There is a widely held assumption in financial services that deploying AI delivers a competitive edge. Give people a powerful model, connect it to your data, and productivity should follow. 

In practice, it rarely works that neatly. 

The constraint is no longer the model. It is how the information shows up. 

The Research Is Catching Up 

Earlier this year, researchers from UC Santa Barbara and KAIST published a paper that deserves more attention in Capital Markets than it has received. They asked a group of finance professionals to complete a complex valuation task using GPT-4o, tracking hundreds of interactions and measuring cognitive load at each step. 

The headline is simple. AI helped. Output improved. But what was more interesting is what got in the way. The issue was not the quality of the answers, but how those answers were delivered. 

The researchers focused on extraneous cognitive load, the effort created by presentation rather than the task itself. This had a significantly greater impact on output quality than the underlying difficulty of the work. 

Put plainly, people were spending too much time deciphering the answer rather than using it. 

The behaviours will feel familiar. The model wandered into topics that were never asked about. It tried to do several things at once. It produced long blocks of text where structure would have been more helpful. And once a thread became messy, it stayed messy. 

None of this looks fatal on its own. Over time, it compounds. Performance slips. 

The least experienced users were hit hardest, precisely the people who stand to benefit most from AI assistance. 

Why This Matters in Capital Markets 

This problem is more acute in Capital Markets because the work is neither neat nor linear. 

Professionals combine structured data with unstructured inputs such as filings, news and research. They do this under time pressure, where hesitation has a cost. Outputs need to be shared, defended and audited across clients, risk teams and compliance. 

Different roles also want different things. A portfolio manager, an analyst and a risk officer may start with a similar question, but they do not want the same answer in the same format. 

Generic AI tools ignore this. They treat every query as a text-in, text-out exercise and leave the user to turn it into something usable. 

That is where the friction sits. 

You notice it in small ways. A response comes back. You pause. You read it again. You strip out what matters. Occasionally, you rewrite the whole thing before you can move on. 

Individually trivial. Collectively expensive. 

Designing for How People Work 

This is not a prompt engineering problem. It is a design problem. 

The research highlights unsolicited task switching as a major driver of cognitive load. Avoiding it requires real context. Who is asking, what they are trying to do, and where they are in the workflow all matter. 

That context has to be built in from the start. 

At Engine AI, we start with the domain rather than trying to retrofit it. User-specific workflows are not a feature, they are the foundation. 

A portfolio manager asking about fund flows needs a different response to a risk officer looking at the same position. Not just different content, but different structure, sequencing and level of detail. 

The research also points to progressive disclosure. Users should be able to go deeper when they choose, rather than being presented with everything at once. Our interfaces are designed so users can shape what they see in real time, without needing technical support. 

Underneath that sits a unified data layer, combining structured and unstructured sources so outputs remain coherent, explainable and auditable. In this industry, that is not optional. It is expected. 

The Broader Point 

Access to better models will not be the differentiator for long. They are improving quickly and becoming widely available. 

What will separate firms is whether their AI fits the way their people actually work. 

The research makes this hard to ignore. Poorly delivered information is not just a user experience issue. It shows up directly in output quality, and it hits hardest where AI should be helping most. 

AI is only useful in context. 

The real question is not whether you have AI. It is whether your AI understands the job well enough to make the person doing it genuinely more effective. 

That remains the gap. And it is the one that matters.