Building AI agents in sensitive financial enterprises

On This Podcast

In this conversation, we examine why Qamar Hussain's team at Webio chose to fine-tune their own models rather than rely on foundation models or third-party APIs for debt collection conversations. He explains that out-of-the-box models scored only 20 to 30 percent accuracy on real classification tasks such as spotting vulnerability, payment intent or callback requests, far below the high bar required before anything reached production. We also discuss how having years of high-quality, anonymised real conversation data, rather than synthetic examples, was the decisive factor in getting the models to a usable standard. Qamar notes this gives incumbents with existing conversation data a real advantage over AI startups that must build a dataset from scratch before their models can perform reliably.

In this conversation, we examine the compliance architecture that lets Webio operate AI in a sensitive financial domain: there is no public-facing AI interface, every interaction is mediated through a human-configured platform, and strict guardrails limit the AI to narrowly defined tasks like classification and summarisation rather than open conversation. Qamar describes a deliberate anonymisation process, part manual and part automated, that strips personal information before any data reaches a training set, and models built so that no underlying information can ever be extracted from them. We also discuss how this differs from the open-ended chatbot model most people associate with AI risk, and why constraining what the AI is allowed to do actually raises quality rather than limiting it. The result, Qamar argues, functions less like an autonomous agent and more like a closely supervised co-pilot for human agents.

In this conversation, we examine why Qamar sees maintaining AI systems as fundamentally different from maintaining traditional software: the same input, such as a message to classify or summarise, can produce different outputs from one run to the next, unlike a deterministic function that either works or fails. He walks through a real example where a new complaint intent had to be created after the team noticed messages being misclassified as vulnerability, illustrating how continuous evaluation catches drift that traditional testing would miss. We also discuss why Qamar now believes a QA mindset, the instinct to hunt for failure points and tolerate ambiguity, matters more when hiring for AI roles than existing AI or engineering credentials. He argues he would rather hire a skilled QA engineer with no AI background than an AI specialist without that detail-oriented, sceptical approach to quality.