What a machine decides at BNKA and what a person decides

"We use artificial intelligence" is now a marketing phrase with no content. This article attempts the opposite: to say exactly where models sit in our product, what they decide, what they do not decide, and what limits we set for ourselves.

We publish it because if we are going to ask you to trust a platform with your money, the least we owe you is the ability to know which part of the process is automated.

The three places where AI sits

1. Identity verification

When you open a euro account with a European IBAN, we have to check that the document is genuine and that the person in the selfie is the person on the document. Models do that by analysing the image: detecting tampering, comparing faces, verifying document security features.

Why it matters that this is automated: doing it manually would mean days of waiting. With models, most verifications resolve in minutes.

Why it sometimes asks you to retake the photo: the model assesses sharpness, glare and cropping. Most rejections are image problems, not document problems, and what the rules actually require is set out in opening a European IBAN account without an NIE.

2. Fraud and transaction monitoring

Every transaction passes through rules and models looking for anomalous patterns: a sharp change from your usual behaviour, typical signals of a compromised account, or the patterns that characterise networks recruiting other people's accounts to move third-party money.

Here we have to say something uncomfortable: a fraud system that never gets it wrong does not exist. The technical terms in this section are defined in the glossary. You have to choose between two kinds of error. Calibrate it never to get in the way, and fraud gets through. Calibrate it to let no fraud through, and it sometimes stops legitimate transactions.

We chose the second side. We prefer an annoying false positive, which a review resolves, over completed fraud, which does not get resolved. What happens when that control fires on your account is explained in why an account goes under review.

3. Exchange rate analysis

We use models to process market data and show context on how a currency has moved, such as the euro and Argentine peso pair. That is information, not a recommendation: we do not predict where a currency is going, and we would be sceptical of anyone claiming they can. None of it constitutes financial, legal, tax or investment advice.

Where there is NO automated decision

This is the part we most want in writing.

This is not only a design preference. The General Data Protection Regulation recognises people's right not to be subject to decisions based solely on automated processing where these produce significant effects. A blocked account is a significant effect.

The limits we set for ourselves

1. We do not train general-purpose models on customer personal data. The data you give to verify your identity is used for that.
2. We do not sell data to third parties.
3. We do not automate the final word on compliance. A model prioritises; a person decides.
4. We do not publish content about money without human review. This article included.
5. We do not use AI to simulate people. If you are talking to an automated assistant, it says so.

How we work internally

On the team side, AI is an ordinary working tool: writing and reviewing code, analysing logs to find the cause of an incident, drafting documentation, translating content, preparing analysis. It lets a small team operate a product that spans several jurisdictions, from the Colombia corridor to the euro area.

That has a cultural consequence we would rather say out loud: being a small team is not an excuse. If a control fails or a user waits too long, the problem is ours, not the team size's.

And it has a limit: AI speeds up the work, it does not transfer responsibility. When something goes wrong, a person with a name answers for it, not a model.

What is missing

So as not to sell a polished version of reality, three things we are short on:

• Support response times are not where we want them, although a good share of the questions are already answered in the FAQ. We are rebuilding that part and the current state is on the support page.
• There is not enough in-app explanation when a transaction is held. Today you have to ask, and to tell a normal delay apart from a hold it helps to read how long a transfer really takes. It should be visible.
• The Lithuanian translations of parts of the site need a native speaker review. We know, and it is in progress.

If you want to work on this, there are openings on the careers page.