Skip to content
Interface language: English
Audit / Signal

OCR before AI: how to check that the model reads the right data

Check OCR before AI analysis: numbers, negations, tables, missing pages and the connection to the original document.

Article language: English

An illustration of a gold-coloured lantern lighting stone steps and cracks in a wall.
In this article

Imagine a form that says “not completed”. The short word “not” disappears during recognition, and the analysis proceeds in the opposite direction. If AI analyses a scan, first check the recognised text and document layout. Numbers, negations and the assignment of headings to columns need comparison with the original. Identify important fields and the conditions for passing the material on. Otherwise, a model may efficiently produce an answer to something the document never said.

OCR stands for optical character recognition: reading text from an image. In a business process, treat it as a separate stage with its own acceptance checks. A readable paragraph generated later by a model does not prove the earlier recognition was correct. AI can explain incorrectly read material very fluently; fluency is not a scanner's proofreader.

Identify information whose confusion changes the task

Start with the purpose of the analysis. In a hypothetical service company, AI is to prepare a draft service request from a submitted form. The equipment identifier, symptom description and whether the work has already been done will matter. An error in a decorative heading has a different consequence from a changed number or a missing negation in the description.

Make a list of fields critical to this process. It need not cover every character in the document. It should distinguish information that must be read exactly, information that can be paraphrased and gaps that must not be filled by guesswork. That distinction helps you choose a check instead of applying the same effort to everything.

Before sending the material, also establish the permitted scope. Data minimisation before using AI applies to the OCR tool too when it runs as a separate service. Removing unnecessary information only from the model's answer does not change what was sent for recognition earlier. The Polish article on data minimisation before AI (in Polish) develops this point.

In the materials for an AI audit, retain a description of the document type, the version of the reading method and an example approved for the test. The result should be traceable to the correct original. Without that connection it is hard to tell whether OCR, later analysis or the person selecting the file made the mistake.

Check completeness before correcting typos

Check page count, order and visible cropping. A missing final page of conditions may matter more than a few wrong characters. Pay attention to documents combined from several files, double-sided scans and attachments. A tool may correctly read everything it received while the process still relies on incomplete material.

For a phone scan, inspect corners, glare and areas near folds. Do not assume that a document generally readable to a person will be read equally well automatically. A weaker image does not necessarily disqualify every piece of information either. Base the decision on the importance of the particular passage and whether it can be checked.

Google describes image-quality indicators in Enterprise Document OCR, including blur, small text and glare. They can help route material for further handling, within the documented limits. They do not replace comparing an important field with the document and are not a universal certificate that the whole reading is correct.

If a page is missing or an important passage is cropped, request the right material or stop the affected part of the analysis. Do not ask the model to reconstruct a missing condition from what usually appears in similar forms. Missing information should remain visible as missing.

Compare recognised text and meaning separately

The first check concerns characters: were the right name, number and negation recorded? The second concerns relationships: what does a value belong to, which section does a condition apply to, and was a description joined to the neighbouring column? A correct sequence of words can still be a faulty basis for analysis if its placement has been lost.

In a hypothetical equipment table, the value “105” may belong to a different row from the name read next to it. Checking the characters alone will not reveal that. Prepare a case with similar names and different values, then ask for the source of each answer. Assess both the text and its position in the structure.

Pay particular attention to headings repeated on later pages, footnotes and checkboxes. Do not describe a blank field as a negative answer without an established rule. No selection, an illegible mark and a clearly selected answer are different states. The downstream process should preserve that distinction.

If text enters a company AI search tool, an error can return in response to many later questions. Check a representative result before adding it to the source collection. Correcting successive answers does not remove faulty material the system still uses. See also the Polish guide to checking a company AI search tool (in Polish).

Use confidence scores to support decisions

Some tools return a confidence score for a recognised word, field or passage. Microsoft explains these scores in Document Intelligence, distinguishing levels of the result and noting the option of human review. Check the availability and meaning of a score for the model and service version you use.

Do not copy one threshold across all documents. A number beside a field does not by itself determine the consequence of an error in your process. An unimportant description may follow a different path from an identifier that triggers further action. Choose the rule from your own examples and record its scope.

Test with material whose reading has been confirmed. Examine cases the tool sends for review and a sample of those it passes automatically. If you assess only rejected results, you cannot tell whether errors also pass through the other path. A score should help locate a problem, not give it a convenient place outside the report.

Establish what the reviewer sees. They need the field, the corresponding passage of the original and a way to record a correction or uncertainty. A whole document in a separate window can be awkward if the reviewer must find the right page every time. A well-designed comparison reduces the cost of checking without giving up its purpose.

Build a test from errors that matter

Prepare a small set of different materials: a good scan, a photograph with glare, a table, a long name and a document with a missing page. Use fictional data or data approved for this test. Record the expected result, including cases that should end in a request for new material rather than a finished analysis.

Aeneia conceptual photography

Add a test of a number with a separator, similar-looking characters and a negation. For example, when the document uses a comma as its decimal separator, “10,5” and “105” should stay distinct, and “not completed” must not become “completed”. These are illustrative checks, not a claim that every engine makes those particular errors. The aim is to test the information important to your process.

Then pass the result through the later AI stage. Assess whether the model preserves gaps and can identify its basis. A correct reading in a separate panel is not enough if the application sends a different file or only part of the result for analysis. A data-flow register helps cover the material's entire journey. The Polish guide to mapping an AI data flow (in Polish) describes this check in more detail.

After a tool or setting changes, return to the same important cases. AI process regression tests help check whether improving one format worsened another. Acceptance should also cover the result format and its connection to the original. For related checks, see the Polish article on AI regression tests (in Polish).

Agree on corrections, responsibility and the next step

A human correction should be recognisable as a correction and point back to the source. Do not silently overwrite the original reading if you may later need to explain the cause of an error. Choose how many versions to retain for the task; an unlimited archive of every test is not necessary.

Assign responsibility for documents requiring a new scan and for cases that cannot be read unambiguously. A “low quality” message alone does not finish the job. The team needs a way to return the material, mark it as waiting and prevent further automated analysis until the issue is resolved.

If recognition leads to writing or sending, connect it to approval of AI actions. The decision-maker should see the important data and its uncertainty. Approving a polished summary is not the same as checking the number on which an operation depends. Further guidance appears in the Polish article on human approval of AI actions (in Polish).

Include rescanning, review and corrections in the cost of the AI process. Cover the image, recognised text and diagnostic copies in the retention plan. They are different materials even when they concern one document.

Distinguish a digital file, a scan and a mixed document

A PDF extension does not say where its content came from. One file may contain text from an editor, a photograph of a signed page and a previously recognised text layer. Establish which of these your process uses. There is no need to read the image again if the right text is already available, but you still need to compare it with the visible material.

In a test, copy a selected passage and compare it with what appears on the page. Then trace the passage actually sent for analysis. If the application uses a different layer from the one a person checks, they may be looking at two different materials under the same file name. Record that difference as a specific acceptance condition.

For a mixed document, choose an example with several page types. Check whether the result preserves order and information about origin. Correctly reading the first page does not establish what happens with the photographed portion. Mark pages needing a separate path instead of passing the whole file under one general quality status.

See what happens after recognised fields are exported

A reading may be correct in the panel and change when transferred to a spreadsheet or ticketing system. An identifier with a leading zero, a decimal separator and a date need appropriate handling. This is no longer the same error as character recognition, but the result can look similar to the recipient: a different value appears in later work.

Prepare fictional fields with those characteristics and follow the whole export. Compare the original, the OCR result and the value at the destination. Record the transition at which the meaning changes. This lets the correction go to the right stage instead of endlessly adjusting recognition settings. When the next stage is a model's response in a data structure, also check the meaning of fields in AI-produced JSON, while keeping the comparison with the original document.

Do not standardise notation automatically without a rule. Sometimes a space is only presentation; elsewhere it separates parts of an identifier. A date likewise needs the document's context. The process owner should confirm which transformation is allowed and how the earlier value can remain checkable.

Reassess checks when document types change

A new form, another language or a changed table layout should prompt an assessment of whether the existing test still fits. Not every visual adjustment needs to be treated as a new deployment, but a change in how important fields are presented can undermine previous assumptions. Record review triggers in the maintenance instructions. This is an example of the broader problem of changes in material reaching an operating AI solution, which should be distinguished from a change to the model or configuration.

When the team starts receiving photographs instead of scans, collect representative examples before expanding the automated path. Check the same important information and how uncertain cases are routed to a person. Results for the older format remain useful, but do not automatically become evidence for the new one.

The responsible person should receive a short decision: which type of material was checked, which fields were covered and what remains unknown. An operator should not have to guess whether an unusual document fits the accepted scope. A clear boundary helps use the solution safely without pretending it recognises every possible page.

Accept the source before judging the elegance of the answer

Finally, choose a few questions whose answers can be confirmed directly in the document. Ask someone outside the configuration work to follow the result back to the right passage in the original. Record mistakes, missing pages and unresolved cases. Limitations should inform the decision to use the solution, not disappear into a technical appendix.

Add instructions for the person receiving unusual material. They should be able to tell whether the document fits the checked scope, where to see uncertain fields and whom to tell about a gap. Then ask them to take one case through without help from the person who configured the system. If they must guess the recognition state or search for a private explanatory message, improve the handover. This small test joins technical quality to everyday work and checks whether the procedure functions away from its creator's desk.

If you want to examine this stage in your process, see the Aeneia Audit offer.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear