FOUNDER PERSPECTIVE
Artificial Intelligence Should Help Us Understand, Not Just Answer
An explanation should make its foundations easier to examine.
By Brian Holman · Founder & CEO, Dezota LLC
A clear explanation can be enormously helpful when a subject feels complicated. That is part of what makes artificial intelligence interesting to me. It creates opportunities to help people work with information in more approachable ways. But an explanation also carries influence: the easier it is to read, the easier it may be to accept. For financial software, I believe the quality of that explanation has to include whether its reader can understand what supports it.
My independent research has included work with large language models, retrieval-augmented generation, content transformation, and semantic search. These ideas explore different parts of an important problem: finding relevant material and making it more useful. The terminology matters less to a person using a product than the outcome. Can they find what they were looking for? Does the explanation preserve the meaning of the source? Can they recognize where an interpretation has been added?
Retrieval-augmented generation is a useful example of the distinction I want to preserve. It brings selected source material into the process of producing a response. That makes the selection and interpretation of those sources important parts of the design. I would still want to ask whether the material is relevant, whether it covers the question, and whether the answer goes beyond it. Attaching a source to an explanation should be an invitation to examine the connection, rather than a reason to stop asking questions.
Consider a hypothetical question about why an account balance changed. A fluent answer might connect the movement to a broad market story. But the records could also contain deposits, withdrawals, or information from different dates. I would want a financial tool to work from those records, distinguish the arithmetic from the explanation, and make any remaining uncertainty clear. An answer that sounds reasonable should not quietly take the place of evidence about what actually occurred.
This also shapes how I think about evaluation. A useful demonstration is a beginning. The harder questions arise when information is missing, two sources disagree, or a familiar-looking document contains something unexpected. We need to consider what the system should leave unanswered and when a person should examine the result more closely. Those are design responsibilities I want us to take seriously as the work develops. A polished response is only one part of the experience we should be evaluating.
There is room for ambition here without assuming that every question needs an AI-generated answer. A clear calculation, a well-labeled source, or a simpler screen may sometimes do the job better. As architect of WealthProof, I want the choice of technology to follow the need. AI is valuable to our direction when it helps someone find, interpret, and question information with more understanding. The goal is a person who is better equipped to exercise judgment, with the technology supporting that responsibility.