consumes more staff hours than the scanning does, and it is the part you cannot outsource without also outsourcing judgment about your own records.
- Remove staples, clips, bindings, and sticky notes — and decide in advance whether a sticky note is a record.
- Repair tears, flatten folds, and separate anything that will jam a feeder.
- Batch into logical units that match how the records will be retrieved later, not how they happen to sit in the box.
- Insert separator or target sheets so the machine knows where one document ends and the next begins.
- Flag oversized items — plans, maps, ledgers — for a different device and a different unit price.
- Record what was in the box before it left, so you can prove what came back.
Image specification
Write these numbers into the contract. "Industry standard" is not a specification.
| Material | Resolution | Mode | Notes |
|---|---|---|---|
| Typed office documents | 300 dpi | Bitonal or grayscale | Grayscale survives poor originals better |
| Documents with photos or color seals | 300 dpi | 24-bit color | Color matters for signatures and stamps |
| Small or degraded text | 400–600 dpi | Grayscale | Resolution is what OCR accuracy depends on |
| Engineering drawings and maps | 400 dpi or higher | Grayscale or color | Large-format device; price separately |
| Microfilm and microfiche | 300–600 dpi equivalent | Grayscale | Specialist conversion; sample before committing |
| Photographs | 600 dpi | 24-bit color | Preservation copy, not a working copy |
Two file-format decisions follow from this:
- PDF/A for preservation. A self-contained, archival-profile PDF with embedded fonts and no external dependencies. This is what you keep.
- Tagged, accessible PDF for publication. PDF/A is an archival guarantee, not an accessibility one. A file can be perfectly valid PDF/A and completely unusable with a screen reader. If the document is going on the website, it also needs a tag structure, a reading order, a document title, and a language setting.
These are different requirements. Asking a scanning vendor for "PDF/A with OCR" and expecting an accessible document is the single most common specification error we see. OCR gives you searchable text. Tagging gives you structure. You need both, and most scanning bureaus supply only the first.
In-house or outsourced
The honest comparison is not price per image; it is total cost including staff time, equipment life, and the cost of doing it twice.
| In-house | Outsourced | |
|---|---|---|
| Best for | Ongoing day-forward capture, sensitive records, small steady volume | Large one-time backfile, oversized media, tight deadline |
| Cost shape | Capital plus salary; low marginal cost per image | Per image; scales linearly |
| Speed | Limited by staff availability | High; production floors run shifts |
| Risk | Project stalls when the assigned person leaves | Records leave your custody |
| Quality control | Yours to enforce continuously | Must be contractual and sampled |
Most agencies land on a split: outsource the backfile, keep day-forward capture in-house.
Quality control that actually catches defects
Do not accept a delivery on the basis of a spot check by whoever has time. Write a sampling plan into the contract:
- Sample a fixed percentage of every batch — commonly 2–5%, selected randomly, not by the vendor.
- Inspect against defined defects: missing pages, out-of-order pages, skew beyond tolerance, cropped content, illegible areas, wrong orientation, blank images where content existed, and document boundaries in the wrong place.
- Define batch rejection. A defect rate above the agreed threshold sends the whole batch back for rework at the vendor's cost. Without this clause, sampling has no teeth.
- Verify page counts against the pre-scan inventory. This is why you counted the box before it left.
- Test the text layer, not just the image. Run a sample through a text extraction and read the output. See testing OCR quality.
- Check the metadata. Document title, dates, and index fields are what make the archive usable in five years.
Custody and destruction
If records leave the building, the contract needs to say who is responsible for them: transport method, facility security, whether any work is subcontracted or sent offshore, insurance, and what happens if a box is lost. For records with confidential content, add background-check and confidentiality requirements for the vendor's staff.
Destroying the paper after scanning is a separate legal question from whether the scan is good. In California, destruction of originals after reproduction runs through specific Government Code provisions with conditions attached — do not shred anything until counsel has confirmed that your process satisfies them, and until quality control has actually passed. See our overview of retention schedules.
What to require in the deliverable
- Images at the contracted resolution and mode, in the contracted format
- A searchable text layer, with a stated accuracy target and a sample-based verification method
- An index or manifest: one row per document, with the fields you specified
- A defect report for each batch, including anything the vendor could not capture
- The originals returned in their original order, or documented destruction
- Delivery on media you can actually ingest, with checksums so you can prove nothing was corrupted in transit
Run a paid pilot first. A few thousand images from your worst-condition series, delivered against the full specification, will expose every disagreement about quality while it is still cheap to fix. Agencies that skip the pilot discover the problem at image 200,000.