part one · discovery and research

What should coordinated care actually be?

Obesity and cardiometabolic care runs through five or six services that rarely speak to each other. Guidance in three countries says it should be coordinated. None of it says how.

So I started with the evidence, then mapped the people, the work and the decisions around one patient, to find where the coordination actually breaks.

patient GP Obesity clinic Dietitian Psychology Pharmacy Cardiology Diabetes service Physiotherapy Bariatric follow-up Work and social every thread is a relationship the patient maintains. none of them connect to each other.
The first thing I mapped. In cardiometabolic care the patient usually holds two or three conditions across separate services, so coordination is their unpaid job by default.

start here

A patient sees five or six services. None of them sees the others.

A person with a long-term condition sees a GP, a dietitian, a specialist, a pharmacist, sometimes a psychologist. Nobody is responsible for the join between them.

what it is like now GP Dietitian Specialist Pharmacy Psychology Lab Community patient the patient is the only connection design what coordinated care means EntryAssess PlanMonitor Close one journey GP Dietitian Specialist Pharmacy The coordination layer who owns the next step · what information moves · when a handoff is complete what happens when the situation changes · who is still responsible at the end the service carries the joining work, not the patient
AHRQ describes coordination as a set of activities — establishing accountability, communicating, facilitating transitions, monitoring and responding. Appointing a coordinator does not perform them. backed by guidance

The four things that actually break

Not clinical quality. The four below happen between services rather than inside an appointment.

How this case study is organised

Three parts, in the order the work actually happened. Each one only makes sense because of the one before it.

You can jump straight to the AI in part three. Part three is a list of features. Parts one and two are what makes any of them checkable.

In one sentence. A person can receive care from several services without experiencing several disconnected systems — and once that model is clear, AI can be assessed where it removes unnecessary work, makes important change easier to see, or keeps coordination from breaking. The care model came first. Every AI candidate here was assessed against it, and several were rejected.

the idea

The patient is the only one who sees the whole thing

The patient experiences one journey, even when several people and services deliver it. The coordination layer is what keeps that journey connected.

the patient experiences Entry Assess Plan Treat Monitor Review Close delivered by GP Dietitian Specialist Pharmacy Coordinator shared information Coordination who owns the next step · what information moves · how handoffs complete what happens when the situation changes · how the pathway closes the coordination layer is the designed object — not the appointment
AHRQ treats coordination as work that has to be specified before it can be assigned. Naming a coordinator does not specify it. backed by guidance
The working definition. Coordinated care is the deliberate organisation of the activities, information, decisions, responsibilities and transitions needed to make a patient’s care work as one connected journey — even when several people, teams or organisations are involved.

Synthesised from AHRQ’s care-coordination framework and WHO’s integrated people-centred services work. AHRQ notes the literature holds more than forty definitions and no single accepted one, so this is a design interpretation of converging evidence rather than a settled clinical taxonomy.

what has to stay true

What has to stay true, however the pathway is configured

Read the sentence, not the label. Each one is written the way a patient would recognise it, with the formal term underneath and the evidence it rests on beside it.

These hold across every configuration in this system. What changes is how they are delivered — the route, the roles, the intensity and the safeguards. They are a design synthesis grounded in converging evidence, not a published clinical standard.

the method

How this care model was designed

I started with the evidence. Each stage below was chosen after the previous one, not planned in advance.


each stage answers a question, produces an artefact, and changes the design

Select any stage for the artefact it produced and what it changed.

the evidence

Research already establishes what good care has to achieve

Plenty has already been published on continuity, integrated care, multiple conditions, obesity as a chronic disease, stigma, treatment burden and alert fatigue. My job was to translate it: what should good care achieve, and where does that fall apart in a real service?

what the evidence establishes Continuity and mortality22 studies, 9 countries, observational Obesity as chronic diseaseEASO 2024 · NICE NG246 Eating disorder recognitionNICE NG69 Cardiometabolic riskADA Standards 2026 Alert burden and overrideCDS human-factors reviews what I made the design do about it Build an explicit continuity mechanism — it does not emerge from adding people Assess complications and function, not a single number Let the trigger change the pathway, not just the risk tier Make cross-condition change visible and route it and set the interruption budget from capacity before tuning any detection
Evidence map. What the literature settles, and what I changed in the design because of it.
evidencesystematic review · 2018 · foundational

Continuity is not a decorative service feature

source says
22 eligible studies across nine countries; 18 high-quality studies reported significant mortality reductions with greater continuity. All observational. BMJ Open 2018 — I cite it as foundational, not current.
plain meaning
Being consistently known by the same clinician is associated with living longer.
i asked
Who holds the whole story here, and is that time allocated or done around a caseload?
i designed
An explicit continuity mechanism with a named owner and funded capacity. No AI capability substitutes for it, and I say so in product 03.
evidenceconsensus framework · 2024

The pathway cannot be built around BMI alone

source says
Diagnosis rests on anthropometric and clinical components together. Waist circumference measured where BMI is under 35. BMI ≥25 with waist-to-height ratio over 0.5 plus medical, functional or psychological complications is included in the diagnosis. Ethnicity-specific cut-offs apply. EASO, Nature Medicine 2024
corroborated
NICE asks adults with BMI under 35 to assess waist-to-height ratio, and reduces referral thresholds by 2.5 kg/m² for South Asian, Chinese, other Asian, Middle Eastern, Black African and African-Caribbean backgrounds because risk occurs at lower BMI. NICE NG246
i designed
Multidimensional entry assessment, with staging driving support intensity instead of a threshold value.
evidenceguideline

A mental-health trigger changes the pathway, it doesn't raise a score

source says
Do not use single measures such as BMI or duration of illness to decide whether to offer treatment. Assess physical health, associated mental health problems, and the need for emergency care where physical health is compromised or there is suicide risk. If suspected after initial assessment, refer immediately. NICE NG69
corroborated
The obesity guideline defers to it rather than continuing on its own track. NICE NG246
i designed
A divert, not an escalation. This trigger bypasses the coordination queue entirely and the AI layer does almost nothing in it.
evidenceguideline · 2026

Prediabetes is a cardiometabolic entry point, not a glucose problem

source says
Prediabetes is associated with abdominal obesity, dyslipidaemia and hypertension, and its presence should prompt comprehensive cardiovascular risk factor screening. Referral to prevention programmes including certified technology-assisted programmes is supported. Adults with prediabetes and cardiometabolic risk factors are screened for liver fibrosis risk using FIB-4 even with normal enzymes. ADA Standards of Care 2026, §2–4
i designed
One signal opens several assessment lines at once, each with a named owner and an explicit closure. This is the trigger where closure failure is most costly.
evidencehuman factors

More detection is not more care

source says
Clinical decision support override rates are consistently high across published reviews. Contributing mechanisms include workload, work complexity including patient comorbidity, and alerts with low informational value. Role-based tailoring improves acceptance.
i designed
An interruption budget set from measured capacity before any sensitivity is tuned. A signal nobody has time or authority to act on is liability, not intelligence.

people

Who benefits, who does the work, who carries the risk

Rarely the same person. I mapped this first, because a design can test well and still fail, because the work it adds lands on whoever cannot refuse it.

workcarried authority to change it → high work · low authority — this is where designs fail high work · high authority low work · low authority low work · high authority Nurse / care coordinatorcarries almost all coordination, funded for none of it Patientcarries the information between everyone, and all the risk Obesity physiciandecides treatment, referral and escalation Partner organisationowns the service, the clinicians and the patient relationship Dietitian, psychologist, physioown their intervention, rarely its sequencing GP / primary careholds the whole person; in France, holds the pathway Pharmaceutical companysets what is core and cannot be configured away Neither side owns: whether a referral completes · whether a signal reaches someone who can act · whether the patient still feels held · whether the model stays safe at scale.
Map of who is involved. The top-left quadrant is the design constraint. Coordination work that comes with no extra capacity reaches the coordinator and the patient before anyone else.

journey

Most of what fails happens between appointments, not inside them

what the patient lives — continuous hopefuldoubtplateau"is this working?"quietly stops what the system delivers — discrete assessment dietitian review DNA recorded file closes nobody herenobody herenobody herenobody here A plateau, a side effect, a stressful month, a booking that was too hard, a shift pattern — all of them arrive in the dashed spaces.
my conclusion The dashed spaces are what the coordination layer has to cover.

backstage

Half the work that keeps a service running is invisible

patient Starts treatmentinformed, hopeful Sees dietitianrepeats history Weight plateausdoubts treatment Cancels twicesays nothing Stops replyingfeels it failed Goneno decision made visible clinical work assessment, plan nutrition consult review in 8 weeks slot released marked non-attendance file closes line of visibility hidden coordination work Nurse re-enters the same history into three systemsunfunded · unscheduled · in no report Dietitian notices the drift and tells nobodyno route exists to share a soft signal Referral sits open, unowned, never closednobody is told it didn't happen breakpoints 1History repeated 2Soft signal lost 3Cancellation unexplained 4Eight-week blind gap 5Referral never closed 6Exit recorded as non-attendance rather than investigated as a decision Every breakpoint is a place where the information already exists and does not move. None of them need new clinical knowledge to fix.
discoveryillustrative I reconstructed this from workflow observation. The bottom band is work that keeps the service running and appears in no job description, rota or report.
evidenceHalf of bariatric patients lack two-year follow-up; the GP is often not informed HAS 2024
breakpointReferral open, unowned, never closed
factorReferral routes and tracking
principleCreating a referral is not delivering care
ai opportunityReferral orchestration
measureCompletion by stage · referrer informed

where it breaks

What gets recorded as disengagement usually started as something else

Symptom — patients stop attending, stop replying, stop treatment Immediate — a barrier appeared: plateau, side effect, stress, cost, a booking that was too hard Structural — care is delivered in appointments; the barrier appears between them Organisational — nobody is accountable for the patient between professionals Informational — the signals exist, in four systems, none of which talk Technological — systems record events; they do not represent a patient over time Design opportunity — detect the change, read the cause, route it to someone who can act, and close it This is the move that made an AI layer worth building. The thing to be intelligent about is the system's response, not the patient's willpower.

What I left open

still unresolvedhow I handled it
unknown Which signal combinations predict disengagement, at what lead timeBuilt detection as a hypothesis engine with logged outcomes, not a trained predictor
unknown Whether intervening earlier changes the trajectory or just moves the exit laterDesigned the measurement to answer it. Shipped nothing that assumes the answer
unknown Whether engagement patterns behave the same across income groupsMade subgroup monitoring a launch condition rather than a later phase
unknown Real referral completion baselines in partner environmentsInstrumented the loop first so partners measure their own baseline before anything is tuned

human factors

Every idea I had made work for somebody

So I worked through it in order. Who does the job. What they are trying to get done. What they need to know. What they decide. Where it breaks. And what happens when the system gets it wrong. Technology came last, on purpose.

PeopleNeedsTasksDecisionsWorkflowInformationSystemsAIOversightOrgOutcomes I only reached "AI" at step eight. Anything I put there had to survive the seven steps in front of it. The rule I kept coming back to A signal nobody has time or permission to act on is not intelligence. It is a liability, because now somebody knows about a problem they cannot fix. So I set how many things a person can be interrupted with in a day before I let anyone tune what the system looks for.
what I looked atwhat I found in the current workwhat I changed because of it
WorkloadCoordination is picked up by whoever notices first, usually a nurse, and it shows up in nobody's job planMade it a visible queue with a counted volume, so it can be staffed
Thinking timeRebuilding four months of a patient's story from three systems, before a fifteen-minute appointmentThe summary is ready before they walk in, with the changes first and the source one click away
InterruptionsGeneric alerts arrive mid-consultation and get dismissed without being readA daily limit per role, with things batched into two review windows
SearchingThe thing that would change the decision is in the record, several clicks awayBuilt the view around the decision being made, not around the data available
RememberingOpen referrals live in someone's head and a personal list, and they vanish at handoverOpen items became objects with an owner and an age
HandoversQuality depends on who wrote the note and how tired they wereDrafted from the record, edited by a person, never sent unread
TrustClinicians ignore a score they cannot see inside. They will argue with reasoningEvidence sits above every recommendation, and disagreeing takes one tap
Who is responsibleNobody owns the patient between appointments, so nothing happens and nobody is at faultEvery open item has one named owner. Unowned items escalate on their own
WorkaroundsSpreadsheets, messaging apps, and a nurse's personal follow-up listI treated these as requirements, not bad habits. The nurse's list became the coordinator queue

discovery The last row matters most. People build workarounds where the system fails them, so the workaround is a map of the failure.

designed investigation

Technology came in session six, not session one

We kept the same core group throughout, so each session built on the last one instead of starting again. Every session opened with published evidence. When a session opened with speculation it stalled. When it opened with a published finding, the specialists argued with the finding and then committed to a position.

01–03Find the care gapecosystem · journeys · blueprint 04–05Define coordinationboundaries · next-best-action 06–08Test and killopportunity · data · kill room 09–12Design the layeragents · trust · market · GTM in the room every session endocrinologist · psychologist · dietitian · cardiologist · GP · nurse coordinator · operations lead · data lead · me
#question I askedexercise I ranartefact it produceddecision it changed
01Where does care intelligence break?Reconstruct journeys as event → information available → decision required → decision made → outcome, marking each step known, unknown, delayed, fragmented, manually interpreted or not acted uponCare intelligence gap mapI scoped AI to intelligence gaps inside the journey rather than automation on top of it
02Which signals mean something only in combination?Work longitudinal data across ten signal types, then introduce contradictory signals deliberatelySignal mapThe room rejected an engagement score. Weight improving with engagement declining may mean nothing at all
03Can AI explain why, not just detect that?Same event, four different contexts, one question, would you intervene and whyContextual reasoning mapDetect, explain, recommend an investigation. I rejected prediction without explanation
04What should AI never do?Thirty decision cards into five categories, each challenged with: what happens if the AI is wrongDecision boundary mapI separated do-not-automate from human-only, a standing prohibition, not a capability judgement
05Can it recommend the right next action?Ranked fifteen scenarios, then asked what would change the answerNext best action matrixReframed from predicting dropout to recommending the next care action. It also generated the data requirements
06Where could AI remove work or improve decisions?Every activity classified as understand, decide, create, coordinate, predict or personaliseOpportunity landscapeCoordinate and understand carried the most value. Predict carried the least, which nobody expected
07Which of these is possible with the data?Each opportunity against required data, source, owner, frequency, quality, integration difficulty, privacy sensitivityFeasibility matrixI killed psychological-risk inference on data grounds
08What should we deliberately kill?Five tests: meaningful problem, AI materially better than a rule, data available, someone will act, risk acceptableKill list and survivorsMost died on the same objection — nobody owned the workflow
09Where can an agent actually operate?Map observe → decide → contact → schedule → follow up → escalate → close and find the agentable stepsAgent opportunity mapI separated AI insight from AI orchestration. The operational value sat mostly in the second
10What will patients accept?The same message written four ways, a compliance nudge at one end and an explicit risk statement at the otherAcceptability principlesAI improves care behind the scenes rather than speaking to patients
11What would make clinicians act on it?The same recommendation with progressively more explanation, which would you act onExplainability requirementOutput specified as evidence → interpretation → recommendation. Never score → alert
12What should be funded?Scored on value, readiness, risk, adoption and evidenceOpportunity portfolioBuild, pilot, research or do-not-build, each traceable back to a factor

discovery Method and artefact structures shown; client findings withheld. Where a session conclusion matches published evidence I state it here through the published source. Agreement in a room is a discovery finding, not clinical validation, and I have not presented it as one.

session 11 — the same patient, four ways of saying it a"Patient dropout risk: 78%."nobody would act b"Patient has declining engagement."still not actionable c"Engagement declined over four weeks, alongside a plateau and increased stress."some would d"Possible disengagement driven by treatment frustration. Evidence: plateau, declining nutrition engagement,increased stress. Suggested: psychologist outreach. Human review required."acted on
I produced this finding rather than citing it, four versions, one question. Version D is built as the recommendation screen in product 03.

artefacts

What the work actually produced

These are the objects I actually worked in. They demonstrate the method without disclosing anything confidential, and most of them are reusable as templates on a different problem.

Ecosystem mapEvery organisation touching one patient
Map of who is involvedWork carried against authority held
Patient journeyContinuous experience against discrete episodes
Care-team journeyThe same period from four viewpoints
Evidence mapFindings into design principles
Evidence / assumption boardWhat we know against what we assumed
Root-cause treeSymptom separated from cause
Hidden-work mapWhat keeps the service running and isn't counted
Breakpoint mapSix places where information exists and doesn't move
Scenario cardsSame event, four contexts, four right answers
AI opportunity mapValue against readiness, with the kills visible
ai · human · shared
Human / AI responsibility canvasWho proposes, who decides, who signs
Core / configurable frameworkWhat travels and what cannot
Partner configuration canvasEight things a deployment must declare
Failure-mode boardFailure, detection, response, recovery
Trade-off boardEight tensions with a dial position and an owner
Measurement architecturePatient, clinical, coordination, operational, AI
Before / after blueprintThe same patient month, run twice

what can be inspected

Principles say what good looks like. Factors are what a deployment can be examined for.

I pulled these together from the evidence and the human-factors work. Each one is something a patient's journey either has or does not have. They are my working model, not a validated clinical framework, and I say so everywhere they appear.

d1Continuity of personSomeone knows the whole storyand stays recognisable over time d2Shared understandingEveryone works from the samepicture, including what's missing d3Accountability betweenSomeone owns the patient and theopen work between appointments d4ResponsivenessTime from a change to someoneactually doing something d5ClosureThings that start get finished, andwhoever started them is told d6Patient agencyThey take part without runningthe admin of their own care d7ProportionalityIntensity matches need, risk andthe capacity that actually exists d8Fits where it is being usedIt works within local scope,governance and partner reality a design synthesis — not a validated clinical index

How the factors connect to the principles

The eight above are the properties. The twenty-four below are what I can inspect, measure and redesign to get them. The heatmap in product 02 scores these for a specific deployment.

Structural Informational Relational Operational Population Governance D1 Continuity of person D3 Accountability between contacts D2 Shared understanding D4 Responsiveness D6 Patient agency D5 Closure D8 Fits where it is being used D7 Proportionality six groups · twenty-four factors · eight drivers. no factor affects only one driver, which is why single-issue fixes rarely hold.

the reusable system

Nine capabilities that recur, assembled differently each time

These are not a clinical taxonomy. They are a design synthesis assembled from the coordination activities AHRQ describes, person-centred integrated-care frameworks and the obesity pathway guidance — which is why each one carries the activity it traces back to.

Product 02 assembles these into a configured pathway. A module can be present, conditional or absent depending on trigger, market, partner, population and capacity.

assessed against the model

Where AI could help this care model

The care model came first. This is what was still difficult once it existed.

The team can already detect more than it has capacity to act on. That is the constraint each candidate below was assessed against. Everything below is assessed against a module in the model above — including the things worth refusing to build.

What the evidence does not justify.

These stay labelled as design propositions, contextual hypotheses or open questions throughout the rest of the system.

    The system should still be safe if the AI layer is switched off. That is the test each candidate above has to pass before it earns a place.

    check any of it

    Everything the design leans on, with a link to the original

    Every source chip on this page opens the source. Where a claim is mine rather than the literature’s, it is labelled as a design decision or an inference, and there is nothing to click.

    02 · why this tool exists

    The same care model, built differently for each partner

    The pharmaceutical company doesn't see patients. Partners do, telehealth services, provider groups, pharmacy networks, patient-support programmes. Each has different roles, different data, different rules and different capacity.

    One digital service given to all of them does not fit any of them. Building each from scratch does not scale past the first few partners. So I built a core that stays fixed and a delivery layer that bends. This is the tool that assembles it.

    The standard of care stays the same. What gets built changes, and this tool shows which parts change and why.

    the pharmaceutical company brings
    • The care model and its definitions
    • Evidence and measurement design
    • The intelligence layer and its governance
    • Configuration tooling and market packs
    the partner brings
    • The service and the patients
    • Clinical accountability
    • Care-team capacity
    • The patient relationship and its tone
    neither owns alone
    • Whether a referral completes
    • Whether a signal reaches someone who can act
    • Whether the patient still feels held
    • Whether it stays safe at scale
    why a partner buys it
    • More coordinated capacity without more headcount
    • Fewer referrals lost
    • Less duplicated information work
    • Evidence the model is delivered as intended

    02 · pathway designer

    Set the context. Watch the pathway change.

    Click any step to see where it came from.

    a step in the pathway AI supports here human decision safety gate referral / handoff E backed by guidanceM changes by marketP changes by partnerH human judgement requiredI my inferenceclick any node to see why it is there

    02 · modular architecture

    Every deployment assembles a different set of the same modules

    This is what makes it reusable. Each module has a job, an owner, the data it needs, something to measure it against, the way it usually goes wrong, and where AI is worth using.

    m1
    Entry and triage
    purpose
    Get the person into the right pathway, not just into the service
    owner
    Whoever the market makes the front door — GP, self-referral or partner clinician
    data
    Referral source, presenting signal, existing conditions
    benchmark
    Proportion entering the pathway their trigger indicates
    fails when
    Every trigger is funnelled into one generic pathway
    ai
    Assembling what's already known before the first appointment
    m2
    Assessment and staging
    purpose
    Understand severity, complications and function — not a single number
    owner
    Clinician, with multidisciplinary input where it exists
    data
    Body measurements plus what they mean for health, functional and psychological
    benchmark
    Published: assessment covers medical, functional and psychological dimensions EASO 2024
    fails when
    BMI becomes the decision rule
    ai
    Flagging what's missing from the assessment record
    m3
    Goals and care planning
    purpose
    Agree what success means for this person, in their words
    owner
    Clinician with the patient
    data
    Stated goals, preferences, constraints, social context
    benchmark
    Service: proportion of plans with a patient-stated goal recorded
    fails when
    The plan exists in one system and nobody else can see it
    ai
    Drafting the shared plan so every professional sees the same one
    m4
    Monitor and detect
    purpose
    Notice meaningful change between appointments
    owner
    The coordination layer, within a set interruption budget
    data
    Three or more sources, including scheduling or patient-reported
    benchmark
    Human-factor: items per person per day stay inside the budget
    fails when
    Sensitivity is set by what's detectable rather than what's actionable
    ai
    Grouping signals into one situation with its evidence attached
    m5
    Coordinate and hand off
    purpose
    Get the right person to act, with the context they need
    owner
    Named coordinator, or the clinician where no such role exists
    data
    Role map, what each role is allowed to do, capacity, open items
    benchmark
    Service: handoffs carrying context, questions and open items
    fails when
    Coordination is absorbed informally and appears in no plan
    ai
    Scope-aware routing and drafting the handoff from the record
    m6
    Intervention and support
    purpose
    Deliver the thing that addresses the actual barrier
    owner
    The professional the barrier calls for
    data
    Cause hypothesis, prior interventions, what worked before
    benchmark
    Service: interventions completed, not interventions started
    fails when
    An operational problem receives a clinical intervention
    ai
    Proposing the action with evidence, never enacting it
    m7
    Review and adapt
    purpose
    Check whether it worked and change the plan if not
    owner
    Clinician with the patient
    data
    Outcome of the last intervention, trajectory since
    benchmark
    Design target: every intervention has a recorded outcome
    fails when
    Review is scheduled by calendar rather than triggered by change
    ai
    Summarising what changed since last contact, source-linked
    m8
    Escalate
    purpose
    Move urgent things past the queue
    owner
    Named clinician, with a defined clock
    data
    Risk class, current owner, item age
    benchmark
    Human-factor: unowned item age never exceeds the escalation window
    fails when
    Silence counts as a decision
    ai
    Raising unowned and overdue items automatically
    m9
    Close or transition
    purpose
    Finish things properly, including stopping
    owner
    Whoever started it
    data
    Seven referral stages, outcome, reason for closure
    benchmark
    Published: only half of bariatric patients get two-year follow-up, and the GP is often not informed HAS 2024
    fails when
    Creating a referral is counted as delivering care
    ai
    Chasing to completion and returning the outcome to the referrer

    02 · benchmarks and maturity

    Not every number on this page means the same thing

    I keep these four apart on purpose. Something published is not the same as something I decided to aim for. Neither is a clinical threshold, and I have not invented any.

    published
    From the literature

    Continuity and mortality across 22 studies. Around half of specialty referrals completing. Half of bariatric patients without two-year follow-up. High override rates in decision support.

    service
    From this partner

    Their own referral completion by stage, their own time from change to action, measured over ninety days before anything is switched on.

    human factor
    From capacity

    Items per person per day. Review latency. Unowned item age. Set from what the team can absorb, not from what the detector can find.

    design target
    What I am aiming at

    Every intervention has a recorded outcome. Every referral closes explicitly. Nothing reaches a patient unread. These are my targets, not evidence.

    Coordination maturity

    FragmentedEach service holds its own record. The patient is the integration layer.
    ConnectedRecords can be joined. Someone can see the whole story on request.
    CoordinatedSomeone owns the patient between contacts. Referrals close.
    AdaptiveContact follows change rather than the calendar. Intensity matches need.
    IntelligentChange is detected across conditions and routed with a cause attached.
    LearningMissed signals are found on a schedule and the model is retuned or retired.

    This is a way of describing progress, not a clinical scale. Most services I looked at sit between fragmented and connected, and the jump that matters most is to coordinated, because that one needs a funded person, not software.

    02 · market comparison

    The principles hold across markets. The operating conditions do not.

    Picking a market does not change what good care is. It changes who is allowed to do what, and what information actually moves. That changes where the journey breaks, and which AI work is worth paying for.

    Shared principles — continuity · shared understanding · accountability · responsiveness · closure · agency · proportionality · fits where it is being used operating conditions United KingdomPrimary care gateway into commissionedservices; pathway varies by region FranceMédecin traitant coordinates; eight rolesheets define information sharing back SpainRegional health services differ; GIRO usesEOSS staging and non-BMI indicators what changes in the pathway Entry is the GP; referral criteria arelocally commissioned, so routing mustbe configured below national level Every professional reports back to themédecin traitant. Return-of-informationis a core pathway step, not an extra Staging drives intensity. Configurationis regional, and the model must showwhere integration is thin which factors become critical Structural — pathway shape andscope vary by commissioner Operational — handoff quality andreferral tracking back to one anchor Structural and informational —regional variation plus thin integration which AI opportunity matters most Configuration assistance — manylocal variants to stand up Referral orchestration and outcomereturn — the documented failure here Cross-source reconciliation — joiningrecords is the binding constraint
    evidence UK: NICE NG246. France: HAS care pathway guide, February 2024 update. Spain: GIRO, SEEDO. Each row is a consequence of what the guidance specifies, not a generalisation about the health system.

    The strongest market finding I have. French national guidance itself reports that only half of patients who have had bariatric surgery receive follow-up at two years, and that the treating doctor is often not even informed the operation took place. A closure failure reported by the body that writes the pathway, about its own pathway.

    02 · factor diagnosis

    Where this deployment is most likely to break

    All twenty-four factors, scored for the context set above. Darker means more likely to decide whether this journey works. Click any factor for what it means and what to do about it.

    02 · partner model

    What travels, what changes, and what never automates

    core — identical everywhere
    • Patient state definition
    • Activation state meanings
    • Seven referral stages
    • Human-in-the-loop rules by risk class
    • Audit trail, source labelling and how explanations are written
    • An agreed clinical stop counts as a completed journey
    market configurable
    • Pathway shape and entry points
    • Consent basis and lawful purpose
    • Professional scope mapping
    • Communication norms
    • National or regional routing
    partner configurable
    • Signal thresholds and sensitivity
    • Routing rules and roles
    • Contact cadence and quiet hours
    • Interruption budget
    • Escalation ladders and cover
    human and local
    • Whether this patient needs a call today
    • Difficult conversations
    • Changing treatment
    • Interpreting genuine ambiguity
    • Deciding an appropriate stop

    The last item in the first column was the hardest to hold. A retention-shaped system treats every exit as failure. I made an agreed, clinically appropriate stop a completed journey, counted separately from dropout.

    02 · gtm blueprint

    The commercial output

    inherited from 01 and 02

    Your care model

    This is the model you configured. Nothing below is a general AI proposal — it is assessed against this pathway.

    Change the context in and this pathway, its owners and every AI surface below move with it.

    assessed against the pathway

    Where AI could help

    Each marker sits on a care step where the work is repetitive or information-heavy. Select one for the module it belongs to and what it would do.

    What was considered, and what was rejected

    Four of these were rejected. The reason each one was rejected is listed beside it.

    the commercial case

    Staying in care

    Disengagement gets recorded as a patient behaviour. In the journeys I mapped, it followed something in the pathway failing first.

    Each drop point below is one I could name a coordination cause for. That matters commercially: a partner is paid on people who remain in a pathway long enough to benefit. It also matters for what gets built — the answer to each drop point is removing the friction that caused it, never scoring the person who hit it.

    An illustrative cohort, not a measured result. The lower line is what the pathway loses today; the upper is the design target once each drop point has an owner and a closing loop. illustrative · requires validation
    What this is not.

    Not a retention score, an engagement rank, or a churn prediction. None of those tell anyone what to do, and all of them move the blame onto the patient. Every measure above is a property of the service.

    one product, six moments

    Applications

    Referral coordination is the flagship. The rest are the moments around it.

    These are the same product at six points in one patient’s course, not six separate screens. Every AI output above carries its evidence and can be edited, rejected or escalated by the person who owns the step.

    beyond the referral

    The rest of the application set

    The six moments above follow one referral through one patient’s course. These are the other places the same product shows up — across assessment, planning, accountability, safety, review, capacity and the patient’s own view of their information. Each one is tied to a care module and carries what the human still decides.

    Two of these are deliberately unglamorous. Unowned work and the coordinator day list are the ones a service would feel first, and neither needs a model doing anything clever.

    the boundary

    What AI does. What people decide.

    ai

    seven things the system does assembled context one accountable decision
    Seven system actions on the left, all reversible and reviewable. One human decision on the right, which is not.

    human

    The test. The care model has to remain safe with the AI layer switched off. Anything that fails it has been built as a dependency and needs rescoping.

    before any of this ships

    Safe to use?

    Data has to arrive with its provenance attached, and every suggestion has to clear the same gates before it reaches a person.

    what moves what travels with it

    The gates

    A suggestion that cannot clear a gate is not shown. The gate that stopped it is recorded, because a silent suppression is its own safety problem.