AI Document Classification: How AI Reads Passports, Licences and More

A candidate replies to your onboarding request with four photos in one message. One is a passport. One is the back of a driving licence, taken at an angle. One looks like a bank statement but might be a utility bill. The last is a selfie that was probably sent by mistake. A consultant now has to open each one, work out what it is, decide whether it is good enough, and file it to the right place.

That small act of looking and naming, multiplied across hundreds of candidates, is one of the hidden costs of agency compliance. AI document classification is the technology that does it automatically: reading an attachment, deciding what kind of document it is, and judging whether it is usable. It sounds like a black box, but the principles are straightforward once you see them laid out.

This is part 2 of our AI Document Collection 101 series. Here we go behind the scenes: how AI reads an attachment, how it names the document type, what confidence scores mean, and where a human still needs to step in. By the end you will know what good automatic document recognition looks like and which questions to ask any vendor.

What AI document classification actually does

Document classification answers one question: what is this file? For a recruitment agency, the answer might be "passport", "driving licence", "DBS certificate", "proof of address", "training certificate" or "not a document at all".

It sits between receiving a file and acting on it. Without classification, every attachment is just "image_0423.jpg" until someone opens it. With classification, the system can tell whether the candidate has sent what was asked for, what is still missing, and where the file should go.

Classification is not the same as extraction

Two related ideas often get blurred together:

• Classification decides the type of document.
• Extraction pulls specific fields from it, such as a name, a date of birth or an expiry date.

Many document classification software products do both, but classification comes first. There is no point extracting an expiry date from a document if you have not confirmed it is the right document.

How AI reads an attachment, step by step

Most modern systems follow a similar pipeline, although the details vary by vendor.

1. Safety and format checks

Before anything is read, the file should be checked. Is it an image, a PDF or something else? Is it empty, password-protected or potentially harmful? Programs and installers should never be opened as if they were documents. This step protects your systems and filters out files that could never be valid evidence.

2. Quality assessment

Next, the system asks whether the content is readable at all. Common problems include:

• Blur from a shaky phone camera
• Glare on laminated cards
• Cropping that cuts off the edge of a page or the bottom of an ID
• Low light that makes text unreadable

A document that fails here should be sent back to the candidate straight away, while they are still engaged, rather than being discovered by a person days later.

3. Recognition

Now the AI looks at the content. Modern models consider both the visual layout (photo position, security patterns, the shape of a card) and the text on the page (headings, issuing body, field labels). A passport data page, a photocard driving licence and a bank statement look very different, and models trained on many examples learn those differences.

4. Naming the document type

The system then assigns a label, such as "passport" or "proof of address", and compares it against what was requested. If the candidate was asked for a DBS certificate and sent a training certificate instead, that mismatch is the most valuable thing classification can catch.

Confidence scores and when a human steps in

Classification is rarely a simple yes or no. Most AI systems produce a confidence score: an estimate of how sure the model is about its answer.

What confidence means in practice

A high-confidence result means the document closely matches patterns the model knows well, for example a clear photo of a standard passport data page. A lower-confidence result might mean an unusual document format, a partly obscured image, or a document that could plausibly be two different things, such as a council tax letter versus a utility bill.

A well-designed system uses confidence to decide what to do next:

Where people stay in the loop

Candidate document verification has two parts, and only one of them is about recognition. AI is very good at answering "is this a clear photo of a passport?" It is not a substitute for the checks your compliance team or the relevant authority must carry out, such as confirming a person's right to work, checking a professional register, or judging whether a reference is satisfactory. Those decisions need a trained person, an official service, or both.

The right mental model is that AI handles the sorting and quality control, so people spend their time on decisions rather than on opening attachments.

How GettingDocs uses classification in document collection

Classification on its own is only useful if it sits inside a process that gets documents from candidates in the first place. That is how Amy, the GettingDocs AI assistant, uses it.

• Amy chases candidates on WhatsApp and on email sent from your agency's own domain, written in your voice.
• When a file arrives, Amy checks it is readable and the right document. Blurred photos, cut-off pages and wrong documents are rejected, and the candidate is chased again.
• Every file is screened for viruses and malware. Programs, installers, password-protected files and empty files are refused.
• Accepted documents are filed onto the candidate's record in your ATS, matched by record ID, email or phone, never by name. Rejected documents are never filed to the ATS or Drive.

Live ATS integrations today are Recruit CRM, Crelate, Zoho Recruit and JobAdder, with Greenhouse coming soon. You can see the full feature overview on the GettingDocs homepage and read how files are protected on our security page.

Questions to ask any vendor

If you are comparing tools that promise automatic document recognition, these questions will tell you a lot:

• What happens when a candidate sends the wrong document? The answer should be "the candidate is asked again", not "a person finds out later".
• How are unreadable images handled? Look for blur and cropping checks, not just text recognition.
• Are files scanned for malware before processing?
• Where do rejected files go? They should not end up in your ATS alongside valid evidence.
• How are files matched to candidates? Matching by name alone is risky when two candidates share a name.

Key takeaways

• AI document classification decides what a file is, so the system knows whether the candidate sent what was requested.
• A good pipeline checks safety and quality before recognising and naming the document type.
• Confidence scores help decide whether to accept, re-request or flag a document for review.
• AI handles sorting and quality control; your compliance team keeps the verification decisions.
• Classification delivers the most value inside a process that also chases candidates and files accepted documents automatically.

If you would like to see classification working on your own candidate documents, book a call and we will show you how Amy chases, checks and files them.

---

Please note: this article is general information, not legal advice, and was correct at the time of writing (October 2026). Rules and guidance change, so please check the latest official guidance in the markets you recruit in before relying on it.