In highly-regulated fields like finance and medicine, false confidence can be fatal. Models are getting faster and cheaper—but they still make things up. Until hallucinations are under control, trust is off the table. Navigating the precarious terrain is David Mataciunas, Co-Founder and CTO of AQ22, a firm building AI agents for the financial sector. AQ22's platform enhances loan assessment efficiency for private equity lenders and commercial banks.

Keep it real: Mataciunas' team has seen both the promise of models like GPT-4.1 and the persistent risks that come with them. For AQ22’s core markets, precision isn’t optional; it’s the baseline. "Finance and medicine are huge applications for AI and chat," he says. "It’s very important that everything is as accurate as possible right from the first pass." That demand for reliability is what keeps many regulated sectors cautious. "The biggest problem right now isn't even the cost," explains Mataciunas. "It's the hallucinations, because you need to build so much on top to stay safe and compliant."

Time for a tune-up: Persistent unreliability is driving demand for better tools, with refinement at the top of the list. "I’m really waiting until we can fine-tune 4.1," says Mataciunas. "If it’s a smaller model and we can use our own data, we can get better results. But fine-tuning right now is kind of costly." The desire for more control is widespread, yet access to fine-tuning on advanced models like GPT-4.1 is still limited. And while price and context windows are a major differentiator from open-source models, dependability remains the sticking point.