Give a generic but otherwise capable large language model (LLM) five years of physician productivity data and ask it what happened in primary care. It will come back with a clean finding: work RVUs per provider rose sharply in 2021, compensation did not keep pace, compensation per wRVU fell well below benchmark, and the group looks underpaid. The arithmetic is correct. Every number traces back to a row.

Anyone who ran a physician compensation plan through 2021 will read that and wince. That was the year CMS revalued office visit codes. A level four established visit went from 1.50 to 1.92 work RVUs overnight, with no change in the work. Productivity did not rise. The unit of measurement changed. Because survey benchmarks lagged by a year, the comparison was against a denominator from a different era. A health system that acted on the finding as written would have handed out raises to correct a problem that did not exist.

Nothing in the data says any of this. The knowledge lives in the people who were in the room when it happened.

What a generic LLM cannot know

Healthcare finance is full of “edge cases” like these. If a commercial payer's yield drops four points on a surgical service line, the naive reading is underpayment. But an experienced managed care analyst asks first whether implant cases were moved to a carve-out paid as a percent of charges, because that alone would move the blended rate without a single claim being paid incorrectly. Another example: denials appear to spike in a month, but someone who has actually run a billing office knows to check whether a claim edit was turned on upstream and the "denials" are really rejections that never got sent to the payer.

LLMs are genuinely good at reasoning over data. What they cannot do is know which questions an expert would ask before believing the result. That knowledge isn’t in a textbook. It's not on the internet in any form a model could have learned from. It’s in the heads of people who have sat in a health system's finance department and been wrong before in expensive ways, or been guided through it by a senior colleague with experience and judgment.

The first post in our trust series said that experts define what good looks like. This post is about who those experts are and how we translate their knowledge into our product.

We hire people who have done the job

Translucent has a team whose only job is expertise in the details of healthcare finance. It isn’t a quarterly advisory board or offshore annotation vendors, but full-time, in-house employees who came from revenue cycle, managed care contracting, physician enterprise finance, and health economics, and who now sit in our office next to the engineers building the agents.

We call it the Domain Intelligence team.

One of us spent more than two decades running revenue cycle teams at multi-hospital specialty health systems. One of us has over a decade of experience leading clinical operations for a premier academic medical center, including overseeing surgical, medical, and ancillary departments. One of us has negotiated payer contracts from both sides of the table. Our experience is both deep and wide.

There are two reasons that domain expertise has to be a full-time function.

The first is the feedback loop. When an agent produces something like the erroneous finding above, the person who catches it needs to be reading agent output that day, not reviewing a summary next quarter. The time from "that is wrong" to "that cannot happen again" should be measured in hours. The fix isn’t a note in someone's inbox, but a rule change in the platform’s context layer so that every agent at every health system runs with the fix from then on. That only works if the people who know sit inside the company.

The second is ownership. An outside expert reviews a sample, signs a timesheet, and leaves. Whether the agent was right in a customer's board pack three months later is not their problem. At Translucent, we want to own the entire problem so we can make sure it doesn’t happen again. To that end, the same Domain Intelligence members who have their names on the rubrics that decided an output was ready to ship sit in on the customer readouts and hear firsthand when a number did not pan out. That changes the standard from "plausible to a reviewer" to "would I personally hand this to a CFO and stand behind it?"

That’s a much higher bar, because it also includes whether the finding is something a finance team can actually act on. A correct answer nobody can use is a failure too. Holding that standard on every output is only possible when it’s someone’s entire job to ensure the accuracy and utility of agent output.

Getting what we know into the platform

"Experts in the loop" is what every AI vendor trying to convey trustworthiness says. Internally, we say "encoding our brains" but that’s more of an internal shorthand than a real explanation to customers. In reality it means four things.

1. Opinionated definitions, and the ones that disagree with ours. 

Net revenue, yield, wRVUs, cost to collect—none of these has a single definition in healthcare, and an agent that picks one without knowing which one you use will produce a number that looks right on the surface but never reconciles. 

We write each metric’s definition into the context layer: which inputs it takes, which variants exist in the field, whether wRVUs are counted from claims or from the encounter system, or whether two variants can be reconciled or are simply different measures.

2. What to check, and in what order. 

A CFO sees that cardiology's contribution margin fell six points and asks why. There’s a short list of things that could explain it: fewer cases, a shift toward cheaper cases, more Medicare and less commercial, a contract that reset, a coding change, a cost allocation change, a department that moved between service lines in the ledger. And a combination of any of these.

Some of these happen every quarter and some almost never. An experienced analyst knows which to check first and what evidence rules each one out, and does it so automatically that they could not tell you they had a list. 

We write these lists into the platform’s context layer. Agents work through them in order and report what they ruled out along with what they found, so the CFO sees not just the conclusion but counterarguments and their rebuttals.

3. Turning an analysis into steps an agent can run. 

For every question a finance team asks, there is a way an expert would work it. What should a negotiation-ready contract analysis contain? What will an executive do with a denials finding, which determines what belongs in it?

Making this knowledge usable by an agent means taking the analysis apart: the inputs an expert pulls, the order they work in, the checks they run before they believe a number, and the form the answer has to take. Each of these is written down so an agent can carry it out and a human can see exactly which step went wrong. 

These playbooks are the healthcare finance context our agents carry into every question. They’re written by people who have produced the deliverable for a real CFO, then revised every time an output comes back short. A single revision compounds beyond just that customer, because a gap addressed for one health system's contract analysis is closed for every health system's contract analysis.

4. Evaluation by people who know what a wrong answer costs. 

The same people grade the output. Everyone scoring has watched a wrong number reach a board, a compensation committee, or a payer negotiation, and knows what it took to walk it back. Every agent result is scored against a rubric written by the team, on whether it is client-ready, whether the reasoning holds, and what specifically doesn’t pass muster.

When something material changes, such as an update to the context layer, the team runs a standard set of questions through both versions and scores the results blind, against the same rubric. The outputs are blinded so the scorer cannot tell which arm they are looking at. Scoring uses anchored rubrics rather than gut feel. Runs are repeated to measure how consistent the agent is before any single result is trusted, and the person who holds the key does not score. It’s the scrutiny a good manager gives an analyst's memo, with a scientist's controls added.

Today, the team reviews every Opportunity Agent finding, and their notes on accuracy, tone, and actionability feed back into the context layer. Human review is greatly reduced only where automated scoring consistently matches the experts' judgment.

What Domain Intelligence does not fix

Even encoded knowledge goes stale. CMS changes rules every year, payers change behavior faster than that, and the most likely cause of a margin drop in 2024 won’t be the most likely cause in 2026. Keeping the context layer current is ongoing work, which is why the team is permanent rather than a one-off project.

Some knowledge is specific to a customer, like how one health system defines mission-based P&L, or which departments roll to which service line. We can hold their definition and show the arithmetic under it, but we cannot know it in advance, which is why the intake work with a new customer is largely a list of questions our experts would ask on their first day working in that finance department. Once those definitions are captured, every agent runs with them, and the numbers a CFO gets from us reconcile to the ones they already publish. 

What Domain Intelligence buys a finance team

Domain Intelligence plays a critical role in the two ways the Translucent platform delivers value. The first is time. The workbook that explains a margin drop but took three weeks to prepare arrives as a complete report with the checks that were run and the explanations that were ruled out. The second is findings that nobody would otherwise have uncovered. Human analysts can only chase the variances they’re aware of. Agents work every service line, every contract, and every cost center continuously, and they surface the leakage that was never anyone's project. But to do this without generating false alarms at scale requires agents to have been taught what to be suspicious of in their own work.

What a health system’s finance team does with that is the part that always needed a human. Instead of explaining the past, they spend their judgment anticipating the future: what a contract renewal does to margin at the payer's opening offer, which service line is exposed if the Medicaid mix keeps moving. Agents lay out the options and size the scenarios. The people who know the organization best decide what to prioritize and how to act. LLMs were always going to be good at the math. The work was teaching it what to be suspicious of, so humans can get back to deciding.

‍