TL;DR
From 8 a.m. on 2 October 2026, Jamaica's junior doctors withdrew all services except emergency care for 24 hours. The Jamaica Medical Doctors Association says the Government owes J$31 billion in retroactive overtime. The Government's own calculation is J$23 billion. That J$8 billion gap may be partly a records dispute, and matching records is work AI tools can do. AI cannot decide the case, which belongs to the Industrial Disputes Tribunal. It can help build the shift ledger that both sides need and flag the runs of consecutive hours that make doctors unsafe. A records officer can start this week by matching 100 shifts across three documents and counting the mismatches.
What happened
The Jamaica Medical Doctors Association (JMDA) told the public that it could not guarantee normal services from 8 a.m. on Friday 2 October to 8 a.m. on Saturday 3 October. The Jamaica Observer reported that only emergency care would run, and that the doctors have worked at the same overtime rates since 2021. Dr Jubilee Brown-Service, the association's president, had set the 2 October deadline for a ruling.
The dispute reaches back further. The Observer's report of 30 September says it dates to 2023, was referred to the Industrial Disputes Tribunal (IDT) in 2025, and that the JMDA's case was completed before the tribunal on 1 April. The association puts the arrears at J$31 billion. The Government's figure is J$23 billion. Dr Renee Badroe, the association's immediate past president, told the paper that overtime "can extend up to 80 hours consecutively".
The Government's statement said it is committed to "a fair, lawful and peaceful resolution" and that the matter is before the tribunal for compulsory arbitration under the Labour Relations and Industrial Disputes Act.
The effects showed up on the wards. The Gleaner reported that more than 100 junior doctors at Mandeville Regional Hospital in Manchester stayed off the job. The senior medical officer, Dr Everton McIntosh, said the hospital closed its outpatient clinic and suspended elective surgeries. Dr Brown-Service said health centres had limited staffing and many patients had to reschedule. The IDT acknowledged that unusual delays could erode public confidence in its work, and gave no date for the award. The JMDA's members will meet early this week to decide their next step.
The numbers
Start with the gap. The JMDA's J$31 billion less the Government's J$23 billion is J$8 billion, which is 25.8% of the larger figure. Two parties looking at the same doctors and the same years reached totals more than a quarter apart. The public reports do not say whether the difference sits in the rate applied, the hours counted or who is included. Those three are our guess at where it could be.
The supply of doctors is thin. The World Bank's physicians indicator gives Jamaica 0.46 per 1,000 people in 2023, down from 0.548 in 2018 (the series read 0.428 in 2015, so it rose and then fell). Fewer doctors share the same night shifts, and the arithmetic of overtime follows. The series does not say how many of those doctors work on a public ward at 3 a.m.
Long shifts cost patients. Landrigan and colleagues ran a randomised study in intensive care units, published in the New England Journal of Medicine in 2004. Across 2,203 patient-days and 634 admissions, interns on a traditional schedule with shifts of 24 hours or more made 35.9% more serious medical errors than on a schedule that removed those shifts (136.0 against 100.1 per 1,000 patient-days). They made 5.6 times as many serious diagnostic errors (18.6 against 3.3). The companion study by Lockley and colleagues followed 20 interns: 17 of them worked more than 80 hours a week on the traditional schedule, with a mean of 84.9, against 65.4 on the intervention schedule.
On the paperwork side, a randomised trial in NEJM AI assigned 238 outpatient physicians in 14 specialties at one California health system to Microsoft DAX Copilot (79 doctors), Nabla (79) or usual care (80). Nabla cut time spent in notes by 9.5% (95% confidence interval 1.8% to 17.2%). DAX's 1.7% reduction was not statistically distinguishable from zero. Both groups scored higher on the 10 to 50 Mini-Z wellbeing scale, by 2.83 for DAX and 2.69 for Nabla. The trial ran from 4 November 2024 to 3 January 2025 on English-only encounters, and its time-in-note measure did not count editing done inside the vendors' platforms.
What this means in Jamaica
We apply the method Adrian Dunkley calls People First, Problem Second, AI Last. A test we suggest is to write the human need and the problem in two sentences each with no mention of AI. The need: a patient in Manchester has to be seen on the date the clinic promised, and a doctor has to be paid for the hours worked without being on shift for 80 hours at a stretch. The problem: two parties hold different totals for the same hours, no shared record that both accept has been described in public, and nothing in the roster warns anyone when a run of hours passes a safe limit.
Neither sentence needs the word AI. The need is a clinic that opens and a pay slip that is right. The problem is a missing record and a missing rule. AI comes last, as the tool that builds the record and checks the rule.
That ordering matters for a reason specific to this week. The tribunal will rule, and the ruling will set a rate. Whatever rate it sets, someone will have to apply it to many past shifts, in many hospitals, over several years. If the duty rosters, sign-in sheets and payroll claims do not agree with each other, the award starts a second dispute about the arithmetic. We found no published description of how Jamaica's regional health authorities record doctors' overtime, so we cannot say how far apart those records are. The J$8 billion gap suggests they are not close.
Fix the ledger first
The task is dull and large. Each doctor, each date and each shift appears in up to three places: the duty roster, the attendance record, and the overtime claim sent to payroll. A match means the three agree. A mismatch has a reason: a swap that never reached payroll, a night that ran past midnight and was booked to the wrong day, a grade entered under the wrong rate.
AI helps in two places. Optical character recognition and a language model can read scanned or photographed rosters, which in many wards still start as paper, and turn them into rows in a table. A matching routine can then pair rows across the three sources and sort the mismatches by type. The arithmetic itself belongs in a spreadsheet or a script that anyone can audit, because a language model can state a wrong total in a calm voice. The model reads and sorts, and the spreadsheet adds up.
This is our inference. We found no published test of this approach on Jamaican hospital records, and the studies above come from other settings. What we can say is that it changes the nature of the dispute. Instead of J$31 billion against J$23 billion, each line of the gap carries a reason, and the reasons can be argued one at a time.
Rosters that cap the hours
The Landrigan result is the reason this matters for patients as well as pay. A roster is a set of rules plus a lot of exceptions, and software is good at both. A scheduling tool can refuse a draft roster that contains a run longer than a set number of consecutive hours, show the manager which swaps would fix it, and log each override with a name against it.
Two limits apply to the evidence. The interns in those trials were first-year residents in intensive care units, the studies were published in 2004, and the patients and wards in Manchester differ. We expect the direction to transfer and the size of the effect to differ, and no Caribbean study has tested it. The second limit is the supply figure. At 0.46 doctors per 1,000 people, a cap that removes hours must be paid for with other doctors' hours, and a roster optimiser cannot create a doctor. It can show a minister exactly how many extra posts a cap implies for a given ward, which is a number the negotiation lacks.
Mandeville showed the other half of this on 2 October. With more than 100 junior doctors off the job and consultants covering, the hospital curtailed services. A roster tool that models cover in advance would not have prevented a withdrawal. It would show a hospital manager, a day earlier, which clinics and theatre lists to move first, ranked by clinical urgency, so that patients get a new date by text message instead of a closed door.
Paperwork takes hours too
The NEJM AI trial gives the best recent measure of what a scribe tool does and does not do. One of two products cut time in notes by 9.5%. Both lifted wellbeing scores. The trial did not measure overtime, so it cannot say that a scribe would remove one hour from a Jamaican doctor's week.
There is a second gap. The trial used English-only encounters in California. A ward in Mandeville runs on English, Jamaican Patois and a mix of both within a sentence, and we found no published test of an ambient scribe on that speech. This is where we would apply the idea Adrian Dunkley calls Cognitive Debt, which warns that automating a task without checking its output removes the step where errors get caught. A test we suggest is for one week to have a clinician compare a sample of AI-written notes with what was said, and log how many are wrong. The error rate is the measurement. A tool that fails on Patois should not reach a patient record.
What the Ministry, the regional authorities and the JMDA should ask for
Four requests cost little. Publish how overtime hours are recorded and who signs them off. Agree a single, shared extract of shifts worked, so that the tribunal's award can be applied to one table. Set a maximum consecutive-hours rule in each roster and report breaches monthly. And require any vendor selling a scribe or a scheduling product to report its accuracy on Jamaican speech and Jamaican rosters. Under the idea Adrian Dunkley calls Agent Colonialism, a system imposed on a community with no figures on how it performs there carries unknown accuracy, and the community bears the errors.
What to do this week
These steps are written for a hospital records or payroll officer, a JMDA branch representative and a patient. The test in step two is one we suggest, and it is not a method Adrian has published.
- Choose one department and one month, and gather its duty roster, its attendance record and its overtime claims. Finish by 9 October. Remove doctors' names and replace them with codes before anything leaves the hospital's own systems.
- Run a 100-shift match. By 16 October, pick 100 shifts and check each against all three records. Count the shifts where all three agree. That share is your baseline. Sort the rest into reasons: wrong day, wrong rate, missing claim, missing roster entry.
- Test any AI tool against your hand count. Give it photographs of the same roster pages with names removed, and ask it to list the shifts it reads. Accept it only if it reproduces your hand count on all 100 shifts. Anything less means a person still checks every row. Finish by 23 October.
- Count the long runs. From the same month's roster, list every run of consecutive hours above 24. Then set a cap rule for the next four weeks and count the breaches and the overrides by 1 November.
- If you are a patient, ask the clinic for a written new date before you leave, note your repeat prescription date, and use the emergency department for any symptom that worsens.
Where this can go wrong
A roster optimised for cost can lengthen shifts. A tool told to minimise overtime spend will happily put a doctor on a longer run, and the Landrigan figures show who pays: patients in the ward, who faced 22.0% more serious errors on the traditional schedule (193.2 against 158.4 per 1,000 patient-days), and the doctor on shift. The guardrail is to set the hour cap as a hard constraint that the optimiser cannot trade against cost.
A scribe can get a note wrong. In the NEJM AI trial, clinicians said inaccuracies appeared "occasionally" (a mean of 2.7 for DAX and 2.8 for Nabla on the survey's scale), and the issues they named were omissions (12 reports), structure (11) and pronoun errors (8). The patient bears a wrong note. The guardrail is the sample check described above, with a clinician signing every note.
A shift ledger can become a surveillance record, and an ambient scribe can send recordings of patients to a server outside Jamaica. Neither is documented in a Jamaican hospital that we know of, so these two risks are our inference from how such systems are built. Junior doctors, who have the least say on a ward, would bear the first, and patients the second. The guardrails are a written rule that the ledger is used for pay and safety caps only, with the JMDA able to see the data, and a question to every vendor, before signing, about where audio is stored and for how long.
A confident wrong total is the last risk. Keep the arithmetic in a spreadsheet, and keep the tribunal as the one body that decides what is owed.
The weakest part of this argument
We do not know why the figures differ. The J$8 billion gap may come from a disagreement about the rate, which no software can resolve, rather than about hours, which it can. The ledger idea only helps if the dispute is partly about the count. The safety evidence comes from interns in intensive care units two decades ago, and the scribe trial from one California system. We found no Jamaican study of any of these tools. And a better record does not pay a single doctor. That decision sits with the tribunal, which has not given a date.
Questions readers ask
Who decides how much the doctors are owed, and when?
The Industrial Disputes Tribunal decides, under compulsory arbitration. The Jamaica Observer reported on 30 September that, according to the JMDA, the tribunal's only response was a broadcast statement by its chairman, Errol Miller, that he is not obligated to adhere to any ultimatum given by the union. The Gleaner reported on 3 October that the tribunal has given no date for its award. The 25.8% gap between the two figures is a matter for the tribunal, not software.
Are AI scribes accurate enough for a hospital note?
Not without a check. In the NEJM AI trial, clinicians reported inaccuracies "occasionally", with mean scores of 2.7 for DAX and 2.8 for Nabla. The most common faults were omissions (12 reports), structure (11) and pronoun errors (8). Nabla's drop in work exhaustion, 0.23 on a 0 to 4 scale, had a confidence interval that crossed zero.
Has Jamaica's doctor supply been falling?
Not in a straight line. The World Bank series reads 0.428 physicians per 1,000 people in 2015, 0.471 in 2016, 0.548 in 2018 and 0.46 in 2023. The latest value is about 16% below 2018 and above 2015. The series does not say how many of these doctors work in public wards.
What else did shorter shifts change in the intensive care studies?
Landrigan's team found interns made 20.8% more serious medication errors on the traditional schedule (99.7 against 82.5 per 1,000 patient-days). The total rate of serious errors on the critical care units was 22.0% higher (193.2 against 158.4). Those are further results beyond the 35.9% and the 5.6 times quoted above.
Is it safe to put hospital rosters or patient notes into a public chatbot?
Not with names, ID numbers or clinical details attached. Remove them before testing any tool, and use a hospital-approved system for real records. A roster that shows who is on duty overnight is also sensitive. Test on a made-up month first, then check every figure the tool reads against the paper original.
Supported by StarApple AI, the Caribbean's first AI company. Read more at StarApple AI, AI Boss and the Caribbean AI Risk Management Council. Related reading: Jamaica's nurses, Ghana and the health workforce and AI in Jamaican healthcare.