A staff rota that meets the 1,785-hour limit

Reading time: about 10 minutes

The situation. Every year the rota gets drawn on a sheet of paper, by eye, dragged over from last year. And every year somebody ends up with too many hours and somebody with too few, and it comes out in November.

Here the AI does the boring part — distributing shifts and adding hours — and you do what it cannot: check the result against the collective agreement. It is the most delicate case in this block because a bad rota is a claim waiting to happen, so the weight here is on the checking, not on the generating.

What you need to hand: How many people and their contracts (full time or the part-time percentage), your opening hours, and the figure that rules: the Spanish national agreement’s 1,785-hour annual maximum. If you are under a provincial agreement, check yours first — they do not all match.

Step by step

Give it the hard rules before the problem

A rota is a constraint problem, and a model respects the constraints you write down and no others. If you do not mention weekly rest, it will happily produce seven-day weeks.

The non-negotiables, all of which go in the prompt: the annual cap, rest between shifts, weekly rest, each person’s holidays, and hours already worked or committed.

And do not give it names. "Person 1, person 2" works exactly the same and saves you sending a third party who works in your pharmacy and when. You add the names at the end, on your own sheet.

Help me build the annual rota for a pharmacy in Spain.

STAFF (no names):
- Person 1: full time
- Person 2: full time
- Person 3: 60% part time
- Person 4: 50% part time, afternoons only

OPENING: Mon-Fri 9:00-14:00 and 17:00-20:30; Sat 9:30-14:00
Closed Sundays and 12 public holidays.

RULES THAT CANNOT BE BROKEN:
- Maximum 1,785 hours a year for a full-time contract. Part-timers,
  pro rata.
- Minimum 12 hours’ rest between the end of one shift and the start of
  the next.
- One and a half consecutive days of weekly rest.
- 30 calendar days of holiday per person.
- There must ALWAYS be at least one pharmacist on during opening
  hours.

Give me:
1. The typical weekly distribution, as a table.
2. The ANNUAL HOUR TOTAL each person ends up with under that pattern.
3. The gap between that total and their maximum, in hours.
4. Any week where one of the rules above is broken.

Show me how you calculated the annual hours. If a constraint is
impossible to meet with this staff, SAY SO instead of inventing a rota
that looks like it adds up.

Add the annual hours yourself. Always.

This is the check you never skip here, and it is three lines of arithmetic: hours in the typical week × weeks actually worked. Weeks actually worked are not 52: holidays come off.

With 30 calendar days of holiday you get about 47.7 working weeks. If the typical week is 37.5 hours, 37.5 × 47.6 ≈ 1,785. That is the reconciliation, and it has to come out for you too.

If the model gives you a total you cannot reproduce with that multiplication, do not accept it however good it sounds. And watch for the opposite, which is the commonest: totals well below the cap, meaning weeks were left out of the calculation.
The three figures that must reconcile
- Hours in the typical week. Add them from the table, cell by cell.
- Weeks worked. 52 minus holidays. Not 52.
- The product. Must land at or below the cap, never above.
Check the part-timers separately

Part-time contracts are where it goes wrong most, and in a specific way: it applies the percentage to the weekly hours and then multiplies by full-time weeks, or the other way round.

The check is the same one, done against their own cap: 60% of 1,785 is 1,071 hours. If the model gives them 1,150, that is not rounding: it calculated by a different route.

And there is an agreement detail that is not arithmetic and the model does not know: on part-time contracts the night supplement is not pro-rated. If your rota has shifts running into night hours, that is paid separately and does not come out of this table.
Move it to your own sheet and put the names there

Ask for the table as CSV or Markdown and paste it into your spreadsheet. That is where the names go, where each person’s specific requests go, and where the real rota lives.

This is not only about privacy: the rota will change twenty times during the year, and your own sheet with a running-hours column tells you in March what used to surface in November.
Before you pin it up: what a person decides, not a model

A rota can meet the hours and still be a bad rota. None of this is in any number and you know all of it: who cannot do Tuesday afternoons, who has done every Saturday for three years, and that two particular people do not work well together on a bank-holiday Friday.

The AI gave you the legal skeleton. The version you pin up you adjust yourself — and then you add the hours again, because every adjustment moves the total.

And if there is staff representation or an internal agreement on how shifts are shared, that overrides any rota a model generates. This is not a technical check: it is whose decision it is.

A full example, and what is wrong with it

The four people from step 1 — two full time, one at 60% and one at 50% afternoons — with the opening hours given. This came back.

What it replies

Calculated annual totals:

Person 1 (full time): 37.5 h/week × 47.6 weeks = 1,785 h. Margin: 0 h.

Person 2 (full time): 37.5 h/week × 47.6 weeks = 1,785 h. Margin: 0 h.

Person 3 (60%): 22.5 h/week × 47.6 weeks = 1,071 h. Margin: 0 h.

Person 4 (50%, afternoons): 18.75 h/week × 47.6 weeks = 892.5 h. Margin: 0 h.

"No week breaches the stated rules. The distribution respects the 12-hour rest and the day-and-a-half weekly rest."

And now, what is wrong with that reply:

The arithmetic is impeccable, and that is the problem. 37.5 × 47.6 = exactly 1,785. The part-timers, exactly pro rata. Everything matches to the decimal. And a margin of zero on all four is the signal that something is wrong: a real rota never lands exactly on the cap, because there are absences, swaps, training and busy weeks.
With no margin, the first absence puts you in breach. If somebody gets flu for a week and another covers, that other goes over 1,785 and there is nowhere left to compensate: the year is full. A rota at 100% of the cap is not a tight rota, it is one that cannot survive reality. A sensible buffer is 40 to 60 hours per person.
The 47.6 weeks assume everybody takes holiday at once. 30 calendar days is 4.3 weeks, and 52 − 4.3 = 47.7. Correct for one person. But in a pharmacy holidays are spread across the year, and the weeks somebody is away have to be covered — so the others’ effective weeks are not 47.6. The per-person calculation is right and the whole does not close.
And "no week breaches" is a statement, not a calculation. The prompt asked for "any week where one of the rules is broken". It answered that there are none, without showing the sweep. It may be true; you do not know. Ask explicitly for "show me the week with the least rest between shifts and the one with the least weekly rest" — with day names. If those two comply, the rest do.

None of the four problems is arithmetical, which is why none shows up when you check the sums. They show up when you ask whether the rota survives a real year, a question the model does not ask itself.

Ask again with two more things in the prompt: "leave a 50-hour margin per person" and "holidays do not overlap, spread them between June and September". You get a worse rota on paper and a much better one in January.

And if your pharmacy is not like that

If you do out-of-hours duty. They go separately and you must say so, because an on-call shift is not just a longer normal day: it has its own subsequent rest rules and its own pay. Put it in as an explicit constraint — "Saturday duty is covered by one person and they are off on Monday" — or the rota will add up and be unusable.

If somebody has reduced hours for childcare. It is not an ordinary part-time contract: the person chooses their hours within the ordinary working day, and that is a hard constraint the model will treat as flexible unless told. Write it with the actual hours, not the percentage — "9 to 2, no afternoons" works; "70%" does not.

If you want to try new opening hours before changing them. This is one of the most useful things this case does: ask for the same rota with the hours you are considering and compare the staffing needed. Sometimes opening half an hour later means 180 annual hours you do not have, and that is visible before announcing it rather than after.

If you are not under the national agreement. The cap changes and sometimes the rest rules do too. Look it up in your provincial agreement and put it in the prompt yourself: the model will give you an equally confident number if you do not, and it will be the national one.

When it does not work first time

It says the hours cannot be covered with this staff.

That is a good answer, not a failure — the prompt asked for it explicitly. And it means one of two things: either you are short-staffed, or one of the constraints you set is harder than you thought. Ask "tell me how many hours are missing and which constraint would have to give" — both answers are actionable.

The rota adds up but is horrible to work.

Normal: it optimised coverage and nobody told it anything about how it feels to work. Add explicit human constraints — "nobody does more than two closes in a row", "Saturdays rotate" — and ask again. They are nearly always the missing ones, and they do not appear by themselves because they are in no regulation.

I do not trust the hours it gives me.

Quite right, and that is why step 2 is adding them yourself. The check is three figures: typical-week hours, weeks worked, and the product. If you cannot reproduce its total with that multiplication, do not accept it however pretty the table.

I change one thing and have to redo everything.

That is why step 4 says move it to your own sheet. The rota will change twenty times this year and you will not go back to the chat twenty times: what you need is a sheet with a running-hours column that recalculates itself. The model gives you the starting point, not the tool.

Before you use what it gave you

You added the typical-week hours by hand for each person.
You subtracted holidays before multiplying by weeks.
Nobody exceeds their cap: 1,785 h, or their pro-rata share.
Part-timers are calculated against their own cap, not the full-time one.
There is not a single name in what you sent to the chat.
You adjusted the final version yourself with what the model cannot know.

It multiplies by 52 weeks. The classic and easiest to catch: if holidays are not deducted, the annual total comes out about 160 hours too high. That is why the check is three figures, not one.

It says it complies without having added up. "This rota respects the annual maximum" is a sentence, not a calculation. That is why the prompt asks it to show its working: a claim with no arithmetic behind it is worth nothing here.

It does not know your provincial agreement. The 1,785 hours come from the national agreement. Under a provincial one the figure may differ — and the model will give you a number just as confidently. You supply the cap.

It spreads public holidays as though they were all alike. Local, regional and national holidays are treated differently and sometimes require opening. The model treats them all as closed days and deducts the same for each. If your pharmacy opens on some, list them one by one — it is tedious and it is the only way the annual total means anything.

Now, with your own. Take the rota currently pinned up and do just the check: typical-week hours × weeks worked, person by person. Five minutes. You will quite likely find somebody out of balance — and finding it in January is a conversation, finding it in November is a problem.

← All practical cases Back to the school →

A staff rota that meets the 1,785-hour limit

Step by step

Give it the hard rules before the problem

Add the annual hours yourself. Always.

The three figures that must reconcile

Check the part-timers separately

Move it to your own sheet and put the names there

Before you pin it up: what a person decides, not a model

A full example, and what is wrong with it

And if your pharmacy is not like that

When it does not work first time

Before you use what it gave you