AI may be changing compliance quickly, but building a clever model is only part of the job. Financial institutions also need to know that it continues to work, understand why it has reached a particular decision and be able to explain that decision when it matters.
In our latest Power 50 podcast, host Mark Waker,editorial director at The Fintech Times, is joined byLaksmi Rippe, product marketing manager at Sis ID, and Fintech Power 50 influencer Dr Ruth Wandhöfer to discuss what financial institutions should really be looking for as AI takes on a bigger role in compliance.
One of the challenges they explore is model drift. An AI model that performs well today will not necessarily behave in exactly the same way months or years down the line as data, customer behaviour, regulations and communication channels change. That makes ongoing monitoring just as important as how a model performs when it is first introduced.
Listen to the Power 50: The Real Battle in RegTech Isn’t AI – It’s Auditability here.
The conversation also looks at the sheer volume of data compliance teams are dealing with. Genuine problems can represent a tiny proportion of otherwise legitimate activity, leaving technology with the difficult job of finding meaningful risk without generating so many false positives that teams are overwhelmed. The quality, selection and labelling of the data used to train models therefore matters, alongside human oversight and feedback.
Mark, Laksmi and Ruth also challenge the assumption that bigger AI models are automatically better. For narrowly defined compliance tasks, smaller and more specialised models can sometimes be cheaper and easier to monitor, understand and debug than large general-purpose models.
Ultimately, the discussion comes back to trust. As financial institutions rely more heavily on AI within compliance, knowing that a model has raised an alert is not enough. They need to understand why it was raised, how dependable that decision is and whether there is a clear trail behind it if regulators come asking.
