What an unusually candid conversation about healthcare economics revealed and why AI may change the equation rather than merely rearrange it.
About halfway through an evening of healthcare policy and economics in New York, I found myself having a fairly simple reaction:
I’m not sure the numbers square.
COPE Health Solutions and Ropes & Gray had convened The Value Agenda on September 16, bringing together people who see the healthcare system from very different vantage points: federal and state policy, Medicaid, health plans, safety-net hospitals, independent physician networks, and value-based care. The stated purpose was candid conversation among people making real decisions under real pressure, not another panel agreeing that healthcare is complicated.
As the discussion moved from coverage to reimbursement to hospital economics to social care, no one was really arguing that the goals were wrong. Quite the opposite. Most were individually difficult to disagree with. Keep people covered. Improve access. Pay providers adequately. Keep premiums affordable. Move care out of expensive settings. Preserve safety-net institutions. Address food, housing, transportation, and other social needs that affect health. Put providers at financial risk for outcomes. Invest in transformation. Reduce avoidable utilization.
Taken together, however, they began to sound less like a strategy than a set of simultaneous equations without an obvious solution. That became clearest in the discussion of Medicaid.
Coverage Is an Operating Problem
Beginning January 1, 2027, states must implement the new federal community-engagement requirement for “applicable individuals,” subject to statutory exclusions and exceptions, unless they choose to implement it sooner. New York is preparing for that implementation now. (Medicaid)
Whatever one thinks of the policy itself, implementation immediately creates another question: can someone who satisfies the substantive requirements successfully demonstrate that fact and remain covered?
Brett Friedman, a healthcare partner at Ropes & Gray and former New York State Medicaid Director, kept returning to what he called the individual’s “coverage journey”: eligibility systems, managed-care plans, local agencies, documentation, and human intervention that determine whether someone who qualifies on paper actually remains insured. Theresa Dreyer, CEO of the Health Care Transformation Task Force, a national coalition of healthcare providers, payers, purchasers, and patient organizations focused on value-based care, described the mundane ways that journey can break down. Someone may be working but self-employed and struggle to produce the expected evidence. Someone may qualify for an exemption but still have to establish it. The problem is not necessarily that the person fails the substantive requirement. The problem may simply be that the system requires something the person cannot readily produce.
Jing Shui, COO and Chief Strategy Officer of CAIPA, the Coalition of Asian-American IPA, a large New York network of independent physician practices, then described what that distinction looks like operationally. CAIPA had taken a Point of Care application originally built to surface clinical and risk-adjustment gaps and loaded Medicaid recertification dates into it. When a patient appeared in a physician’s office, the system could flag an upcoming renewal, and the physician could direct the patient toward nearby enrollment help. Shui reported a 93% recertification rate from that effort. CAIPA identifies Shui as its COO and Chief Strategy Officer and describes a network of more than 1,400 independent physicians and healthcare professionals. (CAIPA)
That example stayed with me because it reduces a very large policy argument to its actual unit of execution. Coverage is not achieved merely because a statute says someone is eligible. It is achieved when a person remains successfully connected to the machinery that turns eligibility into insurance and that machinery costs money.
The next problem is what happens after the person is covered.
Value Is Constrained by Where the Money Sits
Errol Pierre, Chief Growth Officer of Healthfirst, a not-for-profit health plan serving roughly two million New Yorkers, speaking from the health-plan side, connected access directly to reimbursement. He described patients with different insurance products experiencing very different appointment availability and argued that reimbursement affects how quickly patients can get into a physician’s office or reach a specialist. He also described value-based arrangements as particularly vulnerable when economics tighten: shared savings and risk feel very different when there is surplus to distribute than when every participant is under pressure.
The broader evidence is more complicated than a simple “higher payment equals better access” formula. The Medicaid and CHIP Payment and Access Commission (MACPAC), the nonpartisan legislative-branch agency that advises Congress, HHS, and states on Medicaid and CHIP policy, has repeatedly examined the relationship among reimbursement, provider participation, and access. Low payment is frequently cited as a barrier to Medicaid participation, while administrative burden, provider supply, specialty, geography, and other factors also matter. (MACPAC)
We often talk about good care, accessible care, and affordable care as though they were three independently adjustable settings. They are not. Payment is somebody’s cost and somebody else’s revenue. Reduce reimbursement far enough and some providers respond through participation, staffing, capacity, service mix, or appointment availability. Increase reimbursement and the additional expense has to appear somewhere else—in premiums, taxes, employer spending, public budgets, patient costs, or another part of the system. “Affordable,” in other words, is an incomplete adjective in healthcare.
Arthur Gianelli, President and Chief Transformation Officer of One Brooklyn Health, put a more difficult version of the problem on the table. Safety-net systems like his are being asked to move care out of hospitals, expand ambulatory and community-based care, reduce avoidable inpatient use, and transform how they deliver care. But some of the revenue streams helping keep those institutions financially viable remain tied to the legacy structure they are being asked to shrink. Gianelli was not simply describing hospitals that need more money. He was describing an organization that can understand what transformation requires and still find that its own economics penalize it for pursuing that transformation. In his telling, the future model is not mysterious. Ambulatory capacity should grow. Hospitals should build stronger relationships with community organizations. Data should move more effectively across institutional boundaries. Care should migrate toward settings better suited to providing it.
Then the finance team runs the numbers.
For his organization, moving away from the legacy inpatient model could also mean changing revenue streams connected to that model, including graduate medical education and state-directed payments. Those financing mechanisms are themselves complex and should not be reduced to simple “payment for beds” formulas. But Gianelli’s institutional point was clear: the economics supporting an incumbent delivery structure do not automatically migrate with the organization when it tries to build the structure intended to replace it.
That was Arthur’s math: the care model can be directionally right while the institution’s economics still punish it for moving there.
Healthcare often describes this family of misalignments as a wrong-pocket problem: the organization that pays for an intervention may not be the organization that receives the economic benefit. But Arthur’s example is broader than that. It can also be the wrong time horizon. The wrong contract. The wrong unit of accountability. The wrong regulatory boundary. Or, more fundamentally, a revenue structure that asks an institution to dismantle part of the economic engine financing the very transformation it has been told to undertake.That is a much harder problem than persuading people to embrace change.
The same tension appeared in the discussion of social care. New York has established nine regional Social Care Networks under its Medicaid 1115 waiver to screen Medicaid members and connect eligible people with health-related social-needs services. The state reports that more than 1.5 million members have already been screened. The current waiver runs through March 31, 2027, and New York is pursuing an extension. (New York State Department of Health)
Shui’s instinct was to stop treating that infrastructure as something running parallel to medical care and begin integrating the two, clinically, through data, and eventually through more durable funding as well.
That seems directionally right to me. But there is an important counterargument: not everything that affects health automatically belongs on the healthcare balance sheet. Housing policy does not become medicine merely because housing affects health.This is not an argument against investing in housing, food, transportation, employment, education, or social connection. I regard those as fundamental to a healthy, functioning civil society. It is an argument about which institutions should finance them.
Sometimes solving the wrong-pocket problem means finding a sustainable healthcare payment model. Sometimes it means acknowledging that the right pocket is not healthcare’s pocket at all. Either way, the recurring pattern is revealing. We identify something that matters. We create a special program to address it. We stand up organizations, contracts, data flows, reporting requirements, and administrative structures. Then everyone asks how to make the funding sustainable once the special program ends.
Blocking and Tackling Is Necessary but Not Sufficient
None of this means execution is irrelevant. In fact, Allen Miller, Principal and CEO of COPE Health Solutions and a longtime advisor to payers and providers on value-based payment, risk arrangements, population health, and clinically integrated networks, made the strongest argument of the evening against allowing structural problems to become an excuse for poor operating performance. His phrase was “blocking and tackling.” Understand the contract you signed. Close care gaps. Know when patients are admitted. Reduce avoidable admissions and readmissions. Use the data you actually have rather than waiting for perfect data. Pay physicians in ways that engage them. Do the unglamorous operating work every day.
What made Miller’s point more than a slogan was its specificity. He described two recurring failure modes in value-based care: providers that negotiate weak contracts, and providers that have workable contracts but do not really understand or operate them. The basics he emphasized—knowing when patients are admitted, closing quality gaps, acting on available health-information-exchange data rather than waiting for perfect data, and aligning physician incentives—are things leaders can control now. His post-event note made the same point: clear plans, goals, and consistent blocking and tackling beat hand-wringing or waiting for policy to “fix” the market. He is right.
A badly run organization does not become well run because the payment system is flawed. A provider that does not understand its risk contract cannot blame every loss on national healthcare economics. Better information, physician engagement, contracting, care management, and operational competence can make an enormous difference. Operational excellence can improve performance inside an economic model; it cannot, by itself, make every contradiction in that model disappear. And this is where I think the healthcare conversation often stops one level too soon.
AI Can Change the Production Function
By production function, I mean something fairly simple: the mix of human labor, technology, facilities, and other inputs required to deliver care. We have spent years trying to improve healthcare economics substantially by changing how money and risk move around the system: fee-for-service, capitation, bundles, shared savings, downside risk, quality incentives, global budgets, directed payments, social-care funding.
These mechanisms can change behavior substantially. Some have also changed how and where care is delivered. Healthcare’s production function has changed before. Ambulatory surgery moved work out of hospitals. Team-based care redistributed professional tasks. Telemedicine moved encounters across geography. Hospital-at-home and remote monitoring have challenged site-of-service assumptions. Yet enormous portions of cognitive healthcare work remain tethered to scarce human professionals. A physician still has only so many hours. Specialists remain scarce. Much of care still requires a trained professional to gather information, synthesize it, interpret uncertainty, make a decision, communicate that decision, document it, and assume responsibility for what happens next.
Payment reform asks: who should pay for that work, how much should they pay, and who should bear the risk? Here’s a different question: what if the amount and kind of human work required to produce healthcare changes? That is why, near the end of the discussion, I asked about AI.
My question was deliberately a thought experiment, not a forecast. Suppose AI does not “replace the doctor.” Suppose instead that it becomes capable of performing 20%, 30%, or 40% of the work clinicians perform today, and that some of that work can happen in the home. What happens then to everything we had just spent the evening discussing: access, reimbursement, affordability, staffing, facility economics, and the allocation of risk when AI-enabled change becomes large enough to alter how care is produced?
History-taking, record synthesis, monitoring, routine follow-up, patient education, documentation, risk identification, triage, portions of diagnostic reasoning, care-plan preparation, these activities do not all have the same technical requirements, regulatory requirements, or need for professional judgment. AI will not become equally capable across all of them at once, nor should capability automatically confer authority.
And task automation does not translate mechanically into equivalent labor savings. If AI removes thirty minutes of work from a clinician’s day but creates twenty minutes of review, exception handling, documentation, or patient counseling, the productivity gain is quite different. A technology can complement scarce professionals rather than replace them. It can expand capacity without eliminating jobs. It can improve quality without lowering total spending.
These are not objections to the production-function argument. They are part of it.
Allen’s response captured another distinction well. There are at least three different gates: whether AI can do something, whether the regulatory system permits it to do that thing, and whether anyone will pay for AI to do that thing. Those gates will not open simultaneously, and they will not open uniformly across fifty states and multiple federal and state regulatory regimes.
AI also cannot answer the fundamentally political questions: Who should be covered? How much should society spend? How should resources be distributed? What obligations do we owe one another? Those remain human choices, but the institutional pathway is only half of what is about to happen.
Patients already have access to increasingly capable AI whether their health system has completed its governance review or not. Pew found this summer that 34% of U.S. adults report using AI chatbots for at least one of eight health or medical purposes it asked about; 25% use them to help figure out what is causing symptoms, 22% to learn more about a physician’s diagnosis, and 20% to help understand laboratory results. (Pew Research Center)
Patients can increasingly arrive having summarized years of records, interrogated longitudinal lab results, assembled months of wearable data, deeply researched symptoms, generated and ranked possible explanations against the available evidence, and prepared highly specific questions before the clinician has even reviewed the patient’s chart.
During the panel, I pushed that example toward an extreme: the patient arrives with the record, symptoms, and an AI-generated interpretation they regard as effectively settled, essentially asking the clinician to validate and operationalize the diagnosis, treatment plan, and referrals. The patient’s AI is becoming part of the clinical pathway whether the institution invited it or not.
There is an important second-order effect here too. A patient with better-organized information may save a clinician time. A patient arriving with five AI-generated diagnoses, sixty pages of synthesized records, wearable alerts, and expectations for action may create more work. Someone has to validate the claims, resolve contradictions, reject bad suggestions, manage incidental findings, communicate uncertainty, and assume responsibility. The patient’s AI becoming part of the pathway does not guarantee that the pathway becomes cheaper. It guarantees that the pathway changes.
Hospitals are approaching the same change from the other direction. Gianelli offered a concrete example from One Brooklyn Health. His organization had acquired an AI-powered system for continuous patient monitoring, 24 hours a day, 365 days a year, including risks such as falls and pressure injuries. That is not “AI replacing a clinician.” It is software taking on part of the observational workload that previously depended more heavily on continuous human attention.
If systems like that allow clinicians and staff to supervise more patients, intervene when needed rather than observe continuously, or redeploy scarce attention toward work requiring human judgment, the economic significance is not simply that one existing task became faster. The mix of human labor and technology required to produce the care has changed. If AI merely makes an existing workflow somewhat faster, we have an incremental productivity story. If it changes which tasks require scarce licensed labor, how many patients a professional can safely oversee, which tasks can occur outside a facility, and when human judgment must enter the process, we have production-model redesign.
However, it does not follow that healthcare spending automatically falls. Lowering the unit cost of producing something can increase demand for it. AI can generate utilization as readily as eliminate it. It can create new monitoring, diagnoses, administrative work, security problems, incidental findings, and low-value care. Clinical authority, liability, evidence, privacy, cybersecurity, and equity do not evaporate because the technology improves.
A cheaper way to produce a service only lowers system cost if the surrounding incentives allow that cheaper production model to replace or materially alter the expensive one. This last condition is why the reimbursement conversation matters.
Imagine an AI-enabled model that can safely resolve part of what today requires a specialist appointment. If the only economically viable pathway still routes nearly every patient through that specialist appointment, the technology may improve triage, throughput, preparation, or quality without realizing much of its potential system-level economic impact. Or imagine a hospital that deploys AI and remote monitoring well enough to prevent admissions and move care into the home, but depends financially on inpatient revenue and supplemental payments tied to the legacy delivery structure. The technology may work. The transformation may work. But the economics may still punish the institution for adopting it.
We are back to Arthur’s math.
Except something important has changed. Much of health-system reform has tried to design better economics around a care model that still depends heavily on scarce human labor and expensive physical infrastructure. AI gives us the possibility of changing those inputs themselves: how much clinician time a service requires, which work software can perform, what can happen outside a facility, and where human judgment actually needs to enter the process.
For leaders, investors, and founders, that should shift the question. Incremental efficiency can be tremendously valuable; there is nothing trivial about making a large recurring workflow materially cheaper or better.
But the most consequential opportunities may be those that go further: changing where professional scarcity enters the system, which pieces of care can be separated and recombined, and where software can safely augment or substitute for scarce professional labor without losing the judgment, accountability, safety, and calibrated trust healthcare requires.
For providers and payers, the implication is equally uncomfortable. If that production function changes materially, today’s reimbursement architecture will not remain untouched. Codes, contracts, professional roles, site-of-service economics, risk arrangements, staffing models, and eventually the economic meaning of a “visit” all become contestable.
None of that makes AI the solution to healthcare. Healthcare does not have one problem, so it is unlikely to have one solution. However, we cannot indefinitely solve healthcare’s arithmetic by changing only who pays whom. Payment reform can align incentives. Policy determines whom we cover and what we choose to fund. Better contracts matter. Blocking and tackling matters.
All of those things can make the system work better. But none of them, by themselves, creates another hour in a physician’s day. They do not make a scarce specialist less scarce. They do not eliminate the labor required to synthesize a record, monitor a patient, answer a routine question, or manage a follow-up. And they do not automatically make care cheaper to produce simply by moving financial risk from one organization to another.
This is the opening AI creates. If AI can safely take on some of that work; if it can let clinicians concentrate their time where human judgment matters most; if it can allow some care to move from expensive facilities into homes; and if it can help one scarce professional safely oversee more care than today’s model permits, then we are no longer only redistributing the same costs and constraints. We are changing them.
That brings me back to the line I wrote in my notebook halfway through the evening: The numbers don’t square.
AI may not solve healthcare, but for the first time in a truly meaningful way, AI just might allow us to change the arithmetic instead of merely rearranging it.
-Marc d. Paradis
About the Author: Marc d. Paradis’ professional journey is a fusion of academic rigor with real-world impact. He began his career over 30 years ago as an academic molecular neurobiologist, instilling in him a deep respect for critical thinking and the scientific method.
Transitioning into industry, he held leadership roles that bridged data and healthcare: as Vice President of Data Strategy at Northwell Health, Marc leveraged one of the world’s most diverse clinical data sets to drive patient-centered innovation via a $100M partnership with Aegis Ventures, launching multiple AI-centered startups; and as Vice President & Dean of Data Science University at Optum, he spearheaded the training of thousands of professionals in practical, product-centric AI, data-driven decision making, and ethical data practices. In each role, he fostered cultures of curiosity, critical thinking, and collaboration – precursors to the Constructive Inquiry ethos.
About SIYOM Consulting: Founded by Marc d. Paradis, SIYOM Consulting is a boutique advisory specializing in Data and AI Strategy for Healthcare and Life Sciences. We help health-system executives, pharma innovators and investors identify, evaluate and execute on high-value data and AI opportunities.
Source note for review: Speaker remarks are paraphrased from personal notes, not presented as speaker- and/or audio-verified quotations.