Second in a series on building trustworthy AI for healthcare finance. The first post laid out three areas where trust has to be established: the data, the math, and the answers. This one is about the math.

Ask a finance team why a service line’s margin moved and you’ll get an answer in about three weeks. What arrives is not really an answer, but a delicate machine that produces one: a multi-sheet workbook built from scratch, wired to a particular data extract, structured around the question exactly as it was asked. If you want to follow up on the initial answer, the machine can’t adapt. A human must rebuild it.

Three things are true of that workbook. It was expensive to build. It’s a snapshot of the moment in time it was built, so it reports on data that stopped being current the day it was finished. And though it's transparent in principle, it’s opaque in practice: thousands of cells with assumptions documented nowhere.

The real cost is not error, though errors happen. The real cost is the questions nobody asks: whether a variance is actionable, what actually drove it, which of several plausible explanations survives contact with the data. When a line of inquiry takes weeks, a finance team learns to ask the questions it can afford instead of the ones that might actually move margin.

None of this is the analyst's fault. The logic in the workbook is sound, and it took real expertise and healthcare experience to build. The problem is that the logic is trapped in the artifact, and the artifact has to be rebuilt every time the question changes.

Where the logic lives instead

With agentic AI, we no longer have to rebuild an artifact for each question. The logic that used to live inside the workbook is encoded in our context layer as rules your organization has already defined. Calculations that work the same way at every health system exist as tested code. What an agent composes on demand is which of these calculations to run and over which data.

There is no model to rebuild.

What has to be true for this to be safe

Calculations an agent composes on demand are only trustworthy if three conditions hold.

The math runs as code, not as language. The agent writes a query and code, and a sandbox runs them. The agent never produces a number itself. It could select the wrong rows or write code that computes something incorrectly, and those are real risks. Wrong rows and wrong logic are both immediately discoverable, because the query and the code are written down and a human can read them. But a number that a model simply asserts is not findable, which is why we let the model decide what to calculate and never let it do the calculating.

The terms are defined upstream. Net revenue, encounter, yield rate: each resolves to mapped fields before an agent touches them. The agent doesn’t infer what an analyst meant by a column name, but instead follows an industry-wide or institutional definition.

The derivation is the deliverable. What travels with the answer is the derivation: the query, inputs, code, and result. Review means a human reading a derivation rather than auditing a spreadsheet.

How agents transform finance work

A human analyst asks a question and an agent calculates and documents its work. The human reviews the agent’s work and signs off. The review and signature don’t change, and they shouldn’t.

What changes is that the reviewer reads a calculation instead of inheriting a workbook. And the full line of inquiry needed to arrive at an action plan for moving margin, which would have taken weeks and therefore never got run, now takes minutes.

Next: how the context layer resolves a term like net revenue to your organization's own definition, and why generic definitions produce answers that look right.

‍