A three-part argument on chronic disease, AI, and the architecture of change in healthcare systems built around the wrong event
A reframing of care trajectories in chronic disease — hard transitions, soft transitions, and the structural blindness of an event-based health economy
How AI can serve the soft transition — and what it means for payers, health systems, and the economics of care
On change vehicles, stuck systems, and where the soft transition might actually be managed first — concluding the soft drift series
A reframing of care trajectories in chronic disease
The dominant framing of preventative care — as a set of clinical activities intended to forestall an acute event — has produced a healthcare economy organised entirely around the acute event itself. The event is what is paid for, measured, audited, and reported. Everything before it is rendered as cost, and everything after it as recovery.
But patients do not live their illness in the binary of well and unwell. They live it as a sequence of transitions, some hard and some soft.
Consider two ways a health system experiences the same patient:
Hard transitions are discrete in time, highly visible, codeable, and exceptionally well served by our acute apparatus. The myocardial infarction, the sudden collapse, the first urgent hospitalisation for an interstitial lung disease (ILD) exacerbation, the formal diagnosis of cancer. Hard transitions have clean edges: the patient was on one side of the line yesterday and is on the other side today, and both sides possess an ICD-10 code. The system is highly competent at mobilising around these edges.
Soft transitions are gradual, distributed, multi-signal, and structurally unrewarded. They represent the slow loss of exercise tolerance in early respiratory disease, where a patient quietly recalibrates their life around their breathlessness long before pulmonary function testing drops below a reimbursement threshold. They are the quiet drift into frailty marked by changes in gait speed and grip strength, the subclinical rise of cardiometabolic markers, the slow erosion of medication adherence, and the gradual narrowing of a patient's social world.
Healthcare systems have learned to detect and pay for hard transitions with immense precision. In the same period, they have become structurally incapable of recognising soft transitions, let alone working with them. The language of prevention is part of the problem: it presupposes the hard event as the anchor of value and reduces everything earlier to a probabilistic averting of that event.
The contemporary vocabulary of prevention is largely descended from Geoffrey Rose's seminal work, which split public health into the high-risk strategy and the population strategy. Rose's central insight — that a large number of people at a modest risk contribute more cases of disease than a small number at extreme risk — remains a pillar of twentieth-century epidemiology.
However, Rose's framing was optimised for a specific kind of problem: a population-distributed risk factor linked to a single, well-characterised acute outcome. Blood pressure and stroke; cholesterol and ischaemic heart disease. In these examples, the population sits on a linear risk curve, the outcome is a discrete event, and interventions are easily calculated by events averted per person treated.
The modern chronic disease burden looks nothing like Rose's worked examples. It is characterised by multimorbidity, frailty, progressive organ dysfunction, and the long, non-linear sequelae of conditions like heart failure, chronic kidney disease, and dementia.
| Rose's original framing | Contemporary chronic disease |
|---|---|
| Single, well-characterised acute outcome | Lifelong trajectory — no singular outcome |
| Linear risk curve across population | Constellation of slowly drifting parameters |
| Discrete events amenable to counting | Non-linear, emergent, multi-morbid sequelae |
| Intervention calculated as events averted | Sustained relational dialogue over years |
When a disease does not generate a clean event, current payment structures either invent an arbitrary proxy threshold — a target HbA1c, a specific spirometry value — or withdraw clinical attention altogether. Both responses distort clinical reality.
Across the OECD, payment systems remain overwhelmingly organised around discrete, codeable, billable episodes. Fee-for-service rewards the encounter; activity-based funding rewards the admission; Diagnosis-Related Group (DRG) systems reward the acute episode. Even value-based purchasing frameworks routinely adjust the price of care based on metrics defined entirely around hard events: 30-day readmissions, complication rates, and mortality.
The value in value-based care is, almost everywhere, merely the value of better-managed hard transitions. Capitation and population-based contracts offer a theoretical escape — yet in practice, because the regulatory machinery used to police these contracts relies on hard-transition metrics, even capitated providers end up steering by the lights of the old system.
The work that matters most for chronic disease — noticing a drop in exercise tolerance that has not yet crossed a protocolised threshold, or adjusting treatment based on a trajectory rather than a single baseline — has no billing code. Frameworks like the Quality and Outcomes Framework (QOF) in the UK or the Merit-based Incentive Payment System (MIPS) in the US reward what can be coded, which is almost always a hard transition or a hard threshold.
Once a hard transition occurs, the system mobilises disproportionately. Acute care is highly visible, well staffed, and well reimbursed. The patient with an acute IPF exacerbation is suddenly visible: scanned, monitored, and discussed in multidisciplinary meetings. The contrast is striking: the same patient was visible for months via declining functional reserves, and the system was blind; they experienced one catastrophic day in a hospital bed, and the system refused to let them out of its sight.
Patients quickly learn these rules. They learn that the ten-minute consultation is structured for triage, and that the slow, accumulating change in how they feel does not fit into a recognised clinical category. They calibrate by waiting until things are bad enough — meaning, until they cross a threshold the system recognises. By that point, the soft transition has completed its course, presenting to the clinician as a hard, expensive, and often irreversible event.
For clinicians, participating in a system that forces a ten-minute triage encounter upon a complex, multimorbid patient who requires forty minutes of unhurried continuity is a profound source of chronic burnout and moral injury.
The hard/soft distinction draws naturally on the broader complexity-science account of healthcare. Using Snowden's Cynefin framework, we can separate problems into the complicated domain (clear cause-and-effect chains that yield to expert protocols) and the complex domain (entangled, emergent patterns where cause-and-effect can only be understood retrospectively).
| Complicated domain | Complex domain |
|---|---|
| Hard transitions (e.g., myocardial infarction) | Soft transitions (e.g., frailty drift) |
| Cause-and-effect is linear | Multivariable, emergent interactions |
| Mode: identify pattern → protocol | Mode: probe → sense → respond |
| Governed by ICD codes and thresholds | Governed by warm data and relational continuity |
Hard transitions are typically complicated, in the Cynefin sense. An acute myocardial infarction yields to a known protocol; the expert applies the guideline, and the mechanism works. Soft transitions are characteristically complex. A patient's slow descent into frailty, or the subtle decompensation of multiple co-morbidities, emerges from the non-linear interaction of biological, psychological, and social factors.
This domain cannot be controlled by acting on a single variable or a static protocol. It demands what Snowden terms a probe–sense–respond approach — the precise mode utilised by an experienced general practitioner or a continuous care team who knows what they do not yet know.
This approach requires what Nora Bateson defines as warm data: contextual, relational, multi-source information that cannot be flattened into an isolated variable. Warm data is precisely what is discarded when a healthcare system insists that only what can be conventionally coded counts. The category error of modern healthcare design has been to treat the complex domain of soft transitions as if it were merely a complicated protocol problem.
The structural challenge of managing soft transitions is not an insurmountable theoretical block; it is actively being solved in pockets of clinical excellence where systems are intentionally redesigned around complexity. The Sussex Kidney Unit provides a clear operational illustration.
When a patient's peritoneal dialysis begins to fail, they enter a classic soft transition. The decline is gradual, continuous, and multi-signal. In a traditional, fractionated system, the soft drift goes uncoordinated and the system misses the accumulating signals of decline until a catastrophic hard event occurs.
The Sussex Kidney Unit counteracted this category error by merging their PD and Home Haemodialysis services into an integrated home unit. This structural redesign allowed the care team to pool warm data and engage in a continuous probe–sense–respond mode. By routinely tracking early soft signals — and proactively asking long-horizon screening questions such as "Would you be surprised if this patient transferred to HHD in six months?" — they created a protective buffer around the patient's physiological reserve.
The operational results are unambiguous. Patients demonstrating early, subclinical signs of inadequate clearance were identified early enough to allow planned vascular access placement (AVF/AVG). This bypassed the need for emergency CVC lines entirely, reduced modality-transition hospitalisations to zero in compliant tracks, and improved long-term patient survival. By dissolving administrative boundaries between distinct therapies, the Sussex model transformed a chaotic, emergency-driven event cascade into an organised, continuous, and deeply relational navigation of a complex illness path.
If a healthcare system is to scale this type of success and take soft transitions seriously, it must overhaul its operational and administrative infrastructure across four domains.
Relational continuity — being known over time by the same clinician or team — is the single most efficient mechanism for detecting soft transitions. It must move from a sentimental virtue to a hard, measured, and paid-for outcome, held in the same administrative esteem as 30-day readmission metrics.
Clinical decision support tools must transition from cross-sectional thresholds to longitudinal trajectories. A patient whose HbA1c has climbed from 5.7% to 6.3% over twenty-four months represents a completely different pathophysiological state than a patient whose HbA1c has sat stably at 6.3% for a decade, yet current guidelines treat them identically.
The patient is the most sensitive sensor of their own soft transitions. Patient-reported outcome measures, when collected longitudinally and analysed as a moving trajectory rather than an episodic snapshot at admission or discharge, can capture functional decline long before a laboratory biomarker surfaces a signal.
The ultimate barrier to reform is not conceptual, but fiscal. Payment models must shift to long-horizon, capitated, risk-adjusted contracts that reward a clinical network for keeping a defined cohort on the favourable side of soft-transition trajectories. Until the metric infrastructure changes, even well-intentioned funding reforms will inevitably regress to the mean.
The traditional vocabulary of prevention has won historic victories against single-source risks. But applied to the contemporary landscape of slow, distributed, and multi-morbid chronic illness, it hides more than it reveals. By anchoring the concept of value to a coded event, it builds a system that only sees, funds, and treats events — leaving the long, invisible transitions where patients actually live entirely unmanaged.
Naming this structural mismatch is the vital first step toward designing a healthcare system capable of navigating complexity. The Sussex model demonstrates that the alternative is not merely theoretical.
How AI can serve the soft transition — and what it means for payers, health systems, and the economics of care
The first article in this series argued that healthcare systems are structurally blind to soft transitions — the gradual, multi-signal drifts that precede most of the catastrophic events we pay so dearly to manage. AI is frequently invoked as the solution. Before accepting that claim, we need to ask a more precise question: which kind of AI, doing what, in whose hands, and governed by what values?
The risk is not that AI fails to detect soft transitions. The risk is that it detects them brilliantly and then delivers that intelligence into the same broken system that failed to act on them in the first place — generating alerts nobody has time to read, flags that reinforce threshold-thinking, and scores that replace clinical judgement with algorithmic confidence. Done poorly, AI will automate the event-based cascade, not transcend it.
Done well, AI can become something genuinely new in the architecture of care: a form of institutional memory that holds the thread of a patient's trajectory over time, surfaces the slow signals that a fragmented system cannot hold, and creates the conditions for a human clinician to attend to complexity rather than fight through administrative noise to reach it.
This article sets out what that distinction looks like in practice — and then examines the implications for payers and health systems, where the most disruptive and underappreciated innovation is now becoming visible.
Any honest account of AI in healthcare must begin with an acknowledgement that is frequently avoided in commercial discourse: AI does not understand a patient. It identifies statistical patterns in data that has been collected, structured, and labelled by humans who did understand — or who were trying to. The model learns from the past. It cannot feel the texture of what the patient describes, notice the shift in tone that signals resignation rather than recovery, or hold the relational context that makes one clinical decision appropriate for this patient and wrong for the next.
Nora Bateson's concept of warm data — the contextual, relational, multi-source information that cannot be flattened into an isolated variable — is precisely what AI cannot generate. It can process warm data, if it has been captured. It cannot create the conditions under which warm data emerges. That is a human function. It requires unhurried, continuous, attentive relationship. The clinician who has followed a patient for three years holds warm data that no algorithm can replicate.
The most dangerous deployment of AI in this context would be one that replaces the relational work of a clinician with an algorithmic score — not because the score is wrong, but because the score is a thin substitute for the kind of knowing that actually manages soft transitions. It would feel like progress while dismantling the very capacity it claimed to support.
This is not an argument against AI. It is an argument for being precise about what AI should be asked to do, and what it should be designed to protect.
The case for AI in this domain is not speculative. It rests on a set of identifiable functions where computational capacity is genuinely superior to unaided human cognition — and where that superiority can be directed at the soft-transition problem specifically.
A clinician seeing a patient for twelve minutes once every six months cannot integrate the full longitudinal trajectory of that patient's functional status, medication behaviour, social context, and physiological markers. A well-designed AI can. When deployed on structured and semi-structured longitudinal data — pharmacy claims, wearable signals, PRO submissions, lab trends — AI can surface the slowly-moving signal that no individual encounter could reveal. It can detect that the patient's gait speed has declined 12% over eight months, that their pharmacy refill intervals have lengthened, and that their self-reported breathlessness scores have climbed — and flag this constellation before any single variable crosses a reimbursement threshold.
This is not replacing clinical judgement. This is giving clinical judgement something it currently lacks: an integrated, longitudinal view of a patient who has been passing through many different clinical hands.
One of the least-discussed costs of the current system is the volume of administrative, documentation, and coordination work that consumes clinician time and attention. A British GP spends an estimated 40% of their consultation time on documentation. A US hospitalist completes six hours of paperwork for every four hours of patient contact. This is not merely inefficiency — it is a systematic depletion of the relational capacity that complex care requires.
AI that takes on documentation, triage sorting, routine protocol adherence checking, and appointment coordination does not automate care. It liberates the time and cognitive space in which genuinely complex, relational care becomes possible. The dividend of AI-assisted administration is clinician presence — and clinician presence is the mechanism by which soft transitions get noticed.
Wearables, remote patient monitoring platforms, and ambient sensor systems now generate continuous streams of physiological and behavioural data. The challenge is not collection — it is signal extraction from noise, and the translation of signal into intelligible, actionable information at the right moment for the right clinician. AI is well-suited to this filtering and contextualisation role, provided the outputs are designed to augment a clinician's probe-sense-respond capacity rather than trigger a protocol-driven alert chain that reproduces event-based logic in a new medium.
The output of an AI monitoring system should not be an alert. It should be a question — framed in terms the clinician can act on: "This patient's trajectory has diverged from their own historical baseline. Would you like to look?" That is a probe. An alert is a threshold crossed. The distinction is not cosmetic.
Conversational AI — carefully scoped and designed — can serve as a continuous low-burden sensing layer between clinical encounters. A patient who would not call their GP about a slow change in breathlessness might mention it to an AI companion that checks in regularly, asks open questions, and logs the qualitative texture of the response. The AI does not interpret or act. It listens, records, and — when it detects a meaningful pattern — it creates the occasion for a human clinician to follow up.
This is proximity care at scale. It does not replace the clinician relationship. It fills the vast, currently unmonitored space between appointments where soft transitions actually happen.
These are not aspirational values. They are functional requirements. An AI system that violates them will, in the long run, degrade the very clinical capacity it claimed to augment.
Every AI function should be evaluated by whether it strengthens or weakens the longitudinal clinician-patient relationship. If it creates efficiency that erodes relational continuity, the efficiency is not worth its price.
AI outputs should raise questions and open conversations, not trigger automatic responses. A system that generates autonomous action — escalating directly to a care pathway without human interpretation — has crossed from augmentation into substitution.
Visualise how a patient is moving, not just where they are. A direction of travel communicates clinical meaning that a single value cannot. Tools that show only current status reproduce the category error of the existing system in new clothing.
Clinicians must be able to understand, interrogate, and override AI outputs. Black-box models that produce a score without explanation are inappropriate in this context. The clinician's sceptical intelligence is part of the system, not an obstacle to it.
AI systems in healthcare are routinely validated on model performance (AUC, sensitivity, specificity) and almost never evaluated on whether deployment changed clinical behaviour or patient outcomes. The latter is the only metric that matters.
The goal is not an AI that manages soft transitions. The goal is a clinician who can manage soft transitions, supported by an AI that makes that possible at scale.
AI in healthcare is not inherently complexity-aware or complexity-blind. It depends entirely on what it is asked to do. The following table makes the contrast concrete.
| Function | Complexity-blind deployment | Complexity-aware deployment |
|---|---|---|
| Risk stratification | Score-based triage; triggers protocol if score exceeds threshold | Trajectory-based flag; invites clinician to probe if direction-of-travel changes |
| Remote monitoring | Alert fired when parameter crosses preset limit; patient called only then | Continuous trajectory monitored; clinician receives contextualised weekly digest, acts relationally |
| Documentation support | AI generates note; clinician validates; relational time unchanged | AI handles full documentation; clinician uses recovered time for unhurried patient conversation |
| Patient engagement | Chatbot delivers protocol-based symptom check; escalates to triage if alarm triggered | Conversational AI listens, records qualitative texture, creates occasion for clinician follow-up |
| PROM collection | Administered episodically at appointment; snapshot compared to population norm | Collected continuously; analysed as individual trajectory; deviation from personal baseline triggers inquiry |
| Multimorbidity management | Condition-specific AI modules fire independently; no integration of signals | Unified longitudinal AI integrates signals across conditions; surfaces emergent interactions |
Healthcare AI discourse is dominated by clinical settings — the hospital, the clinic, the diagnostic imaging suite. The more consequential frontier is the payer. Insurers, NHS commissioners, integrated care systems, employer health plan sponsors, and CMS innovation models all sit at the point where financial incentives are set. If AI reaches the payer before it reaches the clinician — or reaches the payer in a form the clinician cannot see — the result is surveillance infrastructure that serves reimbursement optimisation rather than care quality. If it reaches the payer in the right form, it has the potential to restructure the financial incentives that have made the soft transition systematically invisible.
Current VBC frameworks reward event avoidance — measured by claims data, structured as binary outcomes. The emerging frontier is contracts indexed to longitudinal patient trajectories: risk-adjusted, PROM-informed, and evaluated over rolling 24-month windows rather than 30-day episodes. AI makes this tractable by integrating the continuous data streams that trajectory measurement requires.
Re-insurers and stop-loss carriers are beginning to price AI-detected soft-transition risk into catastrophic coverage products. The logic is straightforward: a payer that can demonstrate its clinical network is actively managing the soft drift will incur fewer catastrophic events. AI-derived trajectory scores become the basis for dynamic premium adjustment — rewarding systems that invest in relational, longitudinal care.
Self-insured employers in the US — who carry catastrophic risk directly — are among the most sophisticated early adopters of trajectory-based AI. Products now offer AI-guided navigation that identifies employees drifting toward high-cost events before they arrive. The employer pays for proactive navigation; the clinical intervention is funded by the savings on the averted event.
CMS's HCC-based risk adjustment is a snapshot model — it codes what has happened, not what is happening. Several CMS Innovation Center models are now piloting AI-derived prospective risk scores that incorporate trajectory data: not what diagnoses a patient carries, but how their functional status is moving.
Integrated Care Boards in England are under structural pressure to demonstrate population health management that goes beyond elective backlog and urgent care performance. Several ICBs are piloting AI platforms that segment their populations not by diagnostic category but by trajectory risk: who is drifting toward complexity, at what rate, and with what intervention options still available.
Pharma companies with progressive chronic disease portfolios — ILD, heart failure, CKD — are increasingly approaching payers as data partners rather than adversaries. The manufacturer provides AI-derived trajectory monitoring infrastructure; the payer provides claims and encounter data; and the resulting signal is used to identify patients at risk of disease progression before the costly exacerbation that triggers a treatment decision.
Each of these innovations carries a version of the same risk: that AI-derived trajectory data is used to restrict coverage, deny claims, or rate-up premiums for patients identified as high-drift rather than to invest in the longitudinal care that would change that trajectory. The governance question — who controls the model, who sees the output, and what decisions it is permitted to inform — is not a secondary concern. It is the central design question.
The most fundamental structural problem in healthcare AI is not technical. It is that the payer and the clinician are currently looking at the patient through different instruments, at different time horizons, and with different incentives for what they find. The payer looks at claims data: periodic, binary, coded. The clinician looks at the patient: continuous, qualitative, relational. When AI is deployed into this fragmented architecture, it tends to reinforce whichever instrument it has been trained on.
The opportunity — and it is a genuine one — is to use AI as the integrating layer that brings the payer and the clinician into the same longitudinal view of the same patient. When the commissioner's risk model and the clinician's care model run on the same trajectory data, indexed to the same patient, evaluated over the same time horizon, the incentive gradient changes. The payer no longer benefits from ignoring the soft drift; the soft drift is now on the balance sheet, moving.
The payer who can see the soft transition and pays for it to be managed will, over a ten-year horizon, systematically outperform the payer who pays only for the acute event. This is not altruism. It is the arithmetic of chronic disease.
For pharmaceutical companies with progressive-disease portfolios, this shift has a specific strategic implication. The conversation about drug value is currently dominated by QALY calculations indexed to clinical trial endpoints — mortality, hospitalisation, forced vital capacity. These are hard-transition metrics. A world in which payers and health systems begin to manage and measure soft transitions creates a new category of outcome that the right therapy — started earlier, maintained consistently, monitored continuously — can demonstrably affect.
That is both a commercial opportunity and an evidentiary challenge. Trials designed around hard-transition endpoints will not generate the longitudinal trajectory data that a trajectory-sensitive payer needs. The implication for trial design, real-world evidence strategy, and patient registry infrastructure is substantial.
It is quieter than the discourse suggests. It does not produce dramatic diagnostic breakthroughs or autonomous clinical decisions. It produces a clinician who has time to listen, a payer who can see the drift before it becomes the event, and a patient whose slow accumulating change is held by a system that is actually designed to notice it.
The technology is largely available. The institutional design, the payment architecture, and the governance frameworks are not. Those are the problems worth solving — and they are, in the end, problems about what kind of healthcare system we are willing to build and pay for. AI will reflect that choice, not make it for us.
On change vehicles, stuck systems, and where the soft transition might actually be managed first
The first two articles in this series named a structural problem — health systems built around the hard transition, blind to the soft drift — and described the conditions under which AI could serve the soft transition rather than entrench the broken architecture. The harder question now is the change question: how does any of this actually happen?
It is the question that is most often answered with the most confidence and the least precision. Change programmes, transformation strategies, ten-year plans, system redesigns, value-based commissioning frameworks — the rhetoric of healthcare reform is enormous, and the actual movement of incentives, behaviours, and outcomes is small. We have been changing healthcare systems for forty years. The patient who quietly recalibrates their life around their breathlessness, long before any threshold is crossed, is still invisible to the system that pays for their care.
That gap — between the reform discourse and the reform reality — is not an accident. It is structural. And understanding why is the precondition for any honest account of how change might happen here.
It is worth being specific about what a conventional consulting approach to this problem would look like, because the conventional approach is the default — and the default is what will fail.
A McKinsey, Bain, or Boston Consulting engagement on managing the soft drift would, with disciplined competence, produce: a stakeholder map, a target operating model, a roadmap with phased milestones, a maturity matrix, a benefits realisation framework, a capability assessment, a list of pilots in three test sites, KPIs anchored to measurable proxies, a steering committee structure, and a governance memo. It would be excellent work of its kind. It would not change anything.
The reason is the category error identified in the first article in this series. The conventional consulting approach treats system transformation as a complicated problem — one with a known solution that can be implemented through expert analysis, planning, and sequenced execution. The soft-transition problem is complex. There is no known solution to implement. There is no off-the-shelf target operating model for a health system that can hold relational continuity at scale, finance trajectory-based outcomes, and integrate longitudinal data across fragmented payers and providers — because no health system in the world has yet built one.
The complicated-tools approach produces beautiful programme architecture that the system politely receives, files, and ignores. Not because the leaders are obstructive, but because the programme is designed for a world in which the answer is known. Here, the answer has to be discovered — and that requires a fundamentally different kind of change vehicle.
What is needed is not a transformation programme. What is needed is a portfolio of safe-to-fail experiments — Snowden's term — running in the systems most able to host them, with the resources, governance, and patience to learn from what emerges. The early Sussex Kidney Unit work described in the first article was not the output of a consulting engagement. It was the work of clinicians who probed, sensed, and responded, sustained by an institutional context that allowed them to keep going. That is the model. The question is where it can scale.
If the change vehicle is a portfolio of complexity-informed experiments, the next question is where those experiments can actually run. The honest answer involves looking at the four largest health economies of the developed world and asking whether they are capable, in their current condition, of hosting the kind of work this requires.
The US is the world's most expensive demonstration of what happens when the soft transition is left unmanaged: chronic disease prevalence, multimorbidity, and the resulting catastrophic-event economy are all visible at scale. The system is so fragmented across Medicare, Medicaid, commercial payers, self-insured employers, and the VA that no single actor can coordinate change. What it does have is incentive heterogeneity. Self-insured employers and integrated delivery networks like Kaiser, Geisinger, and Intermountain carry real risk and are structurally capable of investing in trajectory-based care. The CMS Innovation Center's LEAD model (launching 2027 as the successor to ACO REACH) and the AHEAD state-based total-cost-of-care model are the most promising federal vehicles. The system as a whole will not move. Particular actors within it might.
Germany's structural problem is the depth of its tariff and codification infrastructure. The DRG-based hospital payment system, the EBM physician fee schedule, and the IQWiG-led benefit assessment process are formidable engineering — and all of them are organised around the event, the episode, and the threshold. The 2025 Krankenhausreform and the 2026 Krankenhausreformanpassungsgesetz are restructurings of acute hospital capacity, not of the soft-transition architecture. The €50bn Transformationsfonds (2026–2035) is genuinely large, but it is committed to consolidating hospital capacity, not to redesigning the relational and longitudinal infrastructure of chronic care. The federal-Land divide, the autonomy of the Krankenkassen, and the strength of the medical association lobby make complexity-informed change exceptionally difficult to organise nationally.
China's centralised health architecture under the National Health Commission is, in theory, capable of system-wide redesign in a way no other system can match. In practice, the focus of the past decade — universal coverage extension, hospital capacity expansion, drug pricing reform under volume-based procurement, and the management of an epidemiological transition still partly characterised by acute infectious disease pressures — leaves limited bandwidth for the trajectory-based, relationally-grounded work this requires. The infrastructure for longitudinal observation and warm-data collection at scale is also far less developed than the Western health-tech narrative would suggest. China will eventually be a major actor in this domain, but probably not in the next decade.
Japan has the developed world's most advanced demographic and morbidity profile — exactly the profile that should be driving soft-transition reform. Yet its long-standing fee-for-service architecture, the embedded role of small private clinics, and the limited authority of any single payer or commissioner make systemic change extraordinarily slow. The Long-Term Care Insurance system, in operation since 2000, contains genuinely interesting elements of trajectory thinking — but it is a separate budgetary universe from acute medical care, and the two do not yet operate as an integrated system. Japan is the world's clearest demonstration that demographic pressure alone does not produce system reform.
This is an uncomfortable conclusion, but it is the honest one. The four largest developed-world health economies are, for different reasons, structurally unlikely to lead the change this series has been arguing for. Each has elements that could contribute — particular states, particular Krankenkassen, particular integrated delivery networks — but none is in a position to organise the work nationally.
The change will start somewhere else.
The systems most likely to host the early work are those with three characteristics: a single accountable payer or commissioner with population-level responsibility, a manageable population size, and a current political or fiscal pressure that creates the demand for genuine innovation rather than incremental optimisation.
The NHS in England is currently undergoing its most significant structural reform since 2012, with ICBs consolidating from 42 to roughly 26 through April 2026 and 2027, 50% cuts to ICB running costs, the establishment of integrated health organisations through contractual rather than organisational vehicles, and the 10-Year Health Plan's commitment to shift resources from hospitals to community. The disruption is profound, the bandwidth is limited, and the risk is that all of this energy is absorbed in administrative reorganisation. Yet the architectural ingredients — single accountable commissioner, population-level budget, integrated provider relationships, and a forced reckoning with the financial unsustainability of acute-centric care — are uniquely aligned. If a serious complexity-informed pilot can be embedded in two or three of the new ICB clusters with protected funding and a long enough time horizon, England is the most likely host of genuine soft-transition work at scale.
Moving — fragile
Singapore's Healthier SG initiative, launched in 2023, is the developed world's most explicit policy attempt to build relational continuity into a national primary care system. Patients are encouraged to enrol with a single primary care provider, payment structures shift toward capitation, and structured Health Plans now form the spine of preventive interactions. The 2026 Care Protocols add depression, anxiety, allergic rhinitis, and advance care planning to the framework. The 2026 Health Information Act enables longitudinal data sharing across the care continuum. The implementation is imperfect — physician enrolment, administrative friction, and capacity remain real problems — but Singapore is, in policy terms, further down this road than any other system. It is also small enough to learn from itself coherently. The next five years of Healthier SG will be one of the most important natural experiments in international healthcare policy.
Moving — purposeful
Denmark in particular combines high-quality longitudinal population data, single-payer regional health authorities, and a clinical culture comfortable with measured experimentation. The Dutch system, despite its insurance-based architecture, has unusually strong primary care, GP gatekeeping, and an existing payer appetite for outcomes-based contracts. Sweden's Skåne region and Norway's Trondheim experiments are smaller but instructive. The Nordic systems collectively are likely to produce the cleanest evidence of what trajectory-based care looks like in practice — though their scale limits how directly the learning translates to larger systems.
Moving — incremental
The unsexy candidate. Self-insured employers in the US carry catastrophic risk directly, have no allegiance to the existing fee-for-service architecture, and increasingly purchase care through navigation platforms that are economically incentivised to identify drift early. They are not constrained by CMS rule-making or by traditional payer-provider contracting norms. Their weakness is fragmentation — there is no coherent 'employer view' — and their innovation depends on the quality of the vendor ecosystem they purchase from. But several large employer purchasing coalitions (Pacific Business Group on Health, the Health Transformation Alliance, Catalyst for Payment Reform) are now actively pushing for trajectory-based contracting. This is real and underappreciated.
Moving — opportunistic
The Spanish regional health systems, particularly Catalonia and the Basque Country, have for two decades run some of the most sophisticated population health management infrastructure in Europe — including risk stratification using the GMA (Adjusted Morbidity Groups) framework, integrated chronic disease pathways, and community-based proactive care models. They are small, well-instrumented, and unusually willing to publish what does and does not work. They are unlikely to drive global change but are likely to provide some of its most useful evidence.
Moving — quietly
The systems most likely to manage the soft transition first are not the largest, the richest, or the loudest. They are the ones small enough to learn from themselves, financially exposed enough to need to, and structurally simple enough to act.
The change vehicle for soft-transition reform is not a transformation programme. It is not a national strategy. It is not a model contract or a payment reform announcement. Those are the artefacts the complicated approach produces, and they are the artefacts the existing system is best at absorbing without changing.
What is needed is closer to an extended safe-to-fail experimental portfolio, organised around three principles drawn directly from the complexity-science literature:
First, parallel rather than sequenced. Multiple distinct experiments running simultaneously in different contexts, with explicit permission to develop differently in response to local conditions. The conventional change programme picks the best model and rolls it out; the complexity approach runs several models and learns from their divergence.
Second, time-horizon honesty. The trajectory of a soft transition is, by definition, measured in years. The evaluation of an intervention designed to manage trajectories therefore must be measured in years. Programmes evaluated on 12-month metrics will systematically select for interventions that produce 12-month metric improvements — which, almost by definition, are the wrong interventions.
Third, governance designed for emergence rather than control. The instinct of a system under financial pressure is to add governance, reporting, and oversight to any investment it makes. The instinct of a complexity-aware programme is to protect the experimental space from the very controls that would force it to behave like a complicated implementation. This is the single most difficult cultural shift, and the one most likely to break the work.
The catalyst — and there will be one — is unlikely to be intellectual. It will be a payer or commissioner who has run out of room within the existing model, who can no longer fund the catastrophic event economy at current rates, and who is forced into a more honest reckoning with what the soft transition is actually costing. That financial pressure exists today. What is missing is the political and institutional patience to respond to it with experiments rather than restructurings.
If this work begins, the early signals will not be the ones that make headlines. They will be small, structural, and easy to miss. The following are the indicators worth tracking — not for their individual significance, but as a constellation that would collectively suggest the soft transition is finally entering the operational vocabulary of healthcare.
If three or four of these signals appear within the next five years, the work is genuinely beginning. If none appear, the system has metabolised the discourse without altering its behaviour — which is the more likely outcome, and the one this series has been written to argue against.
Every change programme arrives at this moment. After the diagnostic, the analysis, and the future-state vision, there is always a numbered list of recommendations. This list is the bit that gets photographed, circulated, and printed on the back of a steering committee paper. It is also the bit most likely to be implemented as if the preceding twenty pages had not been written. With that hazard fully acknowledged, here is the list.
The vocabulary commits the category error before the work has begun. Name it for what it is: a portfolio of experiments, run in parallel, designed to develop rather than implement an answer. Protect that framing politically, because it will be eroded.
From the candidate list — selected ICB clusters in England, Singapore under Healthier SG, Denmark or the Netherlands at regional scale, US self-insured employer coalitions — pick the ones where the institutional conditions actually exist. Do not waste effort on systems that cannot move.
Pay for the long consultation, the trusted GP, the named care team — not as a quality bonus, but as the core delivery mechanism. The economics will be defensible over a five-year horizon and indefensible in any twelve-month evaluation. Build the financial commitment to survive that gap.
Longitudinal PROMs collected continuously, integrated with structured longitudinal clinical data, attributable to named clinical teams. Without this infrastructure, no intervention can be evaluated honestly. With it, even imperfect interventions become learnable. The infrastructure is the leverage point.
Every AI deployment in this portfolio should be evaluated on whether it increases or decreases the time clinicians spend in unhurried, attentive contact with their patients. If it doesn't measurably increase that time, it is not serving this work — regardless of how impressive its model performance is.
Companies with progressive chronic disease portfolios need trial designs, registry infrastructure, and real-world evidence platforms that generate trajectory-based outcomes — not just hard-event endpoints. The contracting environment that will require this is forming. The companies that have the evidence ready when it arrives will define the category.
This work will produce its first credible operational evidence in five years. Its first economic vindication in seven. Its first global policy reverberation in ten. Anyone offering faster is selling a programme rather than running one. Plan, fund, and govern accordingly.
It is a numbered list of seven recommendations. It looks exactly like the output of the consulting engagement this series began by criticising. That is the irony, and the trap. The recommendations are correct. Whether they get implemented as a complicated checklist or held as the principles of a genuinely complex experiment is the only question that matters.
This series has argued for a different way of seeing healthcare — one organised around the slow, relational, longitudinal work of managing chronic illness as it actually unfolds, rather than the dramatic, episodic, billable events that currently structure the entire economy of care.
That argument is not new. Versions of it have been made by clinicians, public health scholars, and patient advocates for thirty years. What is changing is the convergence of three things: an unsustainable financial trajectory in the systems most committed to the event-based model, an AI capability finally honest enough to be useful, and a generation of patients and clinicians who can articulate the wrongness of the current arrangement clearly enough to demand something else.
None of this guarantees change. The pressure is necessary but not sufficient. The institutional capacity to respond intelligently to that pressure — rather than reflexively — is the variable that determines the outcome.
And so here, in the time-honoured tradition of change programmes everywhere, we conclude with seven recommendations. They will, of course, fix everything.