Almost every card scanner claims high accuracy and almost none of them say on what. Accuracy on a clean, single-language, high-contrast card is close to solved. Accuracy on the cards people actually hand you at trade shows is a different number, and the gap between those two is where most of the disappointment lives.
This post is about what actually breaks extraction. We are not going to quote an accuracy figure for our own product, because we have not published a measured one yet, and a number without a stated method and sample is marketing rather than information.
Two different technologies, two different failure modes
Card scanning has historically meant optical character recognition: find the text, read the characters, then apply rules to guess which string is the email and which is the job title. Newer tools use a vision model that interprets the card as a whole.
They fail differently, and knowing which you are using tells you what to check.
That distinction matters more than the accuracy headline. A garbled email address is obviously broken and gets fixed. A confidently wrong job title looks correct, gets imported, and turns up in a sales conversation three weeks later.
The seven cards that break extraction
In rough order of how much trouble they cause.
1. Logos printed over text. A watermark or a large logo sitting behind the contact details is the single most common failure. The characters are physically there and partially obscured, which is harder than characters that are absent.
2. Foil, metallic and spot UV finishes. These are designed to catch light, which is exactly what you do not want under a phone camera in a hall with overhead lighting. The reflection lands across a different part of the card in every photograph.
3. Dark stock with thin light type. Low contrast plus a thin weight is hard for any system. Black cards with grey text are common in design and architecture, and they are consistently the worst performers.
4. Two languages on one card. Extremely common in India, the Gulf, Japan and much of Europe. Frequently the two sides carry different information rather than a translation, so photographing one side loses data that is not recoverable from the other.
5. Handwritten additions. A mobile number written on the back because the printed one is the old office. Almost never extracted, and frequently the most useful number on the card.
6. QR codes replacing text. Increasingly the card carries a code rather than an email address. If the tool does not resolve the code, the field is simply absent.
7. Creative layouts. Vertical text, text set in a circle, or details spread across a die-cut shape. Rare, memorable, and reliably unreadable.
What a confidence score is actually telling you
A per-field confidence score is the most useful thing a card scanner can give you, and it is widely misunderstood. It is not a probability that the value is correct. It is the system’s own uncertainty about its reading, which is a different and more useful thing.
The practical use is triage. On two hundred captures you are never going to check every field, so the question is which forty fields to look at. A confidence score answers exactly that, and it turns verification from an all-or-nothing choice into a ten-minute job. This is the mechanism the list cleaning run order depends on at the verification step.
What it cannot do is catch a plausible misread that the system was confident about. A surname read as a similar surname, or a job title assigned to the wrong line, can come through with high confidence and be wrong. That residual is why the original card image needs keeping, and why a tool that discards the photograph after extraction is doing you a disservice.
What to do about the failures
Nothing here is solved by choosing a different vendor. These are properties of the cards. What varies between tools is how gracefully they handle the failure.
The fourth one is worth dwelling on. If a card is genuinely unreadable, the ten-second voice note is a complete substitute: say the name, the company and how to reach them. It takes less time than squinting at the card and it produces a record you can act on. That habit is the subject of why event leads go cold, and it turns out to solve the extraction edge case as a side effect.
How to evaluate a scanner honestly
If you are comparing tools, do not compare their published accuracy claims, because they are measured on different samples and are not comparable. Run your own test, which takes about twenty minutes.
Collect twenty cards from your own last event. Deliberately include the difficult ones: the dark one, the foil one, the bilingual one, the one with handwriting on the back. That selection matters, because a test on twenty clean cards tells you nothing you did not already know.
Run all twenty through each tool, then score field by field rather than card by card. A card with a correct name and a wrong email is not half right; it is unusable, because the email is the field you need. Weight the fields by what your follow-up actually requires.
Then check what each tool did when it was unsure. That single behaviour, flagging versus guessing, will affect your data quality more than a few percentage points of raw accuracy.
What accuracy means for a booth team in practice
There is a gap between accuracy as a benchmark and accuracy as an experience, and the thing that closes it is not a better model. It is where in the process the error surfaces.
An error caught at the booth costs about five seconds: the bot returns the extracted contact, you notice the company is wrong, you reply with the correction and carry on. An error caught during the evening review costs perhaps thirty seconds, because you have to remember the conversation. An error caught after import costs a CRM edit, and an error caught by the prospect costs the lead.
So the same underlying accuracy produces wildly different outcomes depending on how quickly the system shows you what it read. A tool that extracts silently into a dashboard you check next week is materially worse than one with identical accuracy that replies immediately, even though a benchmark would score them the same.
This is why we care more about the reply-to-correct loop than about the headline figure. The practical accuracy of any capture system is the accuracy after correction, and correction only happens if it is nearly free at the moment of capture.
Why we are not quoting a number
It would be easy to put a percentage in this article. We have not, because we have not yet published a measured accuracy figure with a stated method and sample, and a number without those is not information.
What we can tell you is how the system behaves: it returns a confidence score per extracted field, it keeps the original photograph and audio even when extraction fails, and it lets you correct any field by replying to the bot rather than opening a dashboard. Those are verifiable claims about behaviour rather than unverifiable claims about performance.
When the measurement exists, it will be published with the sample size, the card mix and the failure modes included. Until then, the twenty-card test above on your own cards will tell you more than any vendor claim, including ours. The WhatsApp capture flow is the fastest way to run it, and 20 free scans covers the test with room spare.
What accuracy does not fix
It is worth ending on the limit of this whole discussion. Perfect extraction would still leave you with the smaller half of the problem solved.
A flawlessly extracted card gives you a name, a title, a company, an email and a phone number. Every one of those is durable and none of them tells you why the person was worth contacting. A system that reads cards at a hundred percent still produces leads that read identically to each other, because the difference between leads was never on the card.
This is why we treat the voice note as the primary capture and the card as the supporting one, which is the opposite of how the category usually frames it. The card is the easy half and it is the half that decays least. The reason someone mattered is the hard half, it decays within a day, and no amount of extraction accuracy recovers it.
So when comparing tools, weight the question of what happens to the context at least as heavily as the question of how well the tool reads type. A scanner with excellent accuracy and nowhere to put the conversation is solving the part that was not actually broken.
The short version
Extraction on clean cards is close to solved. Extraction on foil, dark stock, logos over text, bilingual cards, handwriting and QR codes is not, and no vendor choice fixes that because it is a property of the cards. What differs between tools is the failure behaviour: whether they flag uncertainty, whether they keep the original image, and whether correcting a field takes seconds or a CRM edit.
Test on twenty of your own cards including the awkward ones, score field by field, and pay more attention to what happens when the tool is unsure than to the headline number.