AI in Compliance and Fraud Prevention

We build language models into compliance and anti-fraud products wherever classic rules and keyword dictionaries generate too much noise. The core principle is a hybrid: deterministic logic finds candidates and applies the rules, while the model weighs the context. The solution stays explainable, and analysts spend less time on false positives.

With AI, I start with the problems where it can deliver tangible value to a bank and its analysts. For our products I chose an approach where AI helps understand context while every decision stays explainable. Together with my team we bring it to life in Merchant Inspector and put it to the test on real banking tasks.

Challenge

Checking merchant websites against dictionaries of prohibited words catches not just violations but everything else too: a product mentioned in an article, in a review, or in a list of things the store explicitly does not sell. Analysts waste time sorting through false positives, and the bank risks turning away an honest merchant. Rules alone also can't realistically tell a merchant's real line of business from its website and compare it with the declared MCC.

Solution

In Merchant Inspector, a language model powers the key checks:

Contextual review of keyword matches — the keyword dictionary finds candidates in page text, text in images (OCR), social media and hidden pages, and the model reads the surrounding snippet to decide whether the merchant really sells that product.
MCC compliance — the model determines the actual line of business from the website content, while programmatic rules make the final determination using groups of permitted MCCs and a list of prohibited MCCs.
Required information — the model checks whether the website actually has contact details, shipping and return terms and a privacy policy, not just similar-looking words.
Merchant assessment — the model drafts an assessment for the analyst, while programmatic rules keep its conclusions grounded in the check results (pilot).

The main model is open-weight and runs on our own GPU server, not in the cloud. It is connected via an OpenAI-compatible API, so the model can be swapped without reworking the product.

Results

on a client bank's portfolio, the model dismissed all keyword matches in 231 of 295 checks that had returned matches — about 78%; how well this carries over to other portfolios is yet to be verified;
after context-aware checking was switched on, the share of checks flagged as violations dropped from 46% to 18% for shipping and return terms and from 29% to 21% for the privacy policy.

Next steps

We are extending this approach to transaction anti-fraud: an AI assistant for analysts that prepares an incident summary and explains which rules fired, and a helper for designing new rules and testing them on historical data.

AI in our development process

Inside the team, an AI analyst clarifies the task by asking the author questions, an AI developer prepares the code and the pull request, and AI takes part in code review. The decision to merge is made by a human.

← Back to projects