What prompt makes an AI do OCR well?
Paste this into Claude, ChatGPT, Gemini or Whizi, attach the image or scan, and fill the brackets. It tells the model to copy rather than fix, and gives it a safe way to say "I can't read this".

Transcribe all the text in the attached [photo, screenshot or scanned page] exactly as written, in reading order. Keep line breaks and headings. Don't correct spelling, grammar or numbers. Where a word or digit is not clear, write [unclear] and never guess. Then list every number and name you are less than sure about.
The first two sentences stop the model from tidying the text, which is how a "7" quietly becomes a "1". The last sentence gives you a short list to check by eye instead of rereading the whole page.
What is the difference between AI OCR and traditional OCR?
Traditional OCR finds the characters on a page, while AI OCR also works out what they mean, such as which block is a table or which date is the due date. That is how Docsumo, a vendor, frames the split in its OCR guide. The AI half now includes vision language models that read the whole image like a person reading a letter.
The classic engines are cheap and narrow. Tesseract's documentation says it is optimized to recognize sentences of words, works best at 300 dpi or more, and has known problems with tables. The cloud services bill by the page, and the price climbs when you ask for structure.
Those are dollars per 1,000 pages. Amazon's page states them per page, at $0.0015, $0.015 and $0.05 for the first million pages a month in US West (Oregon), so the per 1,000 figures are our arithmetic. By the same arithmetic, a 500 page archive costs $0.75 as plain text and $25 as forms.
A chatbot reads differently from either. It predicts likely text, so it handles messy layouts well and can also invent a plausible word where the image is poor. Anthropic's vision guide says Claude might hallucinate or make mistakes with low quality, rotated or very small images. A classic engine would return garbage characters there, and a model may return a clean sentence that is wrong. A clean answer is not the same as a right one.
What do good AI OCR results look like? Three worked examples
Each example gives the prompt and what a good answer must contain. The answer shown is an example written for this page, not the output of a named model.
Example 1, a photographed receipt. Prompt: "Read the attached receipt. Make a table with item, quantity and price, copying each number exactly. After the table, add the line items and compare the sum with the printed total. Say whether they match, and mark any unreadable digit as [unclear]." A good answer has one row per printed line, no invented items and a stated result for the sum. An example of a good ending: "Items sum to a different figure than the printed total, and the difference sits next to the row marked [unclear], so check that price."
Example 2, a scanned bank or invoice table. Prompt: "The attached page has a table with [five] columns. Return it as CSV with the header row first. Keep every cell in its own column, don't merge rows, and copy numbers digit by digit. List any row where the column count is not [five]." A good answer has the same number of fields in every row and flags the exceptions itself. Tesseract's docs list tables as a known weak spot, so this is the case to check line by line.
Example 3, a handwritten note. Prompt: "Transcribe this handwritten note. Write [unclear] for any word you cannot read, and don't guess names or numbers. After the transcription, give your best reading of each [unclear] word, labeled as a guess." A good answer keeps your uncertainty visible. If every word comes back confident on messy handwriting, distrust it.
What are the best AI OCR prompts?
These twelve are grouped by the job. Each works as written once the brackets are filled.
Read text
- Plain transcription: "Transcribe the attached image exactly as written, in reading order, keeping line breaks. Don't correct anything. Write [unclear] for any word you can't read."
- Two column page: "This page has [two] columns. Transcribe the left column first, then the right, with a label for each. Keep headings as headings and drop page numbers."
- Another language: "Transcribe the [language] text exactly, then give an English translation underneath. Put a question mark after any word you're unsure of in both versions."
- Screenshot text: "Pull the text out of this screenshot. Ignore menus and ads, keep only the [error message, chat or code], and preserve indentation."
Pull out structure
- Receipt to table: "Read this receipt into a table of item, quantity and price. Copy the numbers exactly, add a total row from the items, and show the printed total beside it."
- Invoice fields: "Extract invoice number, date, vendor, subtotal, tax and total as JSON. Use null for anything not printed. Copy each number digit by digit."
- Table to CSV: "Return the table in the attached scan as CSV with a header row. One cell per column, no merged rows. List any row whose column count differs from the header."
- Business card: "Read this business card into lines for name, title, company, phone, email and address. Leave a field blank if it isn't printed."
Check the result
- Second look: "Here is the image and your transcription. Compare them line by line and list every number and name where you're less than sure, with the line it is on."
- Sum check: "Add up the line items from your table. Compare the sum with the printed total and tell me if they differ and by how much."
- Handwriting: "Transcribe this handwriting. Use [unclear] for words you can't read. Then give a guess for each, labeled as a guess, and mark any name or number as needing my check."
- Scanned PDF, page by page: "Transcribe the attached scanned PDF page by page under the heading Page [n]. Stop after page [five] and wait for me to say continue."
Is there free OCR software, and what does each AI OCR tool cost?
Yes. Tesseract is free, Amazon, Google and Azure each list a free allowance, and chatbots have free plans that accept images. The table lists what each tool states on its own page as of November 2026, except the ABBYY row, which is Tweak Library's report. A cell says "not stated" where the page did not show the figure.
| Tool | Free use | Sign up | Price | Best at |
|---|---|---|---|---|
| Tesseract | Free, open source engine | None, you run it yourself | Free | Clean printed text at 300 dpi or more |
| Google Document AI | First 1,000 pages free | Google Cloud account | $1.50 per 1,000 pages for Enterprise Document OCR | Pages in bulk |
| Amazon Textract | Free Tier lasts three months, 1,000 pages a month of plain text and 100 with forms or tables | AWS account | $0.0015 a page for text, $0.015 for tables, $0.05 for forms | Forms and tables at scale |
| Azure Document Intelligence | 500 pages free a month | Azure account | $1.50 per 1,000 pages for Read, pay as you go, first million pages | Teams already on Microsoft cloud |
| ABBYY FineReader | Not stated | Not stated | $99 a year (Tweak Library's report, not ABBYY's page) | Desktop scanning, with Adobe Acrobat Pro listed at $19.99 a month by the same site |
| ChatGPT | Free plan lists "Limited messages with uploads" | Account | Not stated on the page we could read | One off photos and messy layouts |
| Claude | Free plan lists "Limited usage" every 5 hours and every week; 20 images per turn on claude.ai | Account | Pro $20 a month, or $17 a month billed annually | Long pages and careful transcription |
| Whizi | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card; up to 5 photo uploads a day | Account | Pro $29.99 a month | Asking Claude, GPT and Gemini to read the same page |
The top result for this query, Intelgic's general OCR page, showed a not found screen with a Go to Home link when we read it, so it gives a searcher nothing to try. Tweak Library's list of AI OCR tools, ranked second, reports $19.99 a month for Adobe Acrobat Pro and $99 a year for ABBYY FineReader. Those are that site's figures, and the list says OCR accuracy is up to 95% in its intro and 98 to 99 percent in its FAQ with no source for either.
Whizi loses on three counts. Tesseract needs no account and runs on your own machine. Google and Amazon are priced per page and built for thousands of pages, which a chat app is not. And a dedicated engine returns text without a model's habit of smoothing over a bad scan.
Gemini is left out of the table because we did not read a Gemini page that states its upload limits for this job. Test it with the same prompts if you use it.
Which AI tool has the best OCR?
No page we read gives a fair accuracy ranking, so the honest answer is to test two tools on your own pages. Docsumo's guide gives no comparative accuracy data for most of the tools it lists, and its own 99% claim has no stated method. Treat any single accuracy figure from a vendor as marketing until you have run a sample.
A fast test takes ten minutes. Pick five real pages, including your worst one. Run each through two tools with the same prompt. Count the wrong digits, not the wrong words, because a wrong digit costs money.
Use a dedicated engine when you have thousands of pages, need positions of words or fields, or must keep documents inside your own systems. Use a chatbot when you have a handful of messy pages and want the text cleaned up, translated or put in a table in the same step.
How do you get accurate OCR from AI?
Start with the image. Crop the page, rotate it upright and send the largest clean version you have. OpenAI's vision guide says to enlarge small text, and warns that its models can misread rotated or upside down text.
Then ask for the output you need and give the model a way to admit doubt. Work in small batches, run the second look prompt, and use the sum check. The five steps are above in the how to list, and prompts 9 and 10 do the checking.
For a scanned PDF, tell the model the page count and ask for one page at a time. That keeps the output short enough to compare against the scan, and a skipped page is obvious.
Why is OCR difficult, and is AI replacing it?
OCR is sensitive to the quality of the image. Low resolution, poor lighting and low contrast all reduce recognition accuracy, which is also how Tweak Library answers the same question. Tesseract's guide adds that it works best at 300 dpi or more, so a small phone thumbnail starts at a disadvantage.
AI is not replacing OCR so much as absorbing it. Tweak Library's answer says simple OCR software is being replaced by cloud services such as Amazon Textract, Google Cloud Vision and Tesseract, which muddles a free engine with paid ones. The practical picture is the one in the table above: a classic engine or a cloud service for volume, and a vision model for a few messy pages.
On accuracy, Tweak Library gives 95 percent or more in its intro and 98 to 99 percent in its FAQ, and cites no test for either. Don't trust either number, and don't skip running your own sample.
What mistakes ruin AI OCR results?
Four mistakes cause most bad transcriptions, and each has a quick fix.
- Trusting a clean looking answer. The fix: models may smooth over what they can't read, so ask for [unclear] markers and run prompt 9.
- Uploading a low resolution photo. The fix: rescan or photograph in daylight, flat, at the highest setting. Tesseract's guide wants 300 dpi or more.
- Dropping a whole multi page scan in at once. The fix: send a few pages, check them, then continue.
- Letting the model fix the text. The fix: say "don't correct anything", because an edited transcription is no longer a record of the original.
A fifth is quiet. Reading a table as running text loses the column each number belonged to, so ask for CSV or a table when the layout matters.
Where does Whizi fit as an AI OCR tool?
Whizi is one app instead of a pile of single purpose OCR tools, and it suits you if you want a second model to read the same page. An uploaded image can be attached to a chat message for the model to look at, so every prompt above works as written.
You can also upload PDF, Word and Excel files, and scanned PDFs get a vision transcription pass for pages with no usable text layer. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Side by side Compare puts two models on one prompt, which helps with the second look: have two models read the same receipt and check only where they differ. Compare is a Pro and Powerhouse feature.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Excel and Word files made in chat have no plan gate, so a table you transcribed can come back as a spreadsheet. See Whizi pricing for what each plan holds.
Can OCR be 100% accurate, and where does AI OCR fail?
No page we read promises it, and the vendors' own docs list the cases that fail. Tesseract's guide says there will inevitably be cases where its image processing is not good enough. Anthropic warns about low quality, rotated and very small images. OpenAI lists rotated text, small text and text in other alphabets as weak spots, and says its model may generate incorrect descriptions in some cases.
Handwriting is the least documented. The Tesseract guide does not mention it, and we found no vendor accuracy figure for handwriting on the pages we read. ABBYY's blog describes intelligent character recognition as an AI based extension of OCR that can be applied to handwriting, and gives no accuracy figure for it either. So treat handwriting results as a draft, and test your own writer before you rely on it.
Numbers are the high stakes part. A transcribed account number, dose or invoice total needs a human check against the original every time. For forms you file, or text that goes into a contract, medical record or court document, the original document stays the record. How to fact check AI answers covers the routine, and what is an AI hallucination explains why a model can write a wrong word with a straight face.
Private documents need a decision too. A cloud engine or chatbot receives your page, so read the vendor's data terms before you upload an ID, a medical letter or a client file. For more on what models can do with pictures, see AI that can analyze images, and for a long scanned report, summarize a PDF with AI.
- Crop, rotate and keep the scan at 300 dpi or more
- Tell the model to copy exactly and mark unreadable words [unclear]
- Send a few pages at a time, not the whole scan
- Run the second look prompt and the sum check on every number
- Read the vendor's data terms before uploading a private document
Frequently asked questions
Is there a free AI OCR tool?
Yes. Tesseract is free and open source, and Google lists 1,000 free pages, Azure 500 free pages a month and Amazon a three month Free Tier, as of November 2026. ChatGPT and Claude both have free plans with limits. None of these pages gives an accuracy figure, so test a sample of your own pages.
Can AI read handwriting?
Sometimes, but the pages we read give no accuracy figure for it. Ask for [unclear] markers, never guesses on names or numbers, and check every figure against the original. Treat the result as a draft, especially for messy writing.
Does AI improve OCR?
It improves layout, tables and messy photos, and it adds a new risk. A vision model may write a fluent word where a classic engine would return garbage, and Anthropic warns about this on low quality images. That is why the check pass matters.
Can AI OCR a scanned PDF?
Yes. Attach the PDF, tell the model the page count and ask for one page at a time. On Whizi, scanned PDFs with no text layer get a vision transcription pass. For thousands of pages a per page engine such as Google Document AI or Amazon Textract is cheaper.