Led by our senior engineers and working alongside BCB's own compliance team, we built it in two parts.

Firstly, we developed an AI document-classification service into BCB's onboarding journey, using Azure AI Foundry. As a customer uploads documents, it identifies each one and flags immediately if the wrong document has been submitted - catching errors at the point of upload rather than days later. 

Not every document needed the same approach: reliable recognition already exists for common identity documents like passports and driving licences, so we reused it rather than rebuilt it, and concentrated our effort where there was a genuine gap: complex corporate documents such as certificates of incorporation, which no off-the-shelf tool could read - for this we designed custom classifiers. 

The harder problem was trust. Using Claude, we generated a first set of classification prompts from BCB's own sample documents. Nevertheless, a prompt that fits the samples won't necessarily hold up against the messy variety of real submissions. So we built a feedback loop: as the system classifies a document, a reviewer confirms or corrects it, and that judgement feeds back to sharpen the prompt, so it improves on the documents BCB actually sees. We tracked reliability as a score per document type rather than a single headline accuracy figure, which, in a compliance setting, is exactly what shows where the AI can be trusted, where it still needs a human's eye, and which classifiers to improve next.

Secondly, we developed an AI-assisted application review process. Working alongside BCB's compliance experts, we turned how they actually assess an application - applicant suitability, company registration, sanctioned-country checks, ownership structure - into a set of structured prompts, each returning a confidence rating and the reasoning behind it. 

Surrounding all of this was a robust Evals framework specially designed around validating the document classification and application review outcomes. Feeding in BCB's sample documents and comparing them against expected 'known good' outcomes.

The system is integrated with the wider flexible KYC workflow engine tool which Parallax are continuing to work on. This allows BCB to make entirely new processes or adapt existing ones when legislation or market changes come into play.

Crucially, the system reflected how compliance officers really weigh a decision: some factors are absolute (a company registered in a sanctioned country stops the process), while others are preferences that shape the review rather than block it. It aggregated these factors into a single, transparent steer for the reviewer - always a recommendation, never a decision, with the reasoning visible so a human could see not just what the AI suggested, but why. Demonstrated successfully, it stands as a proven foundation for a future phase.

The live document-classification tool removes friction at the crucial moment in onboarding, eliminating the back-and-forth communication that stretched onboarding out by days. Now those errors surface the moment a document is uploaded, so customers can correct them on the spot. Fewer applications stall, and fewer are abandoned halfway through.

It also tackles a quieter problem: consistency. Manual review meant different reviewers weighing the same application in different ways. By applying the same checks to every submission, the system gives reviewers a consistent starting point - the same standard, whoever is at the desk.

Just as importantly, BCB now has a proven, human-in-the-loop approach to applying AI in a regulated compliance setting - demonstrating that AI can speed up KYC and improve consistency  without compromising the human judgement those decisions depend on.