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.
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.
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.
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?
Continuity is not a decorative service feature
The pathway cannot be built around BMI alone
A mental-health trigger changes the pathway, it doesn't raise a score
Prediabetes is a cardiometabolic entry point, not a glucose problem
More detection is not more care
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.
journey
Most of what fails happens between appointments, not inside them
backstage
Half the work that keeps a service running is invisible
where it breaks
What gets recorded as disengagement usually started as something else
What I left open
| still unresolved | how I handled it |
|---|---|
| unknown Which signal combinations predict disengagement, at what lead time | Built detection as a hypothesis engine with logged outcomes, not a trained predictor |
| unknown Whether intervening earlier changes the trajectory or just moves the exit later | Designed the measurement to answer it. Shipped nothing that assumes the answer |
| unknown Whether engagement patterns behave the same across income groups | Made subgroup monitoring a launch condition rather than a later phase |
| unknown Real referral completion baselines in partner environments | Instrumented 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.
| what I looked at | what I found in the current work | what I changed because of it |
|---|---|---|
| Workload | Coordination is picked up by whoever notices first, usually a nurse, and it shows up in nobody's job plan | Made it a visible queue with a counted volume, so it can be staffed |
| Thinking time | Rebuilding four months of a patient's story from three systems, before a fifteen-minute appointment | The summary is ready before they walk in, with the changes first and the source one click away |
| Interruptions | Generic alerts arrive mid-consultation and get dismissed without being read | A daily limit per role, with things batched into two review windows |
| Searching | The thing that would change the decision is in the record, several clicks away | Built the view around the decision being made, not around the data available |
| Remembering | Open referrals live in someone's head and a personal list, and they vanish at handover | Open items became objects with an owner and an age |
| Handovers | Quality depends on who wrote the note and how tired they were | Drafted from the record, edited by a person, never sent unread |
| Trust | Clinicians ignore a score they cannot see inside. They will argue with reasoning | Evidence sits above every recommendation, and disagreeing takes one tap |
| Who is responsible | Nobody owns the patient between appointments, so nothing happens and nobody is at fault | Every open item has one named owner. Unowned items escalate on their own |
| Workarounds | Spreadsheets, messaging apps, and a nurse's personal follow-up list | I 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.
| # | question I asked | exercise I ran | artefact it produced | decision it changed |
|---|---|---|---|---|
| 01 | Where 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 upon | Care intelligence gap map | I scoped AI to intelligence gaps inside the journey rather than automation on top of it |
| 02 | Which signals mean something only in combination? | Work longitudinal data across ten signal types, then introduce contradictory signals deliberately | Signal map | The room rejected an engagement score. Weight improving with engagement declining may mean nothing at all |
| 03 | Can AI explain why, not just detect that? | Same event, four different contexts, one question, would you intervene and why | Contextual reasoning map | Detect, explain, recommend an investigation. I rejected prediction without explanation |
| 04 | What should AI never do? | Thirty decision cards into five categories, each challenged with: what happens if the AI is wrong | Decision boundary map | I separated do-not-automate from human-only, a standing prohibition, not a capability judgement |
| 05 | Can it recommend the right next action? | Ranked fifteen scenarios, then asked what would change the answer | Next best action matrix | Reframed from predicting dropout to recommending the next care action. It also generated the data requirements |
| 06 | Where could AI remove work or improve decisions? | Every activity classified as understand, decide, create, coordinate, predict or personalise | Opportunity landscape | Coordinate and understand carried the most value. Predict carried the least, which nobody expected |
| 07 | Which of these is possible with the data? | Each opportunity against required data, source, owner, frequency, quality, integration difficulty, privacy sensitivity | Feasibility matrix | I killed psychological-risk inference on data grounds |
| 08 | What should we deliberately kill? | Five tests: meaningful problem, AI materially better than a rule, data available, someone will act, risk acceptable | Kill list and survivors | Most died on the same objection — nobody owned the workflow |
| 09 | Where can an agent actually operate? | Map observe → decide → contact → schedule → follow up → escalate → close and find the agentable steps | Agent opportunity map | I separated AI insight from AI orchestration. The operational value sat mostly in the second |
| 10 | What will patients accept? | The same message written four ways, a compliance nudge at one end and an explicit risk statement at the other | Acceptability principles | AI improves care behind the scenes rather than speaking to patients |
| 11 | What would make clinicians act on it? | The same recommendation with progressively more explanation, which would you act on | Explainability requirement | Output specified as evidence → interpretation → recommendation. Never score → alert |
| 12 | What should be funded? | Scored on value, readiness, risk, adoption and evidence | Opportunity portfolio | Build, 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.
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.
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.
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.
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.
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.