{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"scanning-the-paper-drawer","title":"Turning paper into rows you can check","url":"https://addaly.com/learn/ai-for-small-business/scanning-the-paper-drawer","position":24,"of":88,"minutes":9,"answer":"OCR combined with a language model produces clean, plausible text rather than obvious garbage, so a misread digit arrives repaired and unmarked.","keyIdea":"OCR combined with a language model produces clean, plausible text rather than obvious garbage, so a misread digit arrives repaired and unmarked, which is why extracted figures get checked against the paper and against an external total, never against how sensible they look."},"body":{"format":"markdown","text":"## The job, and the error rate\n\nA shoebox of receipts. A stack of delivery notes. Three years of handwritten job cards. Turning these into something searchable is one of the most satisfying things you can do with a phone camera, and it comes with an error rate you should know before you start.\n\nOptical character recognition on **clean printed text** (a laser-printed invoice, photographed flat in decent light) is very good. Well above 99% of characters, which means a typical page has a few errors or none.\n\nOn a **thermal till receipt** that has faded, or a page photographed at an angle in a dark stockroom, accuracy falls sharply. Numbers suffer worst, because 3 and 8, 5 and 6, 1 and 7, 0 and O are exactly the pairs that degrade first.\n\nOn **handwriting** it depends enormously on the hand. Neat block capitals do reasonably well. Ordinary cursive is far worse, and one person's habitual way of writing a 7 can produce a consistent, systematic error across three years of records.\n\nModern tools that combine OCR with a language model do better than old OCR, because context helps: a model reading \"Total: 1,2SO\" can work out what was meant. That same helpfulness is a hazard, which is the point of this lesson.\n\n## The specific danger\n\nOld OCR gave you garbage that looked like garbage. `T0ta1: 1,2SO` is obviously wrong.\n\nA model reading the same receipt gives you `Total: 1,250`: clean, plausible, and possibly not what the paper says. It has repaired the text using what usually follows, which is the same mechanism as everything else in this course. Sometimes the repair is right. Sometimes it turns 1,280 into 1,250 and there is nothing on the screen to indicate it.\n\nThis means the check has to be against the paper, not against the plausibility of the output.\n\n## How to do it so the checking is cheap\n\n**Photograph properly.** Flat, no shadow, whole document in frame, in daylight if you can. Ninety seconds of care here removes most of the errors downstream.\n\n**Ask for structure, not prose.** Request specific fields (date, supplier, total, tax, invoice number) as a list. Structure is easier to check than a paragraph, and easier to paste into a spreadsheet.\n\n**Insist on a blank rather than a guess.** *If a field is unreadable, write UNREADABLE. Do not estimate.* Without this, illegible fields come back filled in.\n\n**Check totals against something external.** The supplier statement, the bank feed, the card statement. This is the real safety net, and it catches errors regardless of where they came from.\n\n**Sample the rest.** For a batch of 200 receipts, check 20 against the paper properly. If all 20 are right, your error rate is probably low. If two are wrong, that is roughly one in ten across the whole batch, and you have a decision to make about the other 180.\n\n## Free tools that do the job\n\n- **Your phone.** Google Lens, Apple's Live Text, and the scanning built into Google Drive and OneDrive all extract text at no cost.\n- **Tesseract**, free and open source, runs on your own computer and never sends anything anywhere. Excellent on clean print, poor on handwriting.\n- **Your accounting software.** Zoho, QuickBooks, Xero, Wave and others have receipt capture built in, usually included. This is nearly always the best option, because the extracted data lands directly where it needs to be with no second copy.\n\nPaid document-processing services exist and are more accurate on messy input. They are worth it at thousands of documents a month and are hard to justify below that.\n\n## Keep the paper\n\nOr at least keep the images. Most tax authorities accept digital copies, and most require you to keep something for a number of years: commonly five to seven, differing by country.\n\nThe practical reason is narrower than compliance. When a figure is queried in eighteen months, the photograph is the only way to settle it, and by then nobody remembers whether that receipt was faded.\n\n## Two jobs, not one\n\nIt helps to separate them, because they have different failure modes.\n\n**Extraction** is turning marks on paper into characters. This is where photograph quality, faded thermal paper and handwriting decide your accuracy, and where the errors are mechanical.\n\n**Interpretation** is deciding that the number next to the word `TOTAL` is the total, that `12/03` is a date, and which of the three numbers on the receipt is the tax. This is where a language model helps enormously and where it also guesses.\n\nKeeping them separate tells you where to look when something is wrong. A total that is out by one digit is usually extraction. A total that has been read from the wrong line is interpretation, and it repeats across every receipt from that supplier, which makes it both more damaging and easier to spot once you know to sort your check by supplier.\n\n## What to do with the output\n\nStraight into a spreadsheet or your accounting software, with two extra columns: the source image filename, and a blank column for `checked`.\n\nThe filename column is what lets you settle a query in eighteen months without hunting through a phone gallery. The checked column is what stops the sampling exercise turning into a vague memory that most of them were fine."},"check":{"question":"A workshop extracts 200 supplier invoices with a phone scanner and an AI tool. The output is tidy and every field is filled in. Which fact should worry them most?","options":[{"text":"Phone cameras produce lower-resolution images than a flatbed scanner, so character-level accuracy will be reduced across the whole batch.","correct":false,"tempting":"Resolution matters at the margins, but a well-lit phone photograph of printed text reads very accurately, and this does not explain why the tidiness itself is a warning sign."},{"text":"Two hundred documents is too many to check individually, so some error rate has to be accepted as the cost of the exercise.","correct":false,"tempting":"Accepts an unknown error rate as inevitable when sampling twenty of them would measure it in half an hour and turn a guess into a number."},{"text":"Every field being filled in means nothing was reported as unreadable, so any illegible figures were repaired into plausible numbers rather than flagged.","correct":true},{"text":"Invoices from different suppliers use different layouts, so the extraction tool may have mapped fields inconsistently and put totals in the wrong column across parts of the batch.","correct":false,"tempting":"A real problem with varied layouts that a structured request largely solves, and one that tends to produce visibly odd output rather than the uniformly tidy result described."}]},"prev":{"slug":"the-tools-that-hold-your-files","title":"Four ways to keep a small library, and what each costs","url":"https://addaly.com/learn/ai-for-small-business/the-tools-that-hold-your-files"},"next":{"slug":"contracts-and-long-documents","title":"Reading a contract you were going to skim anyway","url":"https://addaly.com/learn/ai-for-small-business/contracts-and-long-documents"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Turning paper into rows you can check', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/scanning-the-paper-drawer (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Turning paper into rows you can check","courseTitle":"AI for a Small Business","module":{"index":3,"slug":"your-own-knowledge","title":"Your own documents, and getting straight answers out of them"},"answer":null,"keyIdea":"OCR combined with a language model produces clean, plausible text rather than obvious garbage, so a misread digit arrives repaired and unmarked, which is why extracted figures get checked against the paper and against an external total, never against how sensible they look.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-job-and-the-error-rate","runs":[{"text":"The job, and the error rate"}]},{"t":"p","runs":[{"text":"A shoebox of receipts. A stack of delivery notes. Three years of handwritten job cards. Turning these into something searchable is one of the most satisfying things you can do with a phone camera, and it comes with an error rate you should know before you start."}]},{"t":"p","runs":[{"text":"Optical character recognition on "},{"text":"clean printed text","b":true},{"text":" (a laser-printed invoice, photographed flat in decent light) is very good. Well above 99% of characters, which means a typical page has a few errors or none."}]},{"t":"p","runs":[{"text":"On a "},{"text":"thermal till receipt","b":true},{"text":" that has faded, or a page photographed at an angle in a dark stockroom, accuracy falls sharply. Numbers suffer worst, because 3 and 8, 5 and 6, 1 and 7, 0 and O are exactly the pairs that degrade first."}]},{"t":"p","runs":[{"text":"On "},{"text":"handwriting","b":true},{"text":" it depends enormously on the hand. Neat block capitals do reasonably well. Ordinary cursive is far worse, and one person’s habitual way of writing a 7 can produce a consistent, systematic error across three years of records."}]},{"t":"p","runs":[{"text":"Modern tools that combine OCR with a language model do better than old OCR, because context helps: a model reading “Total: 1,2SO” can work out what was meant. That same helpfulness is a hazard, which is the point of this lesson."}]},{"t":"h2","id":"the-specific-danger","runs":[{"text":"The specific danger"}]},{"t":"p","runs":[{"text":"Old OCR gave you garbage that looked like garbage. "},{"text":"T0ta1: 1,2SO","code":true},{"text":" is obviously wrong."}]},{"t":"p","runs":[{"text":"A model reading the same receipt gives you "},{"text":"Total: 1,250","code":true},{"text":": clean, plausible, and possibly not what the paper says. It has repaired the text using what usually follows, which is the same mechanism as everything else in this course. Sometimes the repair is right. Sometimes it turns 1,280 into 1,250 and there is nothing on the screen to indicate it."}]},{"t":"p","runs":[{"text":"This means the check has to be against the paper, not against the plausibility of the output."}]},{"t":"h2","id":"how-to-do-it-so-the-checking-is-cheap","runs":[{"text":"How to do it so the checking is cheap"}]},{"t":"p","runs":[{"text":"Photograph properly.","b":true},{"text":" Flat, no shadow, whole document in frame, in daylight if you can. Ninety seconds of care here removes most of the errors downstream."}]},{"t":"p","runs":[{"text":"Ask for structure, not prose.","b":true},{"text":" Request specific fields (date, supplier, total, tax, invoice number) as a list. Structure is easier to check than a paragraph, and easier to paste into a spreadsheet."}]},{"t":"p","runs":[{"text":"Insist on a blank rather than a guess.","b":true},{"text":" "},{"text":"If a field is unreadable, write UNREADABLE. Do not estimate.","i":true},{"text":" Without this, illegible fields come back filled in."}]},{"t":"p","runs":[{"text":"Check totals against something external.","b":true},{"text":" The supplier statement, the bank feed, the card statement. This is the real safety net, and it catches errors regardless of where they came from."}]},{"t":"p","runs":[{"text":"Sample the rest.","b":true},{"text":" For a batch of 200 receipts, check 20 against the paper properly. If all 20 are right, your error rate is probably low. If two are wrong, that is roughly one in ten across the whole batch, and you have a decision to make about the other 180."}]},{"t":"h2","id":"free-tools-that-do-the-job","runs":[{"text":"Free tools that do the job"}]},{"t":"ul","items":[[{"text":"Your phone.","b":true},{"text":" Google Lens, Apple’s Live Text, and the scanning built into Google Drive and OneDrive all extract text at no cost."}],[{"text":"Tesseract","b":true},{"text":", free and open source, runs on your own computer and never sends anything anywhere. Excellent on clean print, poor on handwriting."}],[{"text":"Your accounting software.","b":true},{"text":" Zoho, QuickBooks, Xero, Wave and others have receipt capture built in, usually included. This is nearly always the best option, because the extracted data lands directly where it needs to be with no second copy."}]]},{"t":"p","runs":[{"text":"Paid document-processing services exist and are more accurate on messy input. They are worth it at thousands of documents a month and are hard to justify below that."}]},{"t":"h2","id":"keep-the-paper","runs":[{"text":"Keep the paper"}]},{"t":"p","runs":[{"text":"Or at least keep the images. Most tax authorities accept digital copies, and most require you to keep something for a number of years: commonly five to seven, differing by country."}]},{"t":"p","runs":[{"text":"The practical reason is narrower than compliance. When a figure is queried in eighteen months, the photograph is the only way to settle it, and by then nobody remembers whether that receipt was faded."}]},{"t":"h2","id":"two-jobs-not-one","runs":[{"text":"Two jobs, not one"}]},{"t":"p","runs":[{"text":"It helps to separate them, because they have different failure modes."}]},{"t":"p","runs":[{"text":"Extraction","b":true},{"text":" is turning marks on paper into characters. This is where photograph quality, faded thermal paper and handwriting decide your accuracy, and where the errors are mechanical."}]},{"t":"p","runs":[{"text":"Interpretation","b":true},{"text":" is deciding that the number next to the word "},{"text":"TOTAL","code":true},{"text":" is the total, that "},{"text":"12/03","code":true},{"text":" is a date, and which of the three numbers on the receipt is the tax. This is where a language model helps enormously and where it also guesses."}]},{"t":"p","runs":[{"text":"Keeping them separate tells you where to look when something is wrong. A total that is out by one digit is usually extraction. A total that has been read from the wrong line is interpretation, and it repeats across every receipt from that supplier, which makes it both more damaging and easier to spot once you know to sort your check by supplier."}]},{"t":"h2","id":"what-to-do-with-the-output","runs":[{"text":"What to do with the output"}]},{"t":"p","runs":[{"text":"Straight into a spreadsheet or your accounting software, with two extra columns: the source image filename, and a blank column for "},{"text":"checked","code":true},{"text":"."}]},{"t":"p","runs":[{"text":"The filename column is what lets you settle a query in eighteen months without hunting through a phone gallery. The checked column is what stops the sampling exercise turning into a vague memory that most of them were fine."}]}],"check":{"question":[{"text":"A workshop extracts 200 supplier invoices with a phone scanner and an AI tool. The output is tidy and every field is filled in. Which fact should worry them most?"}],"options":[{"text":[{"text":"Phone cameras produce lower-resolution images than a flatbed scanner, so character-level accuracy will be reduced across the whole batch."}],"correct":false,"tempting":[{"text":"Resolution matters at the margins, but a well-lit phone photograph of printed text reads very accurately, and this does not explain why the tidiness itself is a warning sign."}]},{"text":[{"text":"Two hundred documents is too many to check individually, so some error rate has to be accepted as the cost of the exercise."}],"correct":false,"tempting":[{"text":"Accepts an unknown error rate as inevitable when sampling twenty of them would measure it in half an hour and turn a guess into a number."}]},{"text":[{"text":"Every field being filled in means nothing was reported as unreadable, so any illegible figures were repaired into plausible numbers rather than flagged."}],"correct":true,"tempting":null},{"text":[{"text":"Invoices from different suppliers use different layouts, so the extraction tool may have mapped fields inconsistently and put totals in the wrong column across parts of the batch."}],"correct":false,"tempting":[{"text":"A real problem with varied layouts that a structured request largely solves, and one that tends to produce visibly odd output rather than the uniformly tidy result described."}]}],"order":[1,3,0,2],"why":null},"prev":{"slug":"the-tools-that-hold-your-files","title":"Four ways to keep a small library, and what each costs","minutes":9},"next":{"slug":"contracts-and-long-documents","title":"Reading a contract you were going to skim anyway","minutes":10}}}