You've got a stack of cards from last week's event, a scanner app that made every photo look crisp, and a vague worry that one of those contacts now has the wrong phone number. The honest answer is that you can't tell from the photo, and you can't tell from the app's headline percentage either. You find out by comparing the saved record to the printed card, field by field, on cards you actually receive.
A sharp scan can still save the wrong digit
Computerworld's hands-on review of four scanning apps used just nine cards, and it still caught the kind of slip that matters. One address came back with a 5 read as an S, while the rest of that contact was fine. Nothing about that error looks wrong at a glance. It looks like a contact.
A scanner has two jobs, reading the characters and deciding which field each piece of text belongs in. Either can fail quietly, and a clean preview hides both. So the useful question isn't "is this scanner accurate?" It's "does the record I save match my card, in the fields I'll use?"
Pick cards you really get, and use the same set every time
Your test set should look like your pile. If half your cards are two sided, or use a script other than the one you read most, or have small type on a textured background, the set needs those cards in it. Run the identical set through every scanner you're weighing, so a difference in results comes from the scanner and not from the cards. Also record the app version, device and capture setup, selected language, correction settings, and whether anyone reviewed or edited a scan before you scored it. If one is part of the workflow you're judging, note it rather than treating the tests as identical.
There's no magic number. Habsy's guide to scanner accuracy suggests 30 varied cards. A commissioned comparison from Capacitor Partners used 80 cards across several languages with more than 5,500 fields. Computerworld used nine. Those are examples of scope, not a sample size that guarantees anything, so pick a number you'll actually finish and keep it varied.
Write the answer key before you scan anything
Before any card goes under a camera, type out what's printed on it: name, title, company, phone, email, address, website. If a card has no website, write "absent" instead of leaving a blank, so a value the scanner invents later shows up as an error. This reference record is your answer key, and it keeps the scanner's guess from becoming the truth.
Capacitor Partners did the same thing at scale. They transcribed each card field by hand and recorded the output of each app against it. Microsoft's BusinessCardReader exposes separate properties for names, company, title, email, website and more than one phone field, so a mobile and an office number are different answers.
Score what the text says and where it landed
Give every field two checks. First, does the value match the card? Second, is it sitting in the right field? A scan can pass the first and fail the second, like a person's name filed under company, or a mobile number in the office phone slot. A simple row per field is enough: field, printed value or "absent", scanned value, text matches (yes or no), correct field (yes or no).
Keep the two kinds of miss in separate columns, because they have different fixes. A wrong character is a reading problem. A right value in the wrong place is a field mapping problem, and you may be able to correct that one in the app. Capacitor Partners compared text using a Levenshtein score, a count of single character edits, which measures how close the text is but not whether it landed in the right field. They also treated some punctuation, spacing and capitalization changes as harmless. Decide your own tolerance up front, and write it down.
Then tally two numbers: how many required fields matched exactly, and how many whole records came through with every critical field right. Keep them apart. Google's Document AI documentation declines to offer one accuracy figure for extraction, because optional fields and multiple values make a single percentage less meaningful. A marketing number has the same weakness.
Sort the hard cards into their own pile
An average can bury your worst cards. Pull out the ones with a second language, a rotated layout, tiny type, or a decorative design, and score them as their own group. If your pile has many, they're the part of the test that matters most.
For each miss, note what kind it was: a character swapped, text skipped, a value in the wrong field, two values merged, or a language the app doesn't support. ABBYY's Mobile Capture SDK, for instance, recommends a Full HD camera stream and needs English for email and web addresses on cards in non Latin scripts. That is a note about that SDK, not every app.
Some misses are your photo's fault. Google's Enterprise Document OCR can flag blur, glare and cutoff, and Amazon's Textract guidance asks for at least 150 DPI and says not to downsample images before upload. If your results improve with a better photo, you've learned something about your workflow, so log it that way.
Treat a vendor percentage as a question
A percentage only means something if you know what was counted. Ask three things of any figure:
- Does it count characters, fields, or whole cards?
- Which cards, languages, devices and app version were tested?
- Did a human correct anything before it was scored?
Covve's scanner page advertises 96% accuracy, but on the page we checked on October 16, 2026, it doesn't say what the percentage was measured on: no card sample, language mix, app version or capture conditions, and no word on whether correction was included. ABBYY's 2012 launch of Business Card Reader 2.0 for Windows claimed up to 99.8 percent text recognition, footnoted as internal ABBYY testing. Those numbers come from different products, years and methods, so don't pick the bigger one.
Capacitor Partners' reliability index counted cards with every fundamental field right, which is a much harder bar than an average score. Their BizConnect results also included that app's post scan verification, so the number reflects scanning plus checking.
Some services also hand back a confidence score. Amazon's guidance says to weigh that score against error cost, and to route low confidence results to closer human review. Google notes that higher confidence thresholds generally raise precision and lower recall in its Document AI evaluation, which ignores predictions below the threshold. If your scanner shows confidence, a person should look at the fields it flags.
Set a pass rule, then export one record
Here's a rule we'd suggest, and it's our own rather than an industry standard: no silent error in a critical name, phone or email, and one clean export into the intended fields. "Silent" is the key word. A wrong digit the app flags for review is a workflow. A wrong digit it saves without comment is the failure.
Before you trust a batch, run one record through the real destination, whether that's a CSV import, your phone's contacts, or a CRM:
- Pick one card whose reference record you trust, ideally one with two phone numbers.
- Scan it and correct anything the app lets you correct.
- Export or sync it exactly as you would for the full pile.
- Open the result in the destination and check each field against your answer key. Look for names split strangely, phones landing in the wrong field, and fields the destination quietly drops or reshapes.
If a required value is wrong, fix the field mapping, keep a person reviewing every record, or stop before you import the rest. If the app won't show you its raw fields or won't let you correct them, treat any headline score as unproven and have someone check each name, phone and email before saving.
One shortcut applies only to new exchanges. If someone shares a Zapped card by NFC or a Flexible QR code, a Zapped card can go straight to saving the contact, with no scanning step to get wrong. That works only when their card shows Save contact and has no password, and a Main QR code opens the card page instead. The phone receives the same vCard file as the Save contact button: Safari on iPhone shows the contact for you to add, and Android downloads the file for you to open. It doesn't help with paper cards.
Once you've found a workflow you trust, the next question is where all those contacts live, which is covered in how to organize business cards digitally.
