In-Woo Park
In-Woo Park

When In-Woo Park talks with a health insurance brokerage about artificial intelligence, he starts with a question that has little to do with models or software. Imagine that every employee on the team suddenly had an excellent intern. What work would they immediately hand over? Park believes the answer reveals more about where AI belongs than starting with a list of available products and hoping employees discover reasons to use them.

"People often start by asking which AI tool they should buy," Park said. "I think the better question is what work your people would gladly delegate if they had someone capable sitting beside them every day." For Park, co-founder and CEO of Fidaris, that shift helps explain why promising AI demonstrations so often struggle to become useful inside service businesses. The technology may work, but the implementation can still begin with the wrong problem.

Park has been testing that premise inside employee health-benefits brokerages, where Fidaris builds AI systems around work that already consumes significant staff time. The company has worked with firms ranging from smaller brokerages to some of the largest in the country. Across the agencies where Fidaris has run these systems, teams have spent significantly less time per client.

The Gap Between a Demo and the Workday

Park points to a widely reported MIT NANDA finding that roughly 95 percent of enterprise generative-AI pilots have failed to produce measurable profit-and-loss impact. He does not interpret the figure as evidence that AI lacks business value. His concern is that many pilots never become part of the way work actually moves through an organization.

"A system can look impressive when someone gives it a clean task," Park said. "The real test begins when it has to deal with the client history, the exceptions, the portals, and the handoffs that make up an ordinary day." Brokerage work rarely arrives as a sequence of isolated questions that can be answered and forgotten. A renewal may depend on past decisions, client documents, carrier information, internal systems, and knowledge accumulated by people who have served the account over time.

Park believes many AI projects underestimate that surrounding structure. A model may generate a polished answer while still leaving an employee responsible for gathering information, moving between systems, checking the result, and completing everything that follows. "The answer is sometimes the smallest part of the job," Park said. "If a person still has to carry the workflow before and after the AI responds, you have improved one moment without necessarily changing the work."

Why Memory Matters in Service Businesses

That observation shaped how Park approached Fidaris. Rather than designing the system around isolated prompts, the company built what he describes as a living client profile that retains information about an account over time. The system links facts back to their sources, identifies conflicting information rather than guessing, and uses accumulated context as additional work is completed.

The approach reflects how brokerage teams already operate. Employees rarely begin every client interaction by relearning an account from scratch. They remember previous decisions, understand unusual circumstances, and know which details have mattered before. "If every conversation starts cold, the human ends up spending time teaching the system the same client again and again," Park said. "The technology needs enough continuity to understand the work it is being asked to carry."

Fidaris applies that approach to processes including renewals, requests for proposals, carrier implementations, plan analysis, contract review, and more. The system gathers information it needs, operates within the portals and software where the work takes place, and requests approval when a decision requires human judgment. For Park, that is a more useful test of AI than whether a model can produce an impressive response in isolation.

Where Approval Belongs

Park sees clear approval points as part of a reliable system. Actions that cannot easily be reversed need defined boundaries, and teams need to understand when the technology is allowed to proceed and when a person must step in. He also believes adoption depends on clear workflow ownership, trust, and training.

"Good automation should make the handoff obvious," Park said. "People should know what the system can carry on its own and where judgment still belongs with them." In client-facing work, that clarity matters because an unreliable output can create more work if employees have to re-check every result. Park argues that useful AI should reduce that burden rather than move it elsewhere.

Starting with Delegation Instead of Software

The "great intern" question gives Park a way to bring the conversation back to work employees already understand. A brokerage leader may not know which type of AI system is appropriate, but that person usually knows which tasks keep skilled employees occupied when they could be spending time with clients or making higher-value decisions.

Park advises brokerages to identify those tasks first. From there, they can evaluate whether an AI system understands the client context, fits the existing workflow, operates where the work happens, and knows when to request approval. That sequence changes the conversation from "Which AI product should we adopt?" to "Which work are we actually trying to delegate?"

His argument is less about adding more AI than becoming more precise about what the technology is supposed to accomplish. A company can invest heavily in new tools while leaving its underlying work almost untouched. Park believes the useful starting point is already visible inside the organization, in the tasks employees would immediately hand to someone capable if they had the chance.

"If you know what your team would hand to a great intern tomorrow, you already know a lot about where AI should begin," he said. "The next question is whether the system can actually carry that work, not just talk about it."