The first version of Translucent was a chat interface on a health system's financial data. It wasn't simply slapping a generic large language model (LLM) over raw data: it ran on a context layer we built specifically for healthcare finance. We assumed that if the answers were right, the insights would follow.

The answers were accurate, but they rarely led to an opportunity anyone could act on, and watching people use the product showed us why.

Users didn't know where to start. The first questions were broad, like "tell me my numbers." Someone would type "how is margin doing" and get a number. They'd type "why is it down" and get a paragraph that was true and not useful.

That isn't a failure of the people asking. A chat interface can only answer the question it was given, and the best questions start with a suspicion. You ask about a payer contract because you already think something is off with it. Most of a finance team's suspicions come from the places they already look: the close, the variance report, the board pack. The places nobody thought to look don't generate questions.

Better answers don't produce better questions

Many finance leaders expect that a general-purpose AI assistant pointed at their data will start surfacing insights. Better LLMs and better context do improve the answers, but neither changes who has to think of the question.

Our first fix was to suggest questions: prompt users with things worth asking, so they weren't starting from an empty box. But because the suggestions were written in advance, they just pointed at the same places every health system shares instead of what was unusual about their system’s data. It helped people get started. It didn't change what they found.

Starting from the data instead of the question

When agentic approaches became viable, we stopped trying to improve the question and took the question out of the starting point. The Opportunity Agents we’ve developed work behind the scenes across a health system's contracts, service lines, and cost centers, and surface opportunities to improve margin without anyone asking for them. Instead of answers to questions, finance teams get findings they can directly act on.

Each opportunity arrives sized, root-caused, traceable, and documented, with a recommended action for capturing it. Sized means it carries an estimate of what it's worth, so a CFO can decide whether it deserves a team's time. Root-caused means it says why the gap exists. Traceable means every number in it leads back to the source data. Documented means the report shows the checks that were run and the explanations that were ruled out. The recommended action follows from the root cause, and the decision to act stays with the finance team.

The reason this takes agents is scale, and it's two kinds of scale. Telling a real pattern from noise means looking broadly, across every contract and every service line. Finding the root cause means testing several possible explanations at once. A finance team can do either one for a project it has already chosen, often over days or weeks. It can't do both across everything.

$70M nobody was looking for

At a large academic health system, agents compared what each commercial carrier paid against its peers and surfaced two findings worth just over $70 million a year. Both were payer contract problems, but they had different causes and called for different actions.

The first was a facility rate gap worth $29 million. One carrier's facility yield trailed the peer median across more than 15 outpatient service lines, in specialties with little in common clinically. A gap that broad ruled out problems specific to one service line or to product mix, and pointed to the carrier's base facility rate schedule. The recommended action was to take on that rate schedule in the next negotiation, starting with orthopedics and women's services, where the gaps were largest.

The second, worth $41 million, was in oncology. Three of the four large commercial carriers landed within about a point of one another on oncology outpatient yield, all roughly eight points below the fourth, even as charges rose across every carrier. Three carriers clustering that tightly didn't look like three separately weak negotiations. Together with one carrier's yield falling for two years while its charges climbed, it pointed to how the contracts reimburse oncology drug costs. The recommended action was to make that reimbursement structure the starting point for the next oncology negotiation.

The health system's managed care team confirmed both findings, and both were tied to contracts already under negotiation or coming up for renewal.

The size of the number isn't the most telling part. No one at the health system suspected these contracts were a problem. So no one would have spent a couple of weeks building the analysis to find the cause, or asked a chat assistant to look, because there was no reason to ask.

That is the pattern behind most of what our Opportunity Agents find. The findings that move margin are almost never the ones somebody thought to ask about. They’re the ones that were sitting in the data while everyone was busy closing the month.

Where chat fits now

Chat in our product is a different thing from a general-purpose LLM pointed at your data. It still carries the healthcare finance context it always had, but it no longer starts from an empty box over all the data. It starts from the opportunities the agents surfaced and the KPIs a finance leader is already looking at, which settles the question our first users couldn't answer: where to start. How chat works from there deserves its own post, and we'll write it.

Opportunity Agents decide where to look, across everything, before anyone asks. The people who know the organization decide which findings matter, and chat is how they dig into the ones they choose.