Our documents are scanned, faxed or handwritten. Can we still test AI on them?

Resolution decides how much of each character survives the trip into a file. The National Archives' rule for digitizing permanent federal records, 36 CFR 1236.50, sets a floor of 300 pixels per inch for modern office paper, in color or grayscale, and 400 for records with fine detail, going higher still where the smallest detail that matters measures a millimeter or less. When a scan falls short of that care, the damage shows in small type: a dropped decimal point, a 3 that could be an 8, the unit printed under a quantity.

A fax is coarser by design. The international fax standard, ITU-T T.4, sets its ordinary mode at 3.85 scan lines per millimeter, about 98 to the inch from top to bottom, and 1,728 picture elements across a 215 millimeter line, about 204 to the inch, each one black or white. Finer modes are optional. So a standard fax holds roughly a third of the archive floor vertically and no gray at all, and a fax of a photocopy loses more at every pass, often with the sender's header line printed across the top of the page.

That doesn't mean a faxed page fails. It means faxes should be a group of their own in the test, so your reviewer can see whether misses gather there.

Color often carries meaning a black-and-white scan throws away, and some forms rely on that. The IRS rules for substitute information returns, Publication 1179 (revised July 2026), require Copy A of most Forms 1099 to be printed in a red OCR dropout ink, so the agency's scanners drop the printed form and keep only the entries, and the IRS warns that a Copy A printed from its own PDF can't be scanned. On ordinary paperwork the same effect works against you: a red PAID stamp, a blue approval initial or a yellow highlight can vanish from a black-and-white scan, or turn into a dark block over the figure beneath. The archives rule requires color capture whenever color is needed to understand the record.

Stamps also land on top of values. A RECEIVED stamp across an invoice date leaves two sets of characters in one place, and a good draft separates them or says it couldn't.

Orientation and framing are the quiet problems. The archives rule has inspectors look for improper cropping, unexpected orientation, and flipped, inverted or skewed images. On a desk that means a landscape page among portrait ones, a two-sided original copied on one side, or a page fed at an angle with its bottom line, often the total, cut off.

Handwriting is several problems sharing one name. The federal digitization guidelines known as FADGI (third edition, 2023) treat handprint or handwritten text recognition as a process related to, but separate from, reading machine print. Neat capitals in a labeled box, a quantity crossed out and rewritten, initials beside a changed price and a cursive note across a margin are four different tasks, and they deserve four separate scores.

The Library of Congress still treats hard handwriting as work for people. It started its By the People program in 2018 for handwritten and complex typed pages that automated transcription handled poorly, and every transcription it returns to loc.gov is made, then reviewed, by volunteers. A test should keep that habit: the draft reads, and someone who knows the documents checks.

Carbon and carbonless copies, the pink and yellow sheets of a delivery ticket or work order, can be the hardest pages on a desk. FADGI groups carbon copies with manuscripts as documents with poor legibility or diffuse characters, and lists sheet-fed scanners as not recommended for them. If your team works from the third copy of a form, test the third copy.

A test built only from clean PDFs describes a better inbox than yours. Start from a recent month, sort the pile by how each document reached you, then choose:

• Send each one as it arrived. The fax as the fax server saved it, the photo as it was texted, the scan as the copier emailed it. Don't rescan, straighten or clean anything: a live run would read what really comes in.
• Match the mix, then add the worst. Take roughly the share of faxes, scans, photos and clean files your desk sees, then add a few your team remembers for the wrong reasons.
• Tag every document. One line each: fax, phone photo, sideways, stamp over a value, handwritten change, carbon copy. When a draft misses, the tag says why that page was there.
• Keep each document whole. Every page, the back when it carries writing, and the cover sheet or email it came with. The archives rule asks for related envelopes and notes to be captured with a record; on a desk, the instruction that matters is often on the sticky note.
• Scan paper-only originals the way the live run would. Use the scanner and settings your team uses day to day, ideally in color or grayscale, and mark which pages were scanned for the test.

Your reviewer, the person on your team who knows these documents best, should judge each page before judging the draft. The IRS rules for keeping books and records in electronic storage, Rev. Proc. 97-22 (1997), give two tests worth borrowing. A page is legible when an observer can identify every letter and numeral positively and quickly, to the exclusion of all others. It is readable when the observer can recognize groups of letters and numerals as words or complete numbers. Mark each field that matters as clear, hard or unreadable to a person before the draft is opened.

Two questions belong on the sheet for every document too. Did the draft take the handwritten change over the printed value, as your team did? Did it read a stamp as what it means, a received date or a paid status, rather than stray text? Sort the finished sheet by tag: a result that holds up on scans and falls apart on faxes shows where a live run could start and what stays with a person.

Then give every field in the draft one of four marks:

• Read right. The value matches what your team keyed.
• Wrong, but flagged. The draft missed and said it wasn't sure, so a person would have looked.
• Wrong, and stated as fact. A confident value the page doesn't support. Count these separately: a busy desk lets them through.
• Declined. The draft gave no value and asked for a person. Where your reviewer also marked that field unreadable, this is the right answer.

Most hard documents belong in the first test. A few are better kept for a second round, a decision about order, not a verdict:

• Your own team can't read it reliably. If most of these end in a phone call to the sender, the answer isn't on the page. Even the IRS expects some paper not to survive imaging: Rev. Proc. 97-22 recommends keeping paper copies of records that are illegible or can't be transferred accurately and completely.
• Nobody recorded what was done with it. A handwritten note that shaped a decision but was never keyed leaves nothing to compare a draft with.
• It is really a different job. Free-form letters or field notes in cursive are a separate task from forms with labeled boxes, and mixing the two blurs both scores.
• The only copy has faded. A thermal printout or a carbon that has faded past reading raises a records question for your team before it raises a test question.

Our free AI proof of concept runs on copies of 20 to 30 documents your team already finished for one job, sent as they arrived and only after an NDA is signed. Nothing logs in to your systems, and no draft reaches a customer, a vendor or a ledger. If the pages can't leave the building, we build look-alikes with your reviewer, printed, stamped, faxed or photographed the way your real documents travel.

Before it starts, the owner names the reviewer, the success criteria go in writing with the tags above in them, and you get a price range for the live job. After a yes, the job runs on new documents for about 30 days, with a person approving every item before it posts.