Skip to content
AI8 min

Reading documents with AI: from the ID and the bank statement to the customer file, without typing by hand

By · co-founder and Chief Executive Officer of Hábil ·

How to automate document capture with AI to read IDs, proofs of address and bank statements: confidence per field, review and control.

In 1 minute

Today an AI can read an ID, a proof of address or a bank statement and put its data into the system without anyone typing it. That is the easy part. What separates reliable capture from capture that creates problems is what happens after reading: comparing each piece of data against another source, flagging what was not read well and sending it to a person, instead of filling the gap with something that looks right. This article explains how to require that of every AI capture project you evaluate.

Imagine a company that opens hundreds of new customer files every week. It is fictional, but it looks like many. Each file has an ID, a proof of address and a bank statement, and in this example they arrive as photos taken with a phone: some crooked, with glare or cut off. A team of four people opens them one by one and copies the data into the system.

Leadership sees a demo: the AI reads a sample ID and fills in the visible data: name, address and date of birth. Someone asks the natural question: "so we no longer need data entry?".

The short answer is that typing stops being the job, but the job does not disappear: it moves. This article is for whoever is going to approve a project like this. It makes one point: reading is the easy part; what takes work is comparing, flagging what is doubtful and not filling in gaps silently, with a person approving whatever moves money.

Why reading is not the same as capturing well

A person entering data from a document does three things at once, without thinking about it: reads, notices when something is unclear and asks before making it up. An AI does the first very well. The other two have to be designed.

There are two ways to get it wrong. In character recognition, the simpler one: a misread digit or a missing piece of data. And when the flow uses generative AI to interpret or complete the document, a more delicate one: the generative AI profile from NIST, the United States standards institute, calls confabulation the production of false or erroneous content stated with confidence, known colloquially as "hallucinations" [1]. In a document, it looks like this: the last digit of an account number came out blurry in the photo, and instead of leaving it blank, the system puts one in. The file looks complete. The error shows up weeks later, when a deposit bounces. In both cases the rule is the same: do not complete silently, and send the exception to review.

That is why the useful question in front of a demo is not "does it read it well?" but "what does it do when it does not read it well?".

What to ask of an AI capture project

1. That each piece of data comes with how sure it is

Several document-reading services from the large cloud providers return a confidence score for certain predictions, such as a word or an extracted field: how likely it is that it was read correctly. Microsoft's documentation says that, where accuracy is critical, that confidence can be used to decide whether to accept the data automatically or flag it for human review, and warns that not all fields return one [2]. Amazon's recommends setting a minimum threshold and, below it, discarding the result or flagging it for closer human scrutiny [3].

Before designing the flow, it pays to confirm which scores the provider returns for your documents. With that as a base, the project has to be able to say, field by field, how much confidence it requires before accepting a piece of data on its own, and it does not have to be the same for everything: the account number calls for more than the neighborhood. One more thing: that score measures the reading, not whether the document is authentic, current or consistent with another source. That is checked separately.

2. That it compares before accepting

Reading well does not prove that the data is true or that the document belongs to whoever presents it. A well-read ID may be expired; a well-read address may not match the one on the proof of address. NIST's identity guidelines separate distinct controls: confirming that the document is genuine and authentic, confirming that the data it carries is correct and, when the process requires it, confirming that the person presenting it is its genuine owner [4]. Each control answers a different question.

In a customer file, cross-checking means asking questions like these:

  • Is the name on the ID the same as on the bank statement?
  • Does the address on the proof of address match the one the customer gave?
  • Is the date of the proof of address within the period your policy requires?
  • Does the ID data match what the official source your process consults says?

Each answer that does not match is an exception, not an error someone quietly fixes.

3. That a gap stays a gap

If a piece of data was not read, the system has to leave it empty and say why: "illegible", "cut off", "not in the document". Filling it with the most likely value saves a minute today and costs a clarification tomorrow. This rule is worth writing into the project contract, not leaving to anyone's judgment.

4. That a person sees only what matters

The goal is not that nobody looks at the files. It is that the team stops copying and focuses on what does need judgment: what did not match, what was read with low confidence and what was left empty. If an AI reads most of the documents well, that team no longer spends the day typing, but resolving the doubtful cases.

And whatever moves money —opening a credit line, releasing a payment, changing a deposit account— is approved by a person, even if all the data matched.

5. That a review record remains

For each piece of data in the file it should be possible to see which document it came from, with what confidence it was read, what it was compared against and who approved it if it went through review. Without that trail, when a piece of data turns out wrong there is no way to know whether the photo, the reading, the rule or the person failed.

Personal data is not a detail

An ID and a bank statement contain personal data, and the bank statement, financial data as well. Mexico's Federal Law on the Protection of Personal Data Held by Private Parties, published in the Official Gazette on 20 March 2025, requires the data controller to observe, among others, the principles of purpose and proportionality, and states that financial or asset data require the express consent of their owner, except in the cases the law itself provides [5]. Which notice, which consent and which controls apply to your process is confirmed by your legal team before going live; this article does not replace that.

For a capture project, that leads to questions your legal team, your security team and operations have to be able to answer:

  • What is each extracted piece of data used for, and is everything that is extracted needed?
  • Where are the documents processed, and who can see them?
  • How long are the images kept after capture?

Among the design controls worth evaluating is masking personal data in technical logs and on the screens of anyone who does not need it. An example, with fake data: on the review screen, the team sees "Ana Ejemplo Prueba" and "CURP (population registry key): [fictitious data, masked]"; the full key is seen only by whoever needs it to approve.

Before opening it to everyone

Test the capture with a sample of recent files. Review the highest-risk fields separately —identity, account and address— and count three things: how many files went through untouched, how many went to review and why, and how many pieces of data accepted without review had to be corrected later. That last number does not come out in a demo: it is only known by reviewing by hand part of what the system accepted on its own. The full method for setting the baseline, the sample and the criterion for scaling is in The AI pilot went well in the demo.

Your first diagnosis of customer files

Take fifty recent files and mark in each one which piece of data was missing or did not match, which field needed review and how long the person took to resolve it. With that you have the baseline to measure any proposal and the list of exception rules worth starting with.

And before signing, ask the vendor two questions: what the system does when it cannot read a piece of data —leave it empty or fill it in— and what it compares each piece of data against before accepting it.

What Hábil does

At Hábil we custom-build the AI that works with the data and the rules of your systems.

If your team spends hours copying into your systems the data from IDs, proofs of address, bank statements and invoices, the proposal is in Hábil AI: automated data capture, where the AI extracts the data and your team handles only the exceptions. If the onboarding is for customers who open their account from their phone, we verify that they are a real person in front of the camera, compare their face with their ID and validate their data against the INE (national electoral institute) and RENAPO (national population registry); doubtful cases go to a person on your team. Whatever moves money or data is carried out with a person approving.

The conversation can start with your count of the fifty files, even if it is half done. Together we review where the time goes and where it makes sense to start.

Write to us on WhatsApp; to protect your information, do not share documents or customer data through that channel. In the conversation we will agree on the channel for the case.

Request your first conversation, at no cost.

Consulted on 18 November 2026. Each reference carries its confidence: high (primary source read).

References

  1. NIST, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile" (NIST AI 600-1, July 2024), risk 2.2: "Confabulation: The production of confidently stated but erroneous or false content (known colloquially as "hallucinations" or "fabrications") by which users may be misled or deceived". It is for voluntary use. Opened live on 18-Nov-2026. Confidence: high. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
  2. Microsoft Learn, Azure AI Document Intelligence, "Interpret and improve model accuracy and confidence scores": "A confidence score indicates probability by measuring the degree of statistical certainty that the extracted result is detected correctly", "For scenarios where accuracy is critical, confidence can be used to determine whether to automatically accept the prediction or flag it for human review" and "Currently, not all document fields return a confidence score". Vendor documentation, cited as practice. Opened live on 18-Nov-2026. Confidence: high. https://learn.microsoft.com/en-us/azure/ai-services/document-intelligence/concept/accuracy-confidence?view=doc-intel-4.0.0
  3. Amazon Web Services, Amazon Textract Developer Guide, "Best Practices": "In applications that are sensitive to detection errors (false positives), enforce a minimum confidence score threshold. The application should discard results below that threshold or flag situations as requiring a higher level of human scrutiny". Vendor documentation, cited as practice. Opened live on 18-Nov-2026. Confidence: high. https://docs.aws.amazon.com/textract/latest/dg/textract-best-practices.html
  4. NIST, "Digital Identity Guidelines: Identity Proofing and Enrollment" (SP 800-63A, revision 4), among the steps of identity proofing: "Evidence validation: Confirm that supplied identity evidence is genuine, authentic, and accurate", "Attribute validation: Confirm the accuracy of the core attributes" and "Identity verification: Confirm that the applicant is the genuine owner of the presented evidence and attributes". Opened live on 18-Nov-2026. Confidence: high. https://pages.nist.gov/800-63-4/sp800-63a.html
  5. Chamber of Deputies (Mexico), "Ley Federal de Protección de Datos Personales en Posesión de los Particulares", "Nueva Ley publicada en el Diario Oficial de la Federación el 20 de marzo de 2025"; latest amendment published in the Official Gazette on 14-11-2025. Article 5: "El responsable deberá observar los principios de licitud, finalidad, lealtad, consentimiento, calidad, proporcionalidad, información y responsabilidad en el tratamiento de datos personales"; article 7: "Los datos financieros o patrimoniales requerirán el consentimiento expreso de la persona titular, salvo las excepciones a que se refieren los artículos 9 y 36 de la presente Ley" (financial or asset data require the express consent of the data subject, except in the cases of articles 9 and 36). Opened live on 18-Nov-2026. Confidence: high. https://www.diputados.gob.mx/LeyesBiblio/pdf/LFPDPPP.pdf