Skip to main content

Document recognition (OCR)

When someone uploads a document against a credential — a photo of a card, a scanned certificate, a PDF — Oho can read it and tell you whether it is the document it is supposed to be. In the app this is Settings → Manage → OCR, "Optical Character Recognition".

Availability: Premium. It is also off by default and opt-in per credential type.

What it actually does​

This is the part worth being precise about, because the name invites a bigger assumption than the feature makes good on.

It does: identify what kind of document the file is, compare that against the credential type it was uploaded for, score its confidence, and explain its reasoning — including any concerns it noticed about the image.

It does not: fill in the credential for you. Oho does not read the card number, the holder's name, or the expiry date off the document and write them onto the record. Those fields come from the person entering them, from an HR feed, or from the register when the credential is verified.

So the question document recognition answers is "is this the right kind of document?", not "what does this document say?". It is a check on the evidence, not a data-entry shortcut.

Recognition is not verification

Recognising a file as a Working With Children card says nothing about whether that card is real, current, or belongs to this person. Only a check against the issuing register answers that — see One-off verification. Document recognition sits in front of verification, catching the wrong file before anyone relies on it.

What happens on upload​

  1. The file is uploaded against a credential whose type has OCR enabled.
  2. Oho sends it for analysis and gets back a detected document type, whether that matches the expected type, a confidence score, and any concerns.
  3. The result is stored with the attachment.
  4. If the result is low confidence or an outright mismatch, an item is raised in Review & Decide for a person to look at. A clean, confident match raises nothing.

The low-confidence threshold is 70%. Below that — or where the analysis says the document is simply not the expected type — the attachment goes to review. A stated mismatch is treated as more urgent than a borderline score, on the reasoning that "this is the wrong document" is a more actionable finding than "I'm not sure".

Resolving a flagged attachment​

A flagged attachment appears in the queue with the file name, when it was uploaded, the confidence percentage, what the analysis thought the document was, and its reasoning. A reviewer can confirm the document is right after all, ask for a replacement, or remove the file.

Asking for a replacement sends the person a re-upload link and closes the item honestly — as a request for a better document, rather than as a judgement on the one supplied. The replacement, when it arrives, is analysed in its turn and raises its own review if it needs one.

Turning it on​

Under Settings → Manage → OCR there are two controls, and both must be set:

  1. Attachment scanning — the master switch for the organisation.
  2. Credential types — move types into the OCR enabled list. Only types on that list are scanned. The catalogue covers every built-in type plus any custom credential types you've created.

Enabling it for everything is rarely the right call. The value is highest on credential types where the document is the evidence and a wrong file would go unnoticed; it is lowest on types that get verified against a register anyway, where the register is the real check.

Limits worth knowing​

LimitValue
File size5 MB per file
Image formatsJPEG, PNG, WebP, GIF, HEIC
Document formatsPDF
ScopeOnly credential types on the OCR-enabled list

Analysis is performed by an AI model running in Amazon Bedrock, in AWS's Sydney (ap-southeast-2) region — the same region as the rest of your Oho data. See Data retention & privacy for how attachments are stored and removed.

Treat the confidence score as a prompt, not a verdict

The score reflects how sure a model is that a file looks like a kind of document. A high score on a convincing forgery is entirely possible, and a low score on a poorly-lit photo of a genuine card is routine. Use it to decide what a person should look at — not as evidence in its own right.