The Service Layer Is About to Become Software

For twenty years, US health systems have bought the same product under different brand names: a thin layer of software wrapped around a large room of people. Care-management BPOs. Discharge “concierge” vendors. Revenue-cycle shops with a portal. Referral networks that are really call centers. The SKU on the contract says platform. The P&L says labor.

That model was rational. It is also running out of time.

Artificial intelligence does not mainly threaten the electronic health record, or the claims engine, or the pharmacy benefit manager. It threatens the service layer that sat between those systems of record and a completed episode of care: the labor-priced work of chasing missing fields, dialing vendors, re-keying discharge instructions, and swearing the ride happened. Tech-enabled services had a window where wrapping labor in software looked like product. That window is closing.

Call it service as software: health systems stop buying labor wrapped in software and start buying completed, provable work from software. In patient logistics, that means the coordination bought by the seat today, from chasing a missing field to confirming a ride happened, becomes work that agents complete and prove, with people handling the exceptions.

What health systems actually buy

When a health system or a plan buys “patient logistics,” “care transitions,” or “post-acute coordination,” they are rarely buying a deterministic machine. They are buying exception handling.

The happy path is short. A discharge order exists. A destination exists. A payer rule exists. A transport vendor or home-health agency exists. In a clean world, software would compose those facts and emit a completed transition with proof.

The real world is incomplete data, conflicting payer rules, vendors who do not share status, and EHRs that expose events as documents rather than as state machines. So buyers purchase humans as the integration layer. The vendor’s application is often a case-management UI for those humans: queues, notes, phone logs, SLAs. Automation, where it exists, is rules and RPA on the edges.

This is not a moral failure. It is an engineering consequence. When interfaces are brittle and proof is ambiguous, labor is cheaper than building a reliable control plane. Tech-enabled services scaled because they could sell outcomes of effort (“we will manage your discharges”) without having to sell a closed-loop system.

Why the service layer existed

Three technical gaps made the service layer inevitable.

Incomplete state. Clinical and operational systems recorded what was documented, not what was true in the world. A discharge summary is not the same object as “patient left the building, destination confirmed, handoff accepted.” Without shared state, someone has to call.

Brittle composition. HL7, FHIR, proprietary EHRs, fax, and portal uploads do not compose into one workflow. They produce partial updates. Humans reconcile the partial updates. “Integration” became a permanent job description.

Ambiguous proof. Payment and quality programs asked whether care happened. The evidence was often narrative: a note, a claim, a vendor attestation. When proof is ambiguous, you staff people to argue with other people.

Tech-enabled services thrived in that gap. They productized the reconciliation. Their software made the labor measurable. Their sales deck showed dashboards. Their margin came from utilization of human time.

What is changing, technically

Three shifts arrive together.

Models that operate on structured work, not just text. Foundation models and tool-using agents can read order sets, eligibility responses, vendor rosters, and event streams; propose a next action; call an API; and write back a structured result. That is different from a chatbot that summarizes a chart. It is closer to a planner sitting on top of existing systems of record.

Interfaces that finally expose verbs. SMART on FHIR apps, eventing, and modern APIs turn “look up and dial” into “invoke and confirm.” Not everywhere, not cleanly, but enough that an agent can be given tools instead of a headset.

Proof becoming machine-checkable. GPS timestamps at pickup and dropoff, live credential checks, signed orders, immutable event logs. When proof is a data structure instead of a story, the reconciliation job shrinks. Payment can follow proof without a team re-litigating every trip.

Put those together and the expensive middle of many health-IT services becomes compressible. Not every exception disappears. The distribution of exceptions changes. The median case stops needing a person. The long tail still does. That is enough to break a labor-scaled P&L.

What collapses

The first thing that collapses is the labor-scaled P&L inside tech-enabled services.

If your product is “we staff nurses and logistics coordinators who live in our workflow tool,” then your defensibility was process knowledge and relationships, priced by the seat. AI agents with tool access attack process knowledge directly. Relationships remain, but they are not enough to defend a seat count.

That is a claim about where margin sits, not about the people doing the work. Inside health systems, case managers and transport coordinators live with the phone tree every day and are usually the first to want it gone. When agents take the routine case (ride booked, vendor accepted, status confirmed), those people move to the cases that need judgment: the transfer no vendor will take, the family that changes the plan at the curb, the payer rule that conflicts with the order. The margin that paid for outsourced seats of routine coordination moves to the systems that complete that work.

The second thing that collapses is the software-wrapping-labor SKU. Buyers have been trained to accept a hybrid: some licenses, many FTEs, a shared-savings story. As agents take the median case, the hybrid looks like overpaying for a UI. Procurement will ask a sharper question: what is completed without a human in the loop, and what is the unit of proof?

The third thing that collapses is opacity as a business model. Services businesses often hide how much of the work is phone and spreadsheet. Machine-verifiable completion makes that visible. Visible work gets competed.

Notice what is not on this list: brokers. A broker’s job, holding a network and a contract and answering for the ride, survives cheaper coordination. What changes is what a broker can prove. A broker that runs its trips on a shared completion record can show a plan or a state the same evidence it sees, trip by trip, instead of an attestation. Brokers can buy the completion and proof layer too.

Incumbents will not vanish overnight. They will try to rebrand as AI platforms while keeping the same labor graph. Some will succeed at a smaller scale. Many will discover that their customers can buy the control plane from someone else and keep their own network, their own vendors, their own staff where judgment still matters.

The limited window

Here is the timing argument, without theater.

For a short period, tech-enabled service vendors can use AI to raise their own margins: same contract, less labor per case, better SLAs. That is the salvage window. It is real. It is also temporary.

The lasting shift happens when buyers re-contract. Contracts today buy effort and access. Contracts tomorrow will buy completed transitions and proof objects. Once a health system’s RFPs and MSAs are written that way, the old SKU cannot be propped up with a better chatbot. You either operate a completion system or you are a staffing firm with a login screen.

That is the limited window: the years between “AI makes our outsourced coordination cheaper” and “we do not buy coordination by the seat for this class of work.” Vendors who spend the window extracting efficiency for themselves, without changing what the buyer purchases, will arrive at the rebid with the wrong product.

What buyers should demand instead

If you buy for a health system, a plan, or a broker that answers to one, change the unit of purchase.

Demand completion, not activity. A transition that reaches a terminal state with handoff accepted beats a queue of “in progress” tickets.

Demand machine-verifiable proof where the physics allow it: time, place, actor, credential, order. Narrative notes are for the cases that still need judgment, not for the median ride or referral.

Demand clear failure modes. Agents will fail. Good systems fail into a human queue with the state preserved. Bad systems fail into silent drift.

Demand composability. Keep your network. Keep your EHR. Keep your broker if it is working for you. Buy a rail that hands the proof back to you, visible in your own workflow and exportable when you ask, not a black box that holds your vendor relationships and your trip history.

Do not demand clinical-outcome miracles from logistics software. Mechanism and economics first. Throughput, completion rate, on-time performance, cost per completed transition. Outcome claims without a mechanism are how the last generation of services decks got signed.

What stays defensible when software is cheap to copy

There is an uncomfortable corollary. If agents make coordination cheap, they make software cheap too.

Peter Diamandis put it bluntly in a September 2026 article on X, “What Counts as a Moat in the Age of AI?” The classic moats, he writes, all “assumed something that is no longer true: that copying an existing business is expensive and slow.” His list of what no longer protects you: “Patents. Code. Features. Distribution. Everything that can be described can be copied, and AI can describe anything.” Switching costs, he argues, “evaporate when an AI agent does the work for you.” What survives, in his telling, is speed of reinvention and capital velocity, with brand holding only at the high end.

His test is the useful part. Ask of everything you think protects you: “could a frontier model plus a motivated team rebuild your product or service in ninety days? Anything that fails is not a moat. It’s a head start, and you should price it that way.”

Apply that honestly to patient logistics. A request form, a dispatch board, a vendor portal, a dashboard: all of it fails the test. A good team with a frontier model could rebuild the interface in a quarter. That includes ours. The UI is a head start.

What does not rebuild in ninety days lives outside the code.

Closed-loop proof of completed transports. A completion record only means something when the order, the credentialed vendor, the status events, and the settlement are joined on the same request, trip after trip. That record accumulates from operation, not from a sprint.

A live vendor network. Contracted, credentialed providers who actually accept broadcasts in a given zone on a Saturday night are relationships and compliance records, kept current. A model can write a vendor onboarding screen. It cannot make an ambulance company in a given county accept the trip.

Embedded EHR and payer workflows. An app that is live inside a health system’s EHR, cleared by its IT and security teams, configured to its protocols, and checked against a plan’s benefit rules took months of other people’s approvals. Copying the code does not copy the approvals.

The second idea from Diamandis matters even more here. Citing the “Coasean Singularity” paper, he argues that when agents “search, compare, negotiate and transact for us at near-zero cost, the customer relationship stops belonging to whoever owns the storefront and starts belonging to whoever the agent trusts.” The paper itself is more careful. In an NBER working paper, researchers at MIT, Harvard, and Boston University argue that agents lower search, communication, and contracting costs, and that “by lowering the costs of preference elicitation, contract enforcement, and identity verification, agents expand the feasible set of market designs.” They also flag new frictions, including congestion and price obfuscation, and treat the net welfare effect as an open empirical question.

For patient logistics the implication is concrete. When a discharge planner’s agent, or a plan’s agent, chooses the path for a ride, it will not be charmed by a UI. It will prefer the path where eligibility, credentials, and completion can be checked by a machine. For an agent, trust is verifiable state. Whoever holds the proof holds the relationship.

What this is not

This is not an argument that hospitals will run without people. Clinical judgment, edge-case ethics, and hard conversations stay human. So does the coordinator who knows which vendor actually shows up on a holiday weekend.

This is not an argument that every BPO dies in twelve months. Capital, contracts, and switching costs are slow.

It is an argument about where the margin sits. For a generation, margin sat in the service layer because software could not close the loop. AI plus better interfaces plus checkable proof moves margin into systems that complete work. Tech-enabled services that remain mostly services are living on a depreciating asset: the gap that made their labor necessary.

The buyers who notice early will rewrite the SKU. The vendors who notice early will stop selling coordination by the seat and start selling completed, provable transitions. Everyone else will keep polishing a dashboard on top of a call center, and call it AI, until the contract comes up and the question has changed.