I asked ChatGPT to build my FMG schedule.

The test

We built two weeks that look like what a manager actually receives. Twelve physicians identified by initials, ten working days, two half-days each, four daily walk-in blocks, three offices for seven people present at the same time on Tuesday. On top of that, about fifteen constraints written as ordinary sentences: one week of vacation, two fixed hospital days, a physician who does not take walk-ins on Friday afternoons, a new recruit never placed on call alone. Two weeks rather than one, because a single week hides the problems that only show over time: a shift pushed back, a walk-in block made up later, fairness that cannot be measured across five days.

All of it was pasted as-is into a conversation, asking for both weeks as a table. No real names, no medical reasons, no patient data. We come back to that below, because the precaution was not decorative.

The answer arrived in about forty seconds. A clean table, one row per physician, one column per half-day, a paragraph of explanation underneath. In a screenshot sent to a colleague, it passes for an FMG schedule.

What it did well

It understood the request on the first try

No form, no imposed format, no columns to name. Constraints written in messy prose were read correctly, including the ones phrased in the negative and the one that depended on another. A manager who has spent fifteen years translating emails into spreadsheet cells recognizes the value of that immediately.

It asks the right questions when you ask it to

Asking it, before any schedule, what information it was missing produced a list many FMGs have never written down in full: the number of offices available per day, the rule that settles two physicians wanting the same block, what happens when a block stays uncovered, who has the final say. As an exercise in laying out the FMG's own rules, that is time well spent.

It writes the script nobody will write

Asking it not for the schedule but for the program that computes it produces code calling a real constraint solver. That version works, because the computation is handled by an engine that refuses an invalid solution instead of being written out word by word. It still requires someone in the FMG to run code, fix it when the rules change in November, and answer for what it produces.

Where the schedule broke

It does not reread what it just wrote

The physician on vacation for the second week appeared in the following Wednesday's walk-in block. Point it out and you get an apology and a corrected version, which fixes that case and creates another one three rows down, because the correction is not a recomputation of the whole grid. Two or three exchanges later, you no longer know which version is the right one.

That is the structural limit of the exercise. Text is produced left to right, one cell after another, with no step where the whole is confronted with the whole set of rules. An assignment engine does the opposite: it holds every constraint at once and refuses the impossible cell at the moment it is proposed.

A walk-in block is not a shift

It was treated as one cell to fill among others. Nothing in the answer says that an uncovered block is a different kind of problem from an empty office, that those hours count toward the FMG program's accessibility obligations, or that an approved withdrawal has to trigger the question of who takes the block back. Patients, meanwhile, show up at 1:00 p.m.

Walk-in distribution follows rules the model has no reason to know, since they were never written down anywhere in the public documents it learned from.

Five shifts each, and still no fairness

The proposed split was balanced in the most literal sense: the same number of shifts for everyone. A Friday evening and a Tuesday morning were worth the same there, as in nearly every generic tool. That is precisely the arithmetic that starts parking-lot conversations in an FMG, and getting out of it means giving shifts a value.

Offices do not exist in the answer

We had written three offices for seven people on Tuesday. The returned schedule placed all seven, without mentioning the constraint or flagging that it could not be met. An FMG almost always has more professionals than rooms, and placing offices at the same time as people is what prevents the hallway negotiation on Monday morning.

It states that everything is compliant

This is the part that costs the most. Ask it to reread its own schedule and confirm that it honours the constraints it was given, and it answers that everything is compliant, walking back through the rules one by one to prove it. Meanwhile the physician on vacation is still down for the Wednesday of the second week, in plain sight, in the table it has just validated.

A cautious silence would do less damage than that confirmation. A grid with three conflicts comes out with the same formatting and the same tone as a flawless one, no cell is marked doubtful, and a validation obtained on request is an invitation to skip the review. The only way to know what state the schedule is in is still to recheck all of it by hand.

For twelve physicians over two weeks, that review takes roughly as long as building the period directly would have. This is where the value of a tool is decided: in our pilot FMG in Saint-Hyacinthe, about thirty people managed in Excel took roughly 30 hours of management a month, post-publication fixes included, and around one hour with a dedicated tool. These are our own measurements rather than a network average, but they frame the order of magnitude that manual verification eats into.

The question that comes before the technical one

The test ran on initials because an FMG schedule is full of personal information. A physician's availability, absences, preferences and withdrawal reasons all qualify under Quebec's Law 25, whose main obligations have applied since 22 September 2023 and whose right to portability has applied since 22 September 2024. The absence of any patient record does not change that classification.

Before pasting the real list into a consumer assistant, three things deserve a written answer: where the conversation is retained and for how long, whether its content is used to train the model, and what contractual framework binds your FMG to the provider. A personal account opened in three clicks is nothing like a supplier the FMG has assessed and signed an agreement with.

What a chat thread cannot replace

Even assuming the schedule were perfect the moment it came out, it stays frozen in a conversation only the person who asked for it can read. An FMG schedule is built once a month and read every day by fifteen or so people who were not there when it was built. When a physician withdraws on a Friday afternoon, what matters is how fast the change reaches everyone.

Our eight moments that wreck an FMG schedule happen after publication in six cases out of eight. An assistant that produces a grid answers the part that already worked reasonably well in Excel, and leaves the painful part untouched.

What AI does well once it sits in the right place

The useful split runs between the interface and the engine. A language model is excellent at turning "I do not take walk-ins on Friday afternoons unless Marie is away" into a structured rule, with no form and no dropdown. It reviews well too, on one condition: the question has to be closed. "Give me every half-day where MD-04 appears" returns something usable, because finding a string in a table is a reading task. "Is this schedule correct?" returns the everything is compliant described above. Mechanical checking, one rule at a time, holds up; overall self-assessment does not.

The assignment itself needs something else. It needs an engine that holds every constraint at once and refuses, at the moment of the action, the double booking or the block left blank. Natural-language constraints in Synchro rest on exactly that division: each physician writes their constraint in plain language in thirty seconds, and the scheduling engine decides what can actually be placed.

Where to start if you want to test it anyway

1. Anonymize before pasting. Initials, never full names, never the reason for an absence. A constraint tests just as well with MD-04 as with the real name.
2. Ask for the list of rules, not the schedule. "What information would you be missing to build this schedule?" gives you a document your FMG can keep, which the table does not.
3. Have it check one rule at a time. Paste the schedule you already built and ask a closed question per constraint, along the lines of "give me every half-day where MD-04 appears". That is the one variant of the test that saved time. The open question about overall validity returns a confirmation worth nothing.
4. Time the review. If validating the output takes longer than building the period yourself, the answer has arrived and it is no.

Where Synchro fits

Synchro builds schedules for medical teams in Quebec using the objects of the job: walk-in blocks, weighted on-call shifts, shared offices, constraints written in plain language by each physician, transfers between colleagues with a record of who accepted and when. Natural language is the input, rules decide, and a change becomes visible to the whole team the moment it is made.

What Synchro does not do: no patient data flows through it and it does not replace your EMR, whether that is Ofys, Omnimed, Medesync or Kinlogix, all of which consume the schedule rather than produce it. No engine guesses a rule that was never written down either: what nobody has stated, neither ChatGPT nor Synchro will apply.

Further reading

• 8 moments that wreck your FMG schedule
• FMG schedule management: which tool to choose
• Best scheduling software for a Quebec FMG: 7 tools compared
• FMG scheduling software: the complete guide