The Cap You Can Only Test Afterwards
Japan caps overtime with four tests running at once — a monthly limit, an annual one, a rolling average across two to six months, and a count of how many months in the year went over. Only the first of those can be answered by looking at a month.
Most working-time rules are a number and a period: so many hours, per week or per month. You total the period, you compare, you are inside or outside. Japan's upper limits on overtime are not shaped like that, and the difference is the reason a payroll report can show every month inside the limit while the arrangement as a whole is outside it.
This matters beyond Japan, because the shape is spreading — a rolling multi-period average is now a common way to write a working-time rule, and almost no operational system is built to evaluate one. It is worth understanding on its own terms.
Four tests, not one
Article 36 of the Labor Standards Act is the provision that permits overtime at all, by agreement — the agreement everyone calls a 36 agreement. The limits attached to it read as follows.
| Test | The limit | What period answers it |
|---|---|---|
| The principle | 45 hours per month and 360 hours per year — a lower pair, 42 and 320, applies under certain variable working-hour arrangements | A month, and a year |
| The annual ceiling under a special clause | Fewer than 720 hours in the year | A year, and only at the end of one |
| The single month | Below 100 hours, counting extended work and day-off work together | A month — but a different total from the first row |
| The rolling average | An 80-hour average across rolling periods of two to six months, again counting both | Every window of two, three, four, five and six consecutive months |
Read the right-hand column downwards. The four tests do not share a period, and two of them do not share a definition of what is being counted — the principle limit is about extended working hours, while the 100-hour and 80-hour tests count extended work and day-off work together. A single monthly overtime figure answers one of the five constraints and is the wrong input for two of the others.
A rolling two-to-six-month average is not a report you run monthly. It is five overlapping windows that all have to be true at once, and a month that is fine on its own can be the month that breaks three of them.
The arithmetic that catches people
Take a person at 78 hours in each of four consecutive months. Every month is below the single-month test. The four-month average is 78, below the rolling one. Nothing has gone wrong. Now the fifth month is 95: still below 100, so the single-month test is fine, and the running annual total may be nowhere near the ceiling.
But the two-month average across months four and five is 86.5, and that is the test that has failed — on a month where every individual number was inside its own limit. This is not a trick example. It is the ordinary shape of a busy period after a quiet one, and it is invisible to any process that looks at one month at a time.
One more thing worth knowing about the current state, because it changed recently: occupations that had been given a grace period — construction work, automobile drivers and physicians — came inside the upper-limit rules from April 2024. Those are three sectors where a busy month after a quiet one is the normal operating pattern rather than the exception, which makes the rolling test the binding one rather than a technicality.
What a system actually needs to answer this
Reduced to data, the requirement is unglamorous, and the useful news is that most of it is ordinary record-keeping rather than anything clever.
-
A daily figure per person, not a monthly one
Every derived total in the rules is built from days, and the two halves — extended work and day-off work — have to stay separable, because one test counts them together and another does not. A system that stores only a monthly total has thrown away the distinction before anybody asked for it.
-
A running total inside the current month
This is the one that actually prevents a breach rather than recording it. Knowing on the eighteenth that somebody is at forty hours is an operational decision anybody can act on. Knowing on the second of next month is a note for the file.
-
A window evaluated across months, not a report per month
The rolling average has to be computed over every window of two to six consecutive months ending in the current one. It is five sums, it is not hard, and it is simply a different query from the one a monthly report runs.
-
A counter of months exceeded, held per year
The six-months-a-year constraint is not an hours figure and nothing that totals hours will produce it. It is a count, reset annually, incremented when a month goes past the principle limit.
-
The ability to reproduce last quarter's answer
A rolling test is a statement about a window that has closed. If the underlying records can be edited without a trace, the same window computed twice gives two answers and neither is defensible.
Where we are on that list, item by item
Our overtime record is per employee per day, carrying the hours, a reason, an approval status and who decided it — the right granularity for all five tests, which is the part that is genuinely hard to retrofit. It is derived from attendance against the shift, and the derivation runs monthly, for the previous month, on the second — so the figure exists two days after the month it describes, which makes it a record rather than a control. Our HR alerting evaluates five targets — attendance rate, leave backlog, approvals over their service level, attrition and overdue probations — and overtime is not among them. So a running in-month total, a rolling window and a months-exceeded counter are all builds on top of data that is already the right shape. For the separate question of shift design and premium rates, our Kenyan piece covers the other half of overtime.
None of this is a reading of anybody's obligations. A 36 agreement is a negotiated instrument, the limits interact with the arrangement you have actually agreed, and whether a given pattern is compliant is a question for a Japanese employment lawyer rather than for a product blog. What is worth taking from it is the question, which travels: is the rule we are being measured against a total, or a window?
Four questions for any system that claims to manage working time
Can you show me a person's overtime for the current month, today?
What you will hear
Often a report that runs after month end.
How to read it
Then the system records breaches rather than preventing them. This is the single most useful thing to establish, and it is answerable in a demo in about ten seconds.
Is day-off work stored separately from extended hours?
What you will hear
Usually one number.
How to read it
One number cannot serve two tests that count differently. If they are merged at capture, no amount of reporting separates them afterwards.
Compute the four-month average ending last month.
What you will hear
An export, and a spreadsheet.
How to read it
That is a fair answer and it tells you the rolling test lives outside the system. Fine at forty people; ask what it looks like at four hundred, across five windows, every month.
If an attendance record from March is corrected in July, what happens to the March window?
What you will hear
Silence, usually.
How to read it
A rolling test is a claim about a closed window, so a silent recomputation and a silent non-recomputation are both wrong. What you want is a record that the window moved and why.
Our take
The transferable lesson has nothing to do with Japan: before you configure anything, work out whether the rule binding you is a total or a window. A total is a report. A window is a different data model — daily granularity that keeps its categories separate, a running figure inside the current period, and the ability to recompute a period that has already closed and say why the answer changed. Almost every system handles the first and quietly assumes it has handled the second. If you are subject to a rolling test, the cheapest thing you can do this week is stop asking what last month totalled and start asking what the last six months average, because that is the number that is actually being tested.
What AWRA OpsHub does today
- Overtime recorded per employee per day, with the hours, a reason, an approval state and the person who decided it — the granularity every rolling test needs.
- Derivation from attendance against the shift, run for a whole month at a time and idempotent, so re-running it never duplicates a day already recorded.
- An approval trail on each record, so an hour that counts toward a ceiling is one somebody accepted rather than one a clock inferred.
- HR targets evaluated daily with in-app, Slack and throttled email alerts, on attendance rate, leave backlog, approvals over their service level, attrition and overdue probations.
More we can add to your workspace
- A running overtime total inside the current month, visible per person while the month is still in progress and therefore still changeable.
- A rolling multi-month average across every window of two to six months, evaluated as a standing check rather than assembled from an export.
- Configurable ceilings per workspace — a monthly figure, an annual figure and a count of months allowed to exceed the first — with a breach raised the way the other HR targets are.
- A point-in-time reproduction of a closed window, so a corrected attendance record shows which past figures moved and why.
Where we point you to a specialist
- We will not tell you whether an arrangement complies with a 36 agreement or with the upper-limit rules. That turns on the agreement you negotiated and on facts about your workforce, and it belongs to a Japanese employment lawyer who carries responsibility for the answer.
- We will not infer that hours were worked where the day carries no record, in either direction. A day with nothing recorded against it is a question to ask somebody, and a system that answers it silently has manufactured evidence about a person's working time.
The first two lines are the same build at two scales: a running figure inside the month is the query that prevents a breach, and the rolling window is the same arithmetic pointed backwards over five overlapping periods.
What we can build for Japan on top of the standard product
Everything listed above as something we can add describes what ships in the standard product today — it is a starting point for Japan, not a limit on what AWRA OpsHub can do there. Kenya's eTIMS integration and its maintained payroll engine are in the product because Kenyan clients needed them and commissioned them; neither appeared by itself. The same door is open here. If rounding at the currency's own precision, consumption tax subtotalled per rate, a registration number field, a bank or mobile money feed, a statutory return format, a rule specific to how your operation runs, or a link to a system you already have is what stands between you and a decision, tell us and we will scope it as a build — written spec, timeline and price — before you commit to anything.
The arithmetic before the document
Japan is the market where our own arithmetic is the first thing to fix rather than a field we are missing. The invoice, quotation and point-of-sale paths round money to two decimal places as a hardcoded literal, and the yen has no minor unit — so a three-line invoice at the standard rate produces a consumption tax of ¥423.4, an amount that cannot be invoiced or paid. We already hold the correct number of places for every currency as reference data; that code simply does not read it. On top of that sits the requirement that gives the fix its shape: a qualified invoice must show consumption tax categorized by tax rate, and the rounding is permitted once per rate rather than once per document. Our invoice carries a single tax figure with the rate on the line, so the per-rate subtotal in between does not exist — and it is the same piece of work as the rounding, which is why we would build them together rather than in sequence. Separately and smaller: a tax registration number on the organization, on customers and on suppliers, since no column for one exists anywhere today.
Banks, payments and a currency with no decimals
Bank statement feeds and local payment rails wired into the Payments Register, with documents raised and reported in yen at the precision the yen actually has. What we will not do is treat our copy of the public register of qualified invoice issuers as authoritative for your credit entitlement — we will hold the number you recorded and the date you checked it, and leave the checking where it belongs.
The operational work, which is what most commissions actually are
An extra approval stage in a chain that does not match the standard one, a custom field set on employees or assets that only your sector needs, an expiry that has to block an order rather than send an email, a report your board asks for in a shape nothing produces, or a scanner or weighbridge feeding the goods-in door. These are the commissions we are asked for most often and the smallest ones we quote — and unlike a revenue-authority pipeline, none of them waits on a regulator.
Payroll and statutory returns
A Japanese payroll engine with income tax withholding, the standard-remuneration social insurance grades and the year-end adjustment, computed on live employee records rather than rebuilt in a spreadsheet each December.
Systems you already run
The accounting package, CRM, online store or custom database you intend to keep — connected through our API so a fact is entered once and appears everywhere it is needed.
How it works: you describe the requirement, we return a written scope, timeline and cost, and once agreed it is built into your environment and maintained as part of the product. No roadmap slide, and no pretending in a demo that something exists when it does not.
Tell us what you need integratedSee attendance, shifts and overtime as one record
Attendance against a shift, overtime derived per day with an approval trail, and HR targets evaluated daily with alerts to the people who can act on them.
Explore attendance and timeFrequently asked questions
What are Japan's overtime upper limits, in short?
Article 36 of the Labor Standards Act permits overtime by agreement and attaches limits to it: as a principle, 45 hours a month and 360 hours a year, with a lower pair of 42 and 320 under certain variable working-hour arrangements. Where a special clause is agreed, the annual figure is fewer than 720 hours, a single month must stay below 100 hours counting extended and day-off work together, the average across rolling periods of two to six months must not exceed 80 hours on the same combined basis, and the 45-hour monthly limit may be exceeded in at most six months of the year. Whether any of that applies to a given arrangement is a question for a Japanese employment lawyer.
Why can a month be inside every limit and still cause a breach?
Because the binding test is a rolling average rather than a monthly total. Four months at 78 hours each are individually fine and average fine. A fifth month at 95 is still below the single-month figure, but the two-month average across months four and five is 86.5, which is over the 80-hour rolling limit. Nothing in that sequence is unusual — it is a busy period after a quiet one — and it is invisible to any process that examines one month at a time.
Which occupations changed in April 2024?
Construction work, automobile drivers and physicians had been given a grace period from the upper-limit rules, and that period ran to 31 March 2024. From April 2024 they came within them. It matters more than a date usually does, because in all three the normal operating pattern is a heavy period following a light one — which is precisely the pattern the rolling average is designed to catch, so for those sectors the rolling test is the binding one rather than a technicality.
Does AWRA OpsHub evaluate these limits?
Not today. What it holds is the right raw material: overtime recorded per employee per day, with hours, a reason and an approval trail, derived from attendance against the shift. That derivation runs monthly for the previous month, on the second, so the figure is a record rather than a control, and our daily HR alerting covers five targets — attendance rate, leave backlog, approvals over their service level, attrition and overdue probations — of which overtime is not one. A running in-month total, a rolling window and a months-exceeded counter are builds on top of data that is already shaped correctly, which is the half that is hard to retrofit.
Is this only relevant if we operate in Japan?
The rules are, the shape is not. A rolling multi-period average is an increasingly common way to write a working-time rule, and it demands a different data model from a simple total: daily granularity that keeps categories separate, a running figure inside the current period, and the ability to recompute a period that has already closed and explain why the answer moved. Whatever jurisdiction you are in, the useful question is whether the rule binding you is a total or a window — because almost every system handles the first and quietly assumes it has handled the second.
