Passports, letters, ID photos: in public services, artificial intelligence can already read them for your staff. The real question is who makes the call. Here are the rules Lya & Co follows so that the machine saves time without ever deciding the outcome of a case.

In public administration, time is rarely lost at the moment of decision. It is lost beforehand: retyping passport details, sorting post, checking photos. That is where AI earns its keep, and where the temptation creeps in: if the machine can read, why not let it decide?

Our answer is always the same. In e-Visa and e-Archives, our platforms for institutions, AI reads and suggests; the outcome of a case remains a human decision.

Reading is not deciding

On e-Visa, automatic passport reading fills in the form, but it is the applicant who confirms every detail before moving on.

The machine can say no, but only to unusable files: a photo taken of a paper print, a very dark background, a passport uploaded where a plane ticket was expected. It then asks for the document to be submitted again, nothing more.

Some tasks are deliberately not automated: the consular officer compares the face in the photo with the one in the passport, by eye. They may also ask for a new photo even when it has passed the automatic checks.

As for the decision on the visa, it rests with the consul: sending the reply can be automated, the decision never can.

The same logic applies to e-Archives: the machine suggests the shelf where the paper original should be kept, and the archivist confirms it. The assistant that answers questions about the archive cites its sources, so anyone can check.

When AI is unsure, a person decides

Useful AI knows how to say it is unsure. That doubt still has to reach the right person.

• Flagged fields. When the reading of a passport field is uncertain, the applicant sees “Please verify this value” rather than a value shown as certain.
• A second, targeted reader. The holographic security film sometimes hides a printed field, such as the place of birth. Rather than guess, the reading leaves it blank, then a second engine is called on for that field alone, without overwriting the rest. The applicant then reviews everything.
• Two concurring opinions. A photo is only rejected for a background that is not plain if the AI’s assessment and an objective pixel measurement agree. Otherwise, it goes through with a warning.
• A note for the officer. If applicants carry on despite the warning about their photo’s background or framing, their application is not blocked: the officer sees this in the file and makes the call.

Another finding, on the archives: the AI rated its own confidence as high almost every time, whether it was right or wrong. The confidence a machine gives itself therefore proves nothing. We rely instead on an independent signal: does its suggestion match the shelf of similar documents already filed? When the two agree, the archivist confirms it quickly; otherwise, the item gets a closer look.

Measure before you deploy

Adjustments to automatic reading are measured on a test bench: several dozen real documents of around twenty types (passports, tickets, certificates, vaccination records and so on), whose correct answers were established by hand. We compare what the machine reads with that reference, field by field.

The speed improvements now in service were measured away from the live service: no wrongful rejections, no field getting worse. Most are switched on with a simple setting that can be reversed instantly.

These measurements sometimes defy intuition:

• Newer does not mean better. Newer versions of reading engines were set aside: one read handwritten documents less well, another cut off compound first names. Handwriting, such as vaccination records filled in by hand, remains our known weak spot: human review is essential there.
• Nor does faster. An engine twice as fast found far fewer fields on identity documents and mixed up surnames and first names. It was withdrawn the same day, before a single applicant used it.

Measurement also serves fairness. On the photos actually submitted, a strict rule on background colour would have rejected more than a third of them. Yet blue or grey studio backgrounds are perfectly legitimate: as far as colour goes, only a very dark background is refused.

Because it has not been calibrated in real conditions, the detection of a non-neutral expression only triggers a warning, so that no one is unfairly penalised. The sharpness threshold, for its part, was calibrated on real photos: it stops clearly blurred images without rejecting acceptable ones.

Never embellish, always explain, keep a record

An ID photo is never retouched: no background replacement, no lighting correction, no facial enhancement. For photos submitted online, only framing and orientation are adjusted. This is in fact an official requirement in the United Kingdom: the digital photo for a passport application must be “unaltered by computer software”.

Every refusal or warning is explained in plain language, with the next step. Take this warning: “Your photo’s background is not plain, which may lead to your application being refused. Please retake it against a plain, light-coloured wall.”

As for the record, significant refusals are checked and recorded by our system, not merely shown on the applicant’s screen.

AI in public services: four questions before you start

AI can relieve your staff of reading, sorting and retyping without taking away their judgement. Before any project, ask yourself four questions. Who decides? What happens when the machine is unsure? How has its reliability been measured? What does the applicant see?

The principles described here echo published frameworks. In July 2024, the African Union endorsed a Continental Artificial Intelligence Strategy that promotes an Africa-centric approach and ethical, responsible and equitable practices. In Europe, the AI Act classifies systems that help competent authorities examine visa applications as high-risk (Annex III, point 7(c)). It also provides for effective human oversight (Article 14). For us, these texts are reference points, not a badge.

Are you considering automating how you receive applications or manage your archives? Explore our products for institutions or get in touch to talk it through.