Oct 8, 2026
By Nima Roohi Sefidmazgi, Blooming Health CEO and Co-Founder, and Dr. Chethan Sathya MD, Blooming Health Chief Public Health Officer
For a decade, healthcare's AI conversation has centered on knowledge: better diagnoses, better predictions, better decision support. That focus made sense when the tools were new and the questions were about whether a model could match a clinician's judgment. But it has quietly become the wrong question.
Healthcare does not fail for lack of medical knowledge. It fails because nobody owns the hundreds of operational steps between a diagnosis and an outcome.
A patient needs a specialist, but the referral never becomes an appointment. A Medicaid member remains eligible for coverage, but misses a renewal deadline. A health system identifies a transportation or food-access need, but the referral is never completed. A prior authorization is clinically appropriate, but stalls because documentation is missing. A patient is discharged with a plan that makes sense on paper, yet no one reaches them when the plan begins to break.
These are not edge cases, they are the daily operating system of American healthcare.
This is the distinction we think of as Software 2.0 versus Software 3.0 in healthcare. Software 2.0, the enterprise systems of the last two decades, gave us systems of record, dashboards, task queues, care-management platforms, CRMs, patient-engagement tools, and portals. Even the AI layered on top of that generation of software, the copilots and predictive models, is still fundamentally in the business of producing better information for a human to act on. It tells you the referral is open. It flags the care gap. It drafts the note faster. What it does not do is close the referral, schedule the appointment, or confirm the outcome.
Software 3.0 is the shift from software that informs work to software that performs it. The important change is not how the software is built. It is what we expect it to do.
The Medicaid unwinding made the problem unusually visible. More than 25 million people were unenrolled as states resumed eligibility reviews, and 69% of those unenrollments were for procedural or paperwork reasons rather than a determination of ineligibility. The lesson is uncomfortable: access to care can depend as much on whether a workflow is completed as on whether the right benefit or treatment exists. [1]
Software organized human labor; humans remained the execution layer. A care manager still had to call. A scheduler still had to schedule. A navigator still had to find the resource, make the referral, chase the patient, and document the outcome. That held true even as AI got layered into these systems, the tools got smarter, but completion still ran through a person.
Artificial intelligence creates the possibility of changing that relationship.
We think of this shift as Software 3.0 in healthcare: software moving from primarily assisting work to increasingly executing it. The important change is not simply how software is programmed. It is what we expect software to do. [4]
In this model, software is no longer only a tool a worker uses. It becomes an operating layer that can perform substantial portions of a workflow itself: understand a goal, interact with a person, retrieve information, use tools, take an action, observe what happened, and continue until the task is completed or a human needs to intervene.
That distinction matters more than whether an application has an AI copilot.
A copilot helps a care manager write a note faster. An executing system can identify a patient with an open care gap, contact the patient in the appropriate language and channel, schedule the appointment, send reminders, reschedule after a cancellation, arrange transportation when needed, confirm completion, and escalate the exceptions it cannot safely resolve.
The unit of value begins to shift from software usage to work completed and outcome generated. That is a much bigger change in enterprise technology. That is a direct alignment between technology and value based care principles.
The economic evidence for AI today is still more modest than some of the rhetoric surrounding it. In a large field study of more than 5,000 customer-support agents, access to generative AI increased productivity by about 15% on average, with the largest gains among less-experienced workers. That is meaningful, but it is still primarily evidence about AI assisting people—not software autonomously completing complex workflows. [2]
The next step is different. When an AI system is connected to communication channels, enterprise data, and action-taking tools, it can move from recommending the next action to taking the next action.
In healthcare, that could mean not merely reminding a staff member that a referral is incomplete, but pursuing that referral to resolution. It could mean not merely drafting a prior-authorization packet, but gathering required information, submitting it through approved workflows, monitoring the response, and routing an exception to a human.
The most valuable near-term opportunities are not autonomous diagnosis or unsupervised clinical decision-making. They are high-volume, rules-constrained, language-heavy operational workflows where completion can be measured and escalation paths are clear.
For Medicaid continuity, software can identify members approaching renewal, explain requirements, collect missing documents, route them through the appropriate process, and re-engage until coverage is renewed or a human must step in. A use case the whole country can benefit from immediately in light of recent policy changes from CMS.
For care-gap closure, it can move from generating outreach lists to actually reaching patients, scheduling care, handling predictable barriers, and verifying completion.
For social-care navigation, it can move beyond screening and referral to arranging the service, completing a warm handoff, and checking whether the need was resolved.
For prior authorization, it can coordinate documentation requirements, submission, status monitoring, and exception handling. CMS is already making portions of this workflow more machine-readable: its interoperability rule requires affected payers to implement FHIR-based prior-authorization APIs beginning in 2027, while operational provisions require faster decision timelines and specific reasons for denials. [6]
None of this means humans disappear.
In fact, the human role becomes more important where judgment, empathy, clinical expertise, accountability, or trust is required. The change is that scarce people no longer need to carry every routine step of every case. They can supervise systems, manage exceptions, resolve ambiguity, and intervene when the consequences justify human attention.
That distinction matters in healthcare, where “automation” is too easily interpreted as removing humans. The better architecture is human-supervised execution: automate what is routine, observable, and reversible; escalate what is ambiguous, high-consequence, or relational.
And the hardest part will not be the language model. A 2025 field guide from Mass General Brigham researchers describing the deployment of an AI agent found that less than 20% of the implementation effort went toward prompt engineering and model development. More than 80% went toward the harder work of data integration, validation, demonstrating economic value, managing drift, and governance. [3]
That is an important corrective for healthcare executives. The durable capability is not simply access to an AI model. It is the operational architecture around it.
Healthcare organizations therefore should stop asking only, “Where can we add AI?”
A better question is: “Which workflows are we willing to let software increasingly own?”
That requires a different procurement framework.
Start with workflows that have a clear definition of completion. “Improve engagement” is vague. “Schedule the overdue mammogram and verify attendance” is concrete.
Require actionability, not just intelligence. A system that can summarize a chart but cannot communicate, schedule, retrieve a document, initiate an approved transaction, or update a system of record remains an assistant.
Design escalation before autonomy. Organizations should know which actions the system may take, which require confirmation, and which conditions must immediately route to a person.
Measure outcomes instead of activity. The metrics that matter should increasingly be successful renewals, closed care gaps, completed referrals, resolved prior authorizations, avoided no-shows, and cost per resolved case, not log-ins, clicks, or messages sent.
Govern these systems as operational actors. HIPAA obligations do not disappear because software is doing more of the work. Appropriate access controls, minimum-necessary data handling where applicable, audibility, monitoring, and accountability remain essential. NIST's AI Risk Management Framework similarly emphasizes governance, measurement, and risk management throughout an AI system's lifecycle rather than treating approval as a one-time event. [7]
This shift could also change the economics of enterprise software. For decades, enterprise software has largely been sold as access: licenses, seats, modules, and implementations. If software increasingly performs measurable work, buyers will reasonably begin asking vendors to prove and price value against that work. Some categories may move toward payment per completed task, managed workflow, or outcome rather than simply payment for access to another application.
For health systems and health plans, that should fundamentally change the technology-investment question and be more aligned with value based care operational economics.
Organizations have spent heavily on applications designed to make employees more productive, then spent heavily again on the people required to operate those applications and manually close the gaps between them. AI creates an opportunity to break that pattern, but only if leaders resist the temptation to bolt copilots onto every existing system and call it transformation.
The more consequential opportunity is to redesign the operating model.
Imagine a care-management organization in which software executes much of the routine outreach, scheduling, navigation, documentation gathering, and follow-through, while nurses, social workers, community health workers, and care managers focus their expertise on the patients who genuinely require it. The objective is not to replace those professionals. It is to stop wasting their expertise on work that a well-governed system can reliably complete.
Healthcare has spent decades digitizing information. The next decade should be about digitizing execution.
The organizations that understand this shift will not ask only which AI features their existing vendors are adding. They will ask a more fundamental question:
For every important workflow between identifying a need and achieving an outcome, who, or what, actually owns completion?
That is the promise of Software 3.0 in healthcare. Software that does not simply help us manage the work. Software that helps us get the work done.
[1] [10] Medicaid/CHIP Monthly Enrollment Tracker | KFF
https://www.kff.org/medicaid/medicaid-enrollment-tracker/?utm_source=chatgpt.com
[2] Generative AI at Work | NBER
https://www.nber.org/papers/w31161?utm_source=chatgpt.com
[3] Beyond the Algorithm: A Field Guide to Deploying AI Agents in Clinical Practice
https://arxiv.org/abs/2509.26153?utm_source=chatgpt.com
[4] Launch YC: Nuvi – The AI Agent Builder for Software 3.0 | Y Combinator
[5] Article Types | NEJM Author Center
https://www.nejm.org/author-center/article-types?utm_source=chatgpt.com
[6] Prior Authorization API | CMS
[7] Are business associates required to restrict their uses and disclosures to the minimum necessary? May a covered entity reasonably rely on a request from a covered entity's business associate as the minimum necessary? | HHS.gov
[8] Presubmission Inquiries | NEJM Author Center
https://www.nejm.org/author-center/presubmissioninquiry?utm_source=chatgpt.com
[9] NEJM Author Center
https://www.nejm.org/author-center/home?utm_source=chatgpt.com
