A client sends a scanned contract late on Friday, asks for a translation by Monday, and expects a quote based on the translated word count. The scan has no selectable text, so someone has to create a reliable source file before the translator can start. If that preparation time never appears in the estimate, the agency’s margin starts shrinking before translation begins.
Retyping a scan is easy to mistake for a small administrative task. In practice, it can involve file inspection, OCR, correction, table reconstruction, source-word counting, handoff, and a second check after translation. Each step uses time, and each missed step can create rework or a quality problem.
This article breaks down where those costs come from and how to account for them without treating every scan as the same kind of job.
What retyping a scan means in a translation workflow¶
A scanned document is image content that optical character recognition, or OCR, converts into machine-readable text. OCR makes the words available to downstream tools, but recognition is only one part of preparing a scan for translation. The output can still contain misread characters, broken line order, missing table structure, or text that needs human correction. An EDPB-hosted expert report on OCR and named entity recognition describes OCR as a technology that identifies text in images and raises issues that can matter when organizations process personal information.
The phrase “retyping a scan” often covers two different workflows:
- A person reads the image and types the text into a file.
- OCR produces a draft, and a person corrects it against the source image.
The second workflow is not automatically cheaper. OCR can reduce the amount of text a person must type, but review remains necessary. A scan with a clear page layout and crisp printed text can be very different from a skewed photograph with faint print, unusual fonts, handwriting, stamps, or dense tables.
The distinction matters for quoting. “One page” does not tell you how many source words the page contains, how difficult the image is to read, whether the page includes a table, or how carefully the extracted text must be checked. A one-page certificate and a one-page financial statement can need very different preparation and review.
Translation agencies often price translation by source or target word, but not all project work fits that model. Slator reported that 91% of language service providers surveyed used per-word rates in its 2024 Market Report. The same page reports that 14% charged per document and 64% charged by the hour for certain work, showing why a scan-preparation fee can sit alongside a word-based translation price rather than being buried inside it (Slator’s translation-pricing report, 2024 Market Report).
Slator writes: “For decades, per-word rates have been the standard throughout the translation industry.”
The point is not that per-word pricing is wrong. It is that per-word pricing describes one way to bill translation, while scan preparation can involve non-word-based work. A quote that counts only source words can leave the agency paying for the rest.
Where the cost appears before translation starts¶
The most useful way to understand scan-related margin loss is to follow the file from intake to translator handoff. Each stage should have an owner, a cost, and a clear definition of what “done” means.
Intake and file assessment¶
Someone needs to establish what the agency has received. Is the PDF a scan, a text PDF, or a mixture? Are all pages present and readable? Does the file contain printed text, handwriting, stamps, tables, signatures, or images that must remain in the translated document?
The assessment also checks whether the client has supplied the right version. A page that is rotated, cropped, or photographed at an angle can change the preparation approach. A document with unnecessary pages may not need the same processing as a complete archive. Google Cloud Document AI documents page-range selection as a feature that can avoid processing pages that are not needed, depending on product configuration (Enterprise Document OCR documentation, accessed April 18, 2026).
The intake stage costs money even if the agency never retypes a word. A project manager or production coordinator spends time opening files, checking scope, asking questions, and deciding what needs review. If the quote assumes that intake is free, the time still comes from somewhere: another project, the agency’s overhead, or its margin.
OCR setup and processing¶
OCR tools can identify text and layout rather than simply returning a block of words. Google Cloud documents detection of blocks, paragraphs, lines, words, and symbols from PDFs and images. Its listed features also include image-quality scoring, rotation correction, and language detection (Google Cloud Document AI documentation, accessed April 18, 2026).
Those capabilities help, but they do not make every scan equally easy to process. Rotation correction may help with a skewed page. Quality scoring may help route a blurry page to closer review. Neither feature tells the agency that a name, account number, or amount has been transcribed correctly.
Input quality matters. Amazon Textract’s best-practices page says that an image of at least 150 DPI is ideal and advises against converting or downsampling supported PDF, TIFF, JPEG, or PNG inputs before upload. Those recommendations are guidance from AWS documentation accessed April 18, 2026, not a promise that every file at that resolution will produce accurate text (Amazon Textract best practices).
Human correction and proofreading¶
OCR output needs comparison against the image. A reviewer may correct the text directly, mark uncertain fragments, repair paragraph order, or send unclear passages back to the project manager or client. The reviewer also has to decide whether the text is ready for translation or whether the source itself needs clarification.
This is where a hidden cost can grow. A quote might include “OCR” as if it were a single automated action, while the actual job requires a human to check every page. The agency pays for the tool and the reviewer, but charges only for translation.
Review time also depends on the document’s risk. A minor error in a nonessential heading does not have the same effect as a misread amount in a statement or a wrong date in a contract. A single confidence score cannot settle that question. AWS says confidence scores use a scale from 0 to 100 and recommends considering them alongside the sensitivity of the use case. Its documentation gives financial decisions as an example where a threshold of 90% or higher may be appropriate, while making clear that the right threshold depends on the use case (Amazon Textract best practices, accessed April 18, 2026).
File repair, layout, and handoff¶
A translation team needs a usable source, not just recognizable words. OCR may return the correct text but lose the relationships between headings, columns, tables, footnotes, or labels. A reviewer may have to rebuild a table or label each column before the translator can safely work.
That preparation affects later steps. A translator who receives a clean, ordered source can translate with fewer interruptions. A translator who receives fragments from a two-column page may have to infer reading order or ask production to repair the file. Either route adds time.
Layout work can also remain after translation. Text length can change between languages, and table cells may need adjustment. Slator notes that quotes can account for project volume and complexity, graphics localization, and estimated document-formatting effort, rather than word count alone (Slator’s translation-pricing report, page accessed April 18, 2026).
A practical estimate therefore separates at least three questions:
- What effort will make the source text reliable?
- What effort will translate and review the text?
- What effort will prepare the output file in the required layout?
Combining all three into a single per-word price can make the work harder to track. Separate line items show where a job consumes time and make later estimates easier to compare.
How to calculate the cost without hiding the work¶
A workable estimate starts with the task rather than a guess based on page count. The agency does not need a perfect prediction at intake, but it does need to show which assumptions drive the quote.
A basic cost model can use these components:
Total project cost = intake and file assessment + OCR processing + OCR correction + source preparation and layout repair + translation + translation quality checks + project management and delivery.
The formula is not a universal tariff. It is a way to prevent preparation work from disappearing inside a translation rate. An agency can charge some tasks per word, some per page, some per document, and some by time, as long as the quote describes what each charge covers.
For each scan, record the following before approving a fixed quote:
- Count the pages in scope. Identify pages that are blank, duplicated, or outside the requested translation.
- Estimate source volume. Use OCR or a sample page to get a usable word count. Do not treat page count as a word count.
- Classify image quality. Note blur, small print, glare, skew, low contrast, cropping, and other visible issues.
- Describe layout complexity. Record tables, multi-column sections, footnotes, stamps, forms, and elements that need rebuilding.
- Set the review level. Decide whether the reviewer checks every line, samples clean pages, or gives extra attention to names, numbers, and high-risk fields.
- Assign responsibility. Name who handles OCR, correction, source formatting, translation, final review, and client questions.
- State assumptions and exclusions. Explain whether clarification, extra pages, manual transcription, or substantial DTP work changes the quote.
Source-word count deserves particular care. A word counter cannot count text that exists only as pixels. OCR makes the source countable, but the resulting count is only as useful as the extracted text. Missing columns, repeated headers, or recognition errors can distort the number used for pricing.
Slator’s report also describes different billing bases. A 2024 linguist survey cited on the page found that 77% of surveyed linguists were paid by source word and 13% by target word. Those figures describe the survey, not a rule for every agency or language pair (Slator’s translation-pricing report, 2024 survey reported on page accessed April 18, 2026). The practical lesson is simple: make clear whether the translation fee uses source words or target words, and keep scan preparation visible as a separate task when it is not part of that count.
Slator reports that “14% of LSPs surveyed charged per document,” potentially including project minimums.
A per-document fee can fit short files where a word-based rate would not cover setup and handling. It also gives the client a visible price for a defined deliverable. The agency still needs to state what the fee includes; “per document” is not a substitute for scope.
A sample estimate can show the logic without inventing a market rate:
| Work item | What the estimate should describe | Common cost driver |
|---|---|---|
| Intake and assessment | File check, page scope, quality and layout review | Unclear scope or mixed file types |
| OCR processing | Tool, selected pages, output format | Image quality and required configuration |
| OCR correction | Comparison with the scan and corrections | Recognition errors and risk level |
| Source preparation | Reading order, tables, headings, usable handoff | Dense or irregular layout |
| Translation | Source-word, target-word, hourly, or document basis | Language pair, volume, and complexity |
| Final checks and delivery | Completeness, names, figures, file requirements | Required quality level and output format |
The table is a scope checklist, not a price list. Without labor rates, tool charges, and a defined level of review, a universal cash total would mislead more than it would help.
A useful margin check compares estimated effort with actual effort after delivery. Record the time spent on assessment, correction, layout work, and questions separately from translation. If the agency only tracks total project hours, it cannot tell whether a low-margin job came from difficult OCR, a poor estimate, repeated client changes, or slow translation.
OCR is a workflow choice, not a promise of clean text¶
OCR can reduce manual typing, but an agency should not assume that recognition quality is uniform across a document. Text can be easy to read in one part of a page and unreliable in another. A page-level average can hide a single field that matters greatly.
Amazon Textract describes a confidence score as a number from 0 to 100 indicating the probability that a prediction is correct. The same guidance says to use confidence alongside the sensitivity of the task rather than treating the output as error-free (Amazon Textract best practices, accessed April 18, 2026).
AWS documentation says: “A confidence score is a number between 0 and 100 that indicates the probability that a given prediction is correct.”
That score is a signal for review, not a replacement for it. An agency still needs rules for what gets checked and who checks it. A high score on ordinary prose does not eliminate the need to verify a name or number in a consequential document.
Build quality control around the risk and the page features:
- Names and identifying details: compare spellings directly with the scan.
- Numbers, dates, and monetary values: check each value against the source, especially when a misplaced decimal or digit would change meaning.
- Tables: check row and column relationships, not only whether the words appear somewhere in the extracted text.
- Low-quality sections: route blur, glare, very small fonts, and uncertain characters for closer human review.
- Reading order: confirm that multi-column text and footnotes appear in the sequence a reader expects.
- Completeness: check the first and last lines of each page, headers, footers, and any text near the page edges.
Table structure deserves special attention. AWS warns that table extraction can be inconsistent when cells span columns or when cells, rows, or columns differ from other parts of a table. The documentation suggests text detection as a workaround in those situations (Amazon Textract best practices, accessed April 18, 2026). Text detection may preserve the words without restoring the relationships between cells, so a person still needs to inspect the layout.
Google Cloud documents image-quality scoring across eight dimensions, including blur, unusually small fonts, and glare. Such signals can help a team decide which pages need human attention, but they do not confirm that every extracted name or figure is correct (Google Cloud Document AI documentation, accessed April 18, 2026).
The AWS guidance says the appropriate confidence threshold depends on the use case, and gives a financial decision as an example where a threshold of 90% or higher may be appropriate.
The word “may” matters in that guidance. A threshold that fits one use case does not become a universal standard for legal, medical, financial, or administrative translation. Set review rules to fit the project, and record the reason for those rules.
The cost of a missed OCR error can travel downstream. A wrong source number can become a wrong translated number. A missing line can be mistaken for an omission by the translator, while a broken table can make an otherwise correct amount appear in the wrong row. Early source checks often cost less than tracing an error through a translated file, client review, and a revised delivery.
For a detailed discussion of image-based files and their limits, see what AI OCR can and cannot do with translated PDFs and scans. Agencies quoting a file with no selectable text can also use a process for estimating a scanned document’s word count.
When manual retyping makes sense, and when it does not¶
Manual retyping is not automatically wasteful. A human can read context, distinguish some visually similar characters, and identify when a page is too poor to process safely. The question is whether manual transcription is the right way to reach a reliable source for the specific file.
Manual work can make sense when the scan is too poor for dependable OCR, when the text is handwritten, or when the output needs a level of careful interpretation that automated extraction cannot provide. A reviewer can also combine approaches: use OCR for clear printed sections, then transcribe or verify difficult passages manually.
Manual retyping is a poor default when the document is clear, the text is printed, and OCR output can be checked efficiently. Typing every word from scratch may repeat work that a recognition system can do. But “OCR first” should not mean “send the result straight to translation.” The agency still needs a review step that fits the risk, layout, and image quality.
A simple decision path helps:
- Clear printed text and simple layout: run OCR, inspect the output, and correct any errors before translation.
- Clear text with complex tables or columns: run OCR if useful, then check the relationships and rebuild the structure before handoff.
- Mixed quality: use OCR for readable parts and route weak areas for targeted human transcription.
- Handwriting or severely degraded pages: assess whether a human can read the source reliably; ask for a better copy or clarification if not.
- Consequential names, figures, or dates: verify those fields directly against the image regardless of the overall OCR score.
A scan-to-editable workflow can use intermediate files, each with a clear purpose. The first output is an OCR text draft. The next version is corrected against the scan. The final source file is organized for translation, with uncertain passages marked rather than guessed. Agencies that need to compare preparation options can review ways to turn a scanned PDF into an editable Word file before translation.
The case example below illustrates the cost mechanics without claiming a universal time saving or a specific labor rate.
Imagine a small agency receives a scanned statement with printed text and irregular tables. The project manager counts the pages, checks whether the file includes all pages, and requests a clearer image of one section. OCR extracts most of the text. A reviewer then checks names, dates, and amounts against the scan, corrects an unclear table, and creates a structured source file for translation.
A quote that includes only the extracted source words omits at least three jobs: assessing the file, correcting OCR output, and repairing the table. If the reviewer also has to contact the client and wait for a replacement image, project management time enters the calculation too. None of those steps is an unusual surprise; each follows from the input and the scope.
The same logic applies when a translator manually retypes a page as part of translation. The agency should decide whether that work sits inside the translator’s rate or is a separate preparation task. Either choice can be reasonable. Hiding the task makes it harder to compare actual and estimated effort.
Privacy and responsibility belong in the estimate¶
Scanned documents often contain names, addresses, identification numbers, or financial details. The European Commission explains that personal data includes information relating to an identified or identifiable living person and gives names, addresses, and identification-card numbers as examples. A scan and the OCR text extracted from it can therefore contain personal data (European Commission data protection explainer, accessed April 18, 2026).
The Commission describes processing broadly. Collection, recording, organization, storage, alteration, retrieval, consultation, use, disclosure, erasure, and destruction can all count as processing under the GDPR. A scan-to-OCR workflow can involve several of those actions when personal data is present (European Commission data protection explainer, accessed April 18, 2026).
That does not mean every translation agency has the same legal role. The Commission says a controller determines the purposes and means of processing, while a processor handles personal data on the controller’s documented instructions. Agencies should assess their actual role, agreements, and processing arrangements for the workflow they use. The European Commission also says GDPR protection applies to automated and manual processing, so switching from OCR to human typing does not remove personal data from scope.
The European Commission defines personal data as: “any information that relates to an identified or identifiable living individual (data subject).”
That definition makes the practical issue clear: document format does not decide whether information is personal. A paper scan, an OCR text file, and a corrected Word document may all contain the same personal details.
Privacy work can affect project cost in concrete ways. Staff may need to confirm which tool can process the file, who can access it, how it is transferred, and what happens to working copies. The agency may also need to document client instructions or limit which pages go to a processing service. Those tasks take time and should not be assumed to disappear because the OCR itself is automated.
The Commission says that pseudonymized data that can still be used to re-identify a person remains personal data within GDPR scope (European Commission data protection explainer, accessed April 18, 2026). Removing a name from a file is therefore not enough to assume that every privacy obligation has gone away.
An EDPB-hosted expert project examines privacy risks associated with procuring, developing, and using OCR and named-entity-recognition systems. The page states that external contractors prepared the report and that it does not represent the EDPB’s official position (EDPB-hosted expert report, project completed September 2023). Agencies can use that distinction as a reminder to assess tools and contracts on their own facts, rather than treating a report as legal advice.
Privacy review is part of the workflow, not a reason to promise that a file is safe just because it is a scan or because a person typed it manually. Include the relevant handling steps in project planning and make sure the client understands which work is included.
A practical way to protect margin on scanned jobs¶
The best cost-control step is to make assumptions visible before production begins. The agency does not need a complicated pricing system to do that. It needs a repeatable intake record that shows what the scan looks like, how it will be processed, and who will review the output.
A short project record can include:
| Intake field | What to capture |
|---|---|
| File type and scope | Whether each file is a scan, a text PDF, or a mix, and which pages are in scope |
| Image condition | Blur, glare, rotation, contrast, cropping, and text size |
| Content features | Handwriting, stamps, tables, columns, footnotes, names, and numerical fields |
| OCR approach | Tool or process, selected pages, and any configuration that affects output |
| Review plan | Who checks the output and which content receives direct comparison |
| Source preparation | Required file format, reading order, and layout repair |
| Pricing basis | Translation billing basis plus separate preparation or formatting charges |
| Privacy handling | Tool approval, access, transfer, retention, and deletion instructions as applicable |
The record helps explain the quote if the client asks why a short scan costs more than its word count suggests. It also helps the team revisit the estimate when a file differs from the sample page.
A sample page is useful when the complete document is too large to assess in detail before quoting. The agency can inspect a representative page, estimate the preparation approach, and state that the final scope depends on the remaining pages matching the sample. If the remaining pages contain handwriting, unusual tables, or much poorer image quality, the agency should pause and agree on a revised scope rather than absorb the difference silently.
Document the boundary between included correction and additional work. For example, a fixed preparation fee might include correction of OCR output and a clean source file, while substantial table reconstruction or unreadable passages require a separate estimate. The exact boundary depends on the agency’s service model. The important part is that the client sees it before the team spends unplanned time.
A margin review after delivery closes the loop. Compare planned and actual effort by task. Did OCR correction take longer because the scan was poor? Did the page count hide a large source-word volume? Did the project manager handle several rounds of clarification? Did the final document need more layout work than expected? The answers improve the next quote without relying on a guessed universal rate.
For a wider view of how production and review time affect agency pricing, see how to calculate true translation overhead. For projects where scanned-PDF handling fails inside a CAT tool, the reasons scanned PDFs can fail to import into Trados or memoQ help explain why a separate preparation step may be necessary.
The margin disappears when the agency treats preparation as invisible, not simply when it chooses OCR or manual retyping. A clear estimate names the work, assigns responsibility, and sets a review standard that matches the document. That gives the client a more honest price and gives the production team a usable definition of a finished source file.
FAQ¶
How do translation agencies calculate the cost of retyping a scanned document?¶
Agencies should estimate source volume, OCR processing, correction time, layout repair, translation, review, and project management separately. A per-word translation rate alone does not cover work that happens before the source text becomes usable.
How much time does OCR save compared with manually retyping a scan?¶
There is no universal time-saving figure in the cited evidence. OCR can turn image text into machine-readable text, but the time saved depends on image quality, language support, layout, and how much human correction the output needs.
What hidden costs should an agency include when quoting scanned-document translation?¶
Include file assessment, OCR, correction, source preparation, table and layout repair, translation review, project management, and relevant privacy handling. State which tasks the quoted fee covers so the client can see what changes if the scan is worse than expected.
How should an agency check OCR output before translation?¶
Compare the extracted text with the scan, paying close attention to names, figures, dates, tables, and uncertain sections. Use confidence scores and image-quality signals to guide review, but set the review level according to the document’s risk rather than treating OCR as error-free.
When should an agency retype a scan manually instead of using OCR?¶
Manual transcription can fit handwriting, degraded images, or passages OCR cannot reliably recognize. Agencies can also combine methods by using OCR for readable printed text and manually checking or transcribing difficult sections against the source.