Last updated

Business Card Reader

A card photo becomes a contact you can save — one credit each.

VisionParse reads a business card into contact fields — name, job title, company, every telephone number with the label printed beside it, emails, websites and the postal address.

We measured three cards through five conditions each: clean, photographed on a desk at an angle under uneven light, held sideways, shrunk to 500 pixels wide and blurred. 152 of 160 fields came back correct. Whatever a card prints stays in the script it prints it in — an Arabic name comes back in Arabic, never quietly translated. Where a card carries no Latin form at all, a transliteration is offered in a field of its own, marked as ours rather than the card’s.

Last updated: 2026-08-05 · Measured by VisionParse engineering

  • AI tier · 1 credit a document
  • JPG · PNG · WEBP · PDF
  • Photos auto-compressed · PDF to 25 pages
  • Deleted within 1h

What this tool does, in numbers.

  • ACCURACY

    152 of 160 fields across fifteen runs — three cards, each clean, photographed at an angle, sideways, at 500 pixels wide and blurred.

  • LANGUAGES

    An Arabic-primary card returned its name, title, company, numbers and email exactly in all five conditions, kept in Arabic rather than translated.

  • LIMIT

    Arabic-Indic numerals in a degraded address lost one digit in three of five renders, while Latin phone numbers stayed exact.

  • SPEED

    A card comes back as contact fields in about two seconds, ready to save as a .vcf.

  • COST

    One AI credit a card, whether it carries one number or five.

  • ACCESS

    Needs an account with AI credits: every paid plan includes them monthly, or a one-time pack adds them without a subscription.

  • RETENTION

    The uploaded card is deleted within an hour, the extracted contact within 30 days, and nothing trains a model.

Where the 152 of 160 fields, fifteen runs comes from.

Every figure on this page traces to a dated run on hardware we name. Nothing here is a vendor estimate.

Measured 2026-08-05

Claim
Cards survive being photographed the way people photograph them — at an angle, sideways, small and slightly out of focus.
Method
AI extraction tier (gemini-3.5-flash-lite), business card schema, one credit. Full methodology.
Sample
Three cards: a Western corporate card with three labelled numbers, a Gulf card carrying the same person in English and Arabic, and an Arabic-primary card whose name, title and address are printed only in Arabic with Arabic-Indic numerals. Each through five conditions — clean, desk photo at an angle at JPEG quality 72, rotated 90 degrees, 500 pixels wide, and blurred.
Measured
2026-08-05 on Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4
Limitations
All three are our own fixtures, though the Arabic-primary one exists because a real card exposed the gap. Handwritten cards, vertical Japanese layouts, QR-only cards and cards whose second language lives on the reverse are unmeasured. Send a stack of real cards and we measure them before you rely on it.
Contact
VisionParse@senithu.lk

What this tool gets wrong

HELD UP

An Arabic card comes back in Arabic. Whatever a card prints stays in the script it prints it in. A real Arabic-primary card taught us that the hard way: it came back translated into English text that appeared nowhere on it, and nothing flagged that we had written it ourselves. Now the printed value stays printed, and where a card carries no Latin form at all a transliteration is offered separately, labelled as ours. On a bilingual card both scripts are kept, because in Doha the Arabic name is the one the local office files under.

KNOWN LIMIT

Two things to glance at: a wordmark logo, and Arabic numerals in a bad photo. Our Western fixture prints HALVORSEN large with MARINE LOGISTICS as a smaller strapline, and all five conditions returned just HALVORSEN — reading the dominant type as the company is defensible and still wrong. Separately, an Arabic address lost a single Arabic-Indic digit in three of five degraded renders (٩٩٢٧٤ read as ٩٩٣٧٤), while the Latin phone numbers on the same card stayed exact in all five. Everything else on all three cards was exact every time.

HELD UP

Numbers keep the label the card printed, and nothing is guessed. A card printing Mobile, Office and Fax returns those; a card printing bare M and T returns M and T, not an expansion. Guessing turns a fax into someone’s mobile. And there is no arithmetic on a card to verify, so the test is refusal: send a receipt to this tool and it answers "receipt" with an empty name rather than inventing a person.

Three steps, no account.

  1. Photograph the card

    Straight from your phone — JPG, PNG, WebP or PDF. Angled, sideways and small all measured fine; large photos are compressed in the browser first.

  2. The AI engine reads it

    One credit a card, about two seconds. Name, title, company, every number with its printed label, emails, websites and address.

  3. Save the contact

    Download a .vcf that opens in your phone and Outlook, or take the JSON straight into your CRM.

What this tool accepts.

PropertyFree tool
EngineAI extraction tier, 1 credit per card
Measured result152 of 160 fields across fifteen runs, three cards (2026-08-05)
Bilingual result50 of 50 on a Gulf English/Arabic card, all five conditions
Arabic-primary result48 of 50 — name, title, company and numbers exact in all five conditions
Fields returnedName as printed, second-script name, our Latin transliteration when the card prints none, job title, department, company, phones with printed labels, emails, websites, address
OutputContact fields on screen, JSON, and a vCard 3.0 .vcf file
Input formatsJPG, PNG, WEBP, PDF
AccessAccount with AI credits — any paid plan, or a one-time pack
File retentionOriginals deleted within 1 hour; results within 30 days

When to use the API instead.

This page converts one file now. To convert files from your own software, the same engine runs over HTTP from Starter upward.

# same engine, from your code
curl -X POST https://api.visionparse.app/v1/ocr \
  -H "x-api-key: $VISIONPARSE_KEY" \
  -F "file=@image.jpg"

→ 202 { "jobId": "c0df8e2c…" }

Use the API when

You convert more than a few files, convert on a schedule, or need results inside another system. Batch, webhooks and bounding boxes are API-only.

Frequently asked

Can I add the card straight to my phone contacts?

Yes — every result offers a .vcf download in vCard 3.0, which iPhone, Android and Outlook all open without a conversion step. Numbers keep the type the card printed, so a fax does not land in your contacts as a mobile.

Does it handle Arabic or other bilingual cards?

That is the case we measured hardest, and the one a real card corrected us on. A card printed Arabic-first comes back in Arabic — name, title and address in the script they were printed in, because translating them silently hands you text that is not on the card. Where a card gives no Latin form at all, a transliteration appears in its own field, marked as ours. A bilingual card keeps both scripts.

Do I have to photograph the card perfectly?

No. The measured conditions were a desk photo at an angle under uneven light, a card held sideways, a 500-pixel-wide crop and a blurred shot. All of them returned the numbers and emails exactly. Fill the frame and it will be fine.

What does it get wrong?

Two things, both repeatable. A logo that prints the company as a large wordmark with a smaller strapline returns just the wordmark — ours did in all five conditions. And Arabic-Indic numerals in a degraded photo can lose a digit: our Arabic address dropped one in three of five renders, while the Latin phone numbers on the same card stayed exact. Glance at those two; everything else was exact.

Why is this not free?

Reading a card into labelled fields is the AI engine’s job, and it costs money per call — free OCR gives you a wall of text, not a contact. So it runs on credits: included monthly with every paid plan, or bought once in a pack.

Measured by VisionParse engineering. Method and raw data are published.Updated Benchmark · Method · Data handling