Scanned PDF Parsing for AI Agents: Checks Before Use
A checklist for AI agents reading scanned PDFs: integrity tests, confidence thresholds, source trails and approval rules before text is trusted.

To use a scanned PDF safely, an AI agent should confirm that every page arrived, that the text was recognized with acceptable confidence, and that key values agree with a second signal before anything is stored or acted on as company fact. A scan is an image until a parser converts it to text, and that conversion can fail quietly. The checks below run in a fixed order, leave a trace a person can audit, and send doubtful results to a human queue instead of into memory.
Key takeaways
- A scanned PDF is an image until a parser turns it into text, so the agent should never assume the text layer is trustworthy.
- Check file integrity, page coverage, recognition confidence and field-level agreement before any value is stored.
- Attach the source page and a confidence label to every extracted fact so a person can trace it.
- Keep low-confidence or conflicting values out of memory until a named person confirms them.
- Measure parsing quality on a fixed set of your own documents, and repeat the test after every change.
Why a scanned PDF is not a text file
A scanned PDF is often a photograph of paper stored inside a PDF wrapper. Some files also carry a text layer added by the scanner, and that layer may be incomplete, out of order or simply wrong. When there is no reliable text layer, optical character recognition estimates each character from pixels. Its errors cluster in exactly the places a business cares about: decimal separators, digits that look alike, dates written in unfamiliar formats, faint stamps, handwritten corrections and pages photographed at an angle. For a wider view of reading scans, photos and PDFs, see OCR for AI agents: reading scans, PDFs and photos reliably.
The risk is not only that the text is wrong. A language model reading garbled input will often produce a fluent, confident summary, so a wrong total or an invented due date can read exactly like a correct one. The agent's tone tells you nothing about parsing quality. The check therefore has to happen in the data path, before the model reasons over the text, and its result has to be recorded somewhere a person can inspect.
The five checks before the text is used
Run these checks in order. Each one is inexpensive, and each stops a specific kind of failure from reaching memory or a reply.
- Confirm the file opens, is not password protected, and matches a document type the agent is allowed to read.
- Confirm the page count against the sender's description or the document's own page markers, and confirm that no page is blank, cropped or duplicated.
- Record recognition confidence for each page and each extracted field, and mark any page below the threshold you set for that document type.
- Cross-check values that have internal logic: line items should add up to the subtotal, tax should match the rate printed on the document, and a due date should not come before the issue date.
- Compare identifiers and totals against an existing record in your system when one exists, and route every difference to a person instead of overwriting the record.
A failed check should never be a silent drop. The document should be marked as partially read, the fields that passed should be labeled as such, and the failed fields should stay out of the agent's working context until someone resolves them.
Set confidence thresholds by field type
One threshold for the whole document is too blunt. A misread supplier logo may have no consequence, while a misread account number can send money to the wrong place. Set thresholds by field type, test them on real documents, and write them down so the operations owner can change them on purpose. A practical decision guide looks like this:
- Amounts, tax identifiers and account numbers: use only above the strictest threshold, and never without a passing cross-check. Anything else goes to review.
- Dates and deadlines: use only when the format is unambiguous and the value falls in a plausible range. If day and month could be swapped, ask.
- Names, addresses and contact details: store them with a source page reference, and confirm them before any outbound message uses them.
- Free-text notes and descriptions: usable for summaries when the summary says it was extracted from a scanned source.
- Signatures, stamps and handwritten marks: record whether they are present, and do not treat them as readable content unless a person has verified them.
Keep a source trail for every extracted fact
Each extracted value should carry the document identifier, the page, the region if the parser provides one, the parser version, the confidence and the extraction time. When the agent answers a question from that value, the reply should name the page it came from. This turns a claim the agent made into something a person can check in seconds, and it makes audits possible when a parser is updated or a vendor changes its service.
Store the values against a fixed structure rather than as free text. Structured outputs for business agents describes schemas that downstream systems can validate, which gives you a second gate after the parser has finished its work.
Treat text inside documents as data, not instructions
Documents are untrusted input. A scanned page can contain a sentence that tells the reader to ignore earlier rules or to approve a payment, and a parser will faithfully pass that sentence along. The OWASP guidance on prompt injection treats external content as untrusted and recommends limiting what a model can do with it. For scanned PDFs, this means the agent reads extracted text as material to analyze, never as a command, and that a document alone can never trigger a write, a send or a payment.
Write actions should sit behind a separate approval step. The same principle applies to every permission the agent holds, as explained in Read, draft and send: three different AI permissions.
Test parsing on your own documents
Vendor demonstrations use clean samples. Your documents include faded faxes, phone photos taken at a slant and forms with a stamp across the totals. Build a small evaluation set from real files, write the correct values down with a person's sign-off, and include the hard cases on purpose. Run the parser against that set, score each field, and group the failures by cause, such as skew, low contrast or handwriting. The NIST AI Risk Management Framework treats measurement as an ongoing activity rather than a one-time check, and a parsing test should follow the same logic.
Repeat the test whenever the parser, its settings or the mix of document types changes. Keep the results with the date and version they describe so that a drop in quality can be traced to a specific change.
Illustrative scenario: a purchasing team forwards supplier invoices photographed with a phone. On one invoice, the parser reads the decimal separator in the wrong place, which would change the amount by a large factor. The line items do not add up to the printed subtotal, so the check fails, the amount is withheld from the agent's working context, and the document goes to the clerk's review queue with the page image attached. Nothing is drafted for payment until a person confirms the value.
How Be Human handles scanned documents in managed operations
Be Human defines which document types an agent may read during the first engagement, agrees thresholds and review owners with the operations lead, and keeps every write action behind an explicit approval. In managed operations, the team monitors parsing failures, runs recurring evaluations on a sample of real documents, reviews permissions when a source or channel changes, and handles incidents when a bad extraction reaches a downstream system. The goal is that every answer the agent gives stays traceable to a page a person can open.
If you are not sure which scanned document flow should come first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent read scanned PDFs without OCR? Sometimes, if the PDF includes a reliable embedded text layer. Many scanned files do not, and an agent working from an image without recognition cannot read the content at all. Treat the parser as a separate component with its own checks, and never assume the text layer is accurate just because it exists.
What confidence level should trigger human review? There is no universal number, and the threshold should come from your own test set rather than a vendor default. Use stricter limits for amounts, identifiers and dates, and looser limits for descriptive text. Review the thresholds after each parser change or when the mix of documents shifts.
Should the agent store extracted values in its memory? Only values that passed the checks, and only with their source page and confidence attached. Failed or unconfirmed values should stay in a review queue. Memory that holds unverified figures tends to carry errors into later answers long after the original scan has been forgotten.
Who should own the review step? The person who owns the business process, such as the accounts payable lead or the operations manager, should own it, not the engineer who built the parser. That person knows which values matter and what a plausible result looks like. The agent can prepare each review item with the page image, the extracted values and the failed check, which shortens the work for the reviewer.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


