A health system CFO asked me recently why he should let software act on his money without asking him first.

It is the right question, and it is worth noting that it is not a technology question. He was not asking whether the software could find the underpayment. He assumed it could. He was asking what he would be signing up for if he let it do something about it.

That is a delegation question, and we have answered one before.

You already learned to trust a machine with something that mattered

Twenty years ago you printed directions. The answer was only as good as the person reading the map, and if the map was wrong or the driver missed an exit, that was the end of it. You trusted the driver.

With the advent of Google Maps and Waze, the system started watching the road. Live traffic, rerouting around a closure before you reached it, an arrival time that updated itself. A real change, and an incomplete one, because a human still made every decision and took every action. You trusted the guidance.

In Midstream’s hometown of San Francisco, you can’t throw a rock without hitting a self-driving car. Robotaxis like Waymo and Zoox perceive, decide, and move autonomously inside defined guardrails, in a defined zone, and gets better with every mile driven. You trust the system.

That progression did not happen in one step. Each stage earned the next by being proven: you could see what the system saw, check its reasoning against your own, and watch it be right enough times that checking stopped being worth the effort. Trust was not granted. It was accumulated, in public, with the steering wheel still in reach.

Financial operations is walking the same ladder right now, and most health systems are further down it than they believe.

The levels

Most health systems today sit at Level 0 or Level 1. A quarterly margin review assembled by hand is Level 0. A dashboard that shows contract compliance and waits for someone to read it is Level 1. Both can represent excellent work by excellent teams, but neither moves a dollar without a person deciding to move it.

The level that everyone mistakes for progress

Level 2 is where the market is crowded and where self-assessment goes wrong.

Level 2 feels transformative from the inside. Ask a question in plain language, get an answer in seconds that used to take an analyst three days. Root cause explained. Variance attributed. The margin analysis that used to be rebuilt every time an assumption changed now regenerates itself.

But the dollar still has not moved. Somebody still has to validate the finding, decide it is worth pursuing, find out who owns it, contact the supplier, escalate when the supplier does not respond, track the credit, and confirm it landed. Seven handoffs, each one a place where a real finding quietly stops being a real dollar. This is why the gap between identified savings and realized savings is the most reliably disappointing number in healthcare finance, and why it barely narrows when a faster analytics layer gets installed. Speeding up the first step does nothing about steps two through seven.

There is a simple diagnostic. Ask what your savings number means. If the number your team reports is identified, you are at Level 2 or below, whatever the software was sold as. If it is banked, meaning verified against the contract and reconciled to cash received, you are at Level 3 or above. The two numbers are usually different by a factor that nobody wants to say out loud.

Visibility is easy. Autonomous action is hard.

Here is the part that gets skipped, and it is the reason the market is about to fill with Level 4 claims that are Level 2 products.

The hard problem in autonomous driving was never seeing the pedestrian. Cameras have managed that for a decade, which is why nearly every car sold today can beep at you when a pedestrian walks behind when backing out of a parking space. Very few of the cars sold today can drive themselves. The distance between those two facts is the entire industry.

The hard problem in autonomous driving was never the decision to brake. It was building an atomic-level model of the world detailed enough that braking was the right call: every pedestrian, lane marking, sign, occlusion, and edge case, labeled and modeled in granular detail, and then years of encoding what to do in each scenario. Acting autonomously is where the difficulty lives.

Financial autonomy works the same way. Identifying an underpayment or missed rebate is not actually the hard part. Claiming that underpayment autonomously is a different discipline. The system has to know the clause that governs the tier, the purchase history that establishes the threshold, the eligibility status of the buying entity at the time of purchase, the unit of measure reconciled across three systems that disagree, whether the compliance form was ever filed, and the reimbursement on the other side that determines whether the item was worth buying at all. Then it has to be right, because a claim filed in error costs credibility with a supplier you will negotiate against every year for the next decade. Then it has to know which form, which window, which contact, what to do when the supplier does not respond, and how to argue the appeal after the window has closed.

Almost none of that is model capability. It is operational knowledge, and today it lives in the heads of people who have run contracting, rebates, and P2P inside complex health systems. Getting it into software requires those people working directly alongside the engineers building the agents. That pairing is the scarce input. Either half on its own produces something the market already recognizes: technologists ship a beautiful dashboard, operators ship a services engagement, and in both cases a human is still the thing that moves the dollar.

That foundation takes years and it cannot be shortcut. Which leads to an uncomfortably useful test: when a vendor describes autonomous action, ask what the world model underneath it is, how long it took to build, and how the system knows a finding is true. Visibility is table stakes. Autonomous action is the hard part, and it is the only part that produces a dollar.

Four requirements before you delegate anything

Level 3 and above should be earned, not assumed. Four properties are worth insisting on regardless of who you buy from.

A defined action space. Autonomy inside a boundary is a system. Autonomy without one is an incident waiting to be written up. The set of actions an agent may take without asking should be written down, approved, and narrower than you think you need at the start.

Traceability to source. Every action should resolve to the clause, invoice, purchase order, or claim that justified it. This is what makes an autonomous system auditable, and auditability is the entire basis of accumulated trust.

Reversibility. Prefer actions that can be unwound over actions that cannot. Filing a rebate claim, submitting an appeal, and flagging a price variance are recoverable. Committing volume is not. The order in which you delegate should follow that line.

Accountability for the result. Someone should own the outcome, in a way that appears in a number they are measured on. Autonomy that diffuses responsibility produces activity rather than dollars.

Level 5 does not mean finance stops deciding

The misreading of Level 5 is that the humans leave. What actually happens is that the work is reassigned along the line where machines and people are respectively good.

Your team runs on 80/20 today, because that is the only rational way to spend scarce attention. They work the big rocks: the system-wide device negotiation, the payer contract, the service line decision. Those are judgment calls involving relationships, clinical preference, political capital, and strategy, and they should stay judgment calls.

What 80/20 leaves behind is the long tail. A tier discount unapplied. A compliance form unfiled. An invoice overpaid by three percent. Individually too small for anyone to be assigned to chase, because the return does not justify a person's afternoon and the headcount does not exist anyway. Collectively, in a multi-billion-dollar cost base, they are the difference between hitting the margin target and explaining why you did not.

Level 5 is a workforce of agents working that tail continuously while finance supervises the whole and keeps its people on the decisions that require a human. Automate the twenty. Augment the eighty.

The cost of staying where you are

None of this is urgent because the technology is exciting. It is urgent because of arithmetic.

Total hospital expenses grew 7.5 percent in 2025, more than double the 3.3 percent rate of hospital price growth. Supplies outpaced that at 9.9 percent and drugs at 13.6 percent. Expense growth compounds annually against revenue often outside your control, and the industry's historical answers have all been tried at scale. Nonprofit hospitals spent $7.8 billion on management consulting between 2009 and 2023, at an average of $15.7 million per engagement, and researchers publishing in JAMA this May found no evidence of meaningful improvement in finances, operations, or quality of care at the hospitals that hired them compared with matched hospitals that did not. The likeliest explanation has little to do with the quality of the analysis. Their findings were real and they likely arrived as a deck. But the deck was often the end of the engagement. The consultants turned in their badges, and somebody still had to validate each finding, assign an owner, contact the vendor, escalate the nonresponse, and confirm the dollar landed, while the consultants were already embedded at the next health system.

The levels are not a maturity score to feel good about. They are a series of decisions about what you are willing to delegate, and each one you defer has a price that shows up in a number your board already looks at.

The question the CFO asked me was the right one. It deserves the same answer that driving got: show me what the system sees, let me check it, keep the wheel within reach, and let it earn the next level.