Leave rules count days and hours. AI-assisted engineering has loosened the link between those counts and what a team delivers. This is what the research says about that, and a proposal that replaces "days off" with declared availability and a protected floor of real time off. It is the second of three articles on running a small team that works with AI, after the developer scorecard; the third covers pay.
One week away, nothing missed
This is a real account from one of our people, not a composite.
In the past few weeks they attended two industry events and spent five working days of one week out of the office. They still joined every meeting. The work was reviewed and committed as usual. They sent an hourly status report, so when anyone called, they could answer with a current status on everything in flight. Nobody on the team felt the absence.
Over the last fifteen days they were also sick for two. They kept working from bed, with a laptop running AI agents on their tasks. The only day they took fully off was the day their voice gave out. By "off" they mean one thing: nobody calls them.
On paper, the event week was work travel and the two sick days were sick leave. Neither label described what happened: the work kept flowing and the person never stopped. The one day that mattered, the day nobody called, has no label at all.
That is not a model to copy. It is a sign that the rules no longer describe the work. Our leave rules were built for a world where output equals hours at a desk. We think they should be rewritten around availability, and we propose how.
What the old rules measure
A leave request assumes that presence is capacity. A day away is a day of output lost, so it needs approval, a balance and a record.
That was a fair model when one person typed the work, one hour at a time. In a team where every developer works with AI, a day's output depends on how well the work was specified, how many agents ran, and how good the review loop was. Hours of human attention matter less than they did. Days and hours are inputs, and we are still managing people by inputs.
A caution before going further. One person's week is an anecdote. This person is senior, controls their own calendar, and had no new hires to mentor that week. Treat the story as a question, then test it against evidence.
What the research says
Hours were never a good proxy for output
John Pencavel studied British munitions workers in the First World War. Per summaries of the study, output moved in proportion to hours up to roughly 49 a week, then rose at a falling rate. Output at 70 hours differed little from output at 56, and working six days produced slightly more than seven. These are factory workers a century ago, so read the shape, not the numbers.
The health side is stronger. A WHO and ILO analysis estimated that long working hours were linked to 745,000 deaths from stroke and heart disease in 2016. Working 55 hours or more a week carried an estimated 35 percent higher risk of stroke, and 17 percent higher risk of dying from heart disease, than working 35 to 40.
Outcome-based schedules hold up
Erin Kelly and Phyllis Moen studied a results-only work environment at Best Buy's head office. Over eight months, about 6 percent of participants left, against about 11 percent of a comparison group, a 45 percent lower turnover after controls. The much-quoted productivity gains come from the programme's creators, not from peer review, and Best Buy dropped the programme in 2013.
Nicholas Bloom and colleagues randomly assigned 1,612 employees at a travel technology company to office-only or hybrid work for six months. In their Nature paper the hybrid group quit about a third less, and the software engineers wrote about the same amount of code in both groups. It is one company, but it was randomized.
The four-day week trials point the same way, with a caveat. In the 2022 UK trial, 56 of 61 companies were still on the shorter week afterwards, and 18 had made it permanent. Participants volunteered, the results are largely self-reported, and the trial was run with a campaign group. Iceland's trials covered more than 2,500 workers and reported stable or better productivity, but the report was written by a think tank that campaigns for shorter weeks.
Together, these suggest that when the unit of control moves from hours to results, people stay and output does not visibly fall. They do not test leave.
Unlimited vacation is not the answer
The obvious fix is to drop the balance. The evidence on that is poor. Namely's own data disagree across years: unlimited plans were reported at 13 days against 15 for traditional ones in 2018, and 12.09 against 11.36 in 2022. A Resume Builder survey found 12 percent of unlimited-plan workers took no days off in a year, against 4 percent on set plans. Being too busy was the most common reason. Not wanting to look lazy came from 26 percent.
Basecamp dropped its unlimited policy after it caused more stress than flexibility. GitLab keeps flexible time off but strongly recommends at least two consecutive weeks and 25 days a year, and handles sick time separately. Netflix is reported to have stopped tracking vacation in 2004, but a company with a strong performance culture is a hard case to copy.
No limit with no floor drifts towards no leave.
Working sick costs more than it feels like
This is where our colleague's story needs the most care. Remote work makes working through illness easier. A 2023 cross-sectional survey of German remote workers found about four days per respondent of working while ill, weakly linked to poor ability to switch off and weak manager support. Survey links like these do not show cause. A Trinity College Dublin study of remote workers found they were more likely to work while sick when they had not met their daily goals, and that it hurt their performance the next day.
So working from bed can look like dedication on the dashboard and cost the next day. A sick-day rule has to protect the person from their own goals.
Rest is not the same as absence
Sabine Sonnentag's research on recovery points to psychological detachment, actually stopping thinking about work, as the most consistent recovery factor. Reviews of vacation studies find the benefits depend on detaching, not merely being away. In a study of teachers by Kühnel and Sonnentag, which we read second-hand, the benefit of a vacation had faded within about four weeks of returning.
This is the hidden cost in the opening story. The week at events was not rest. It was work in a different building. A team can run well on that for a while, but the person's recovery account is not being refilled. A dashboard of commits will not show the overdraft.
AI changes the pressure, not the need
The hope was that AI would free time. An eight-month study of about 200 employees at one US technology company, published in Harvard Business Review in February 2026, found the opposite: people worked faster, took on broader tasks and extended work into more hours. The authors describe it as ongoing research, and one company is a small sample, but it matches what we see. Freed time gets filled, and the filler is rarely rest. This is why a leave model that depends on goodwill alone will lose to workload creep.
The law is moving the same way
France introduced a right to disconnect in 2017. Belgium, where we have a presence, has had one for private employers since 1 April 2023, but only for employers with at least 20 employees. Australia added one from August 2024, with small businesses following later. Details differ in each, as comparisons by law firms show. In 2019 the EU Court of Justice ruled in case C-55/18 that employers must have a reliable system to record daily working time. EU law also guarantees at least four weeks of paid annual leave. So the direction is both ways: more protection for off time, and more recording, not less.
Why people push past their limit
We have said that people work while sick. We have not said why. Dedication is part of it. In our experience two other causes are common, and both are in the company's hands.
Unclear career and raise milestones
If an employee cannot see what separates their level from the next one, or what earns a raise or a bonus, they will guess. The usual guess is "more". More hours, more weekends, more visible effort, including the sick days worked from bed. They are trying to meet a requirement that nobody has written down.
The evidence here is thin and mostly opinion or survey. A TU Delft piece by the board of its young academy describes unstable promotion criteria and nights and weekends becoming the norm. That is the board's reading of its own survey, not a measured result, and it concerns academics, not software teams. A stronger source is the job demands and resources literature. A meta-analysis by Crawford, LePine and Rich found that demands go with burnout and resources go against it. Role ambiguity is a hindrance demand in that literature: it goes with more burnout and less engagement. We read that paper only through its abstract.
The fix is cheap: write down, for each level, what the person must deliver and when it will be reviewed. Put the raise and bonus milestones next to it. If the answer is "it depends on the budget", say that too. An employee who knows the target can stop padding it.
Unclear direction from management
The second cause sits one level up. When the roadmap is vague, milestones move and deadlines appear without warning, people cannot tell how much work is enough. They protect themselves by doing everything, always. In startups this is common, and the people at the top often feel it too.
Our recommendation is to avoid both extremes. Do not over-plan: a startup's plans change every few weeks, and a six-month roadmap is fiction by month two. Do not leave it unplanned either, because then every day is an emergency. Plan in a short rolling window of two to four weeks, and refresh it on a fixed day. Each person should be able to answer three questions at any moment: what is expected of me in this window, what is the next deadline, and what is not expected.
Basecamp's Shape Up method uses six-week cycles with a two-week cool-down. We suggest a shorter window, but the principle is the same: a fixed horizon and a planned break. We found no startup-specific evidence on how well a rolling window works, so we treat it as a practice to try, not a proven rule. Off days are easy to plan inside a plan, and hard to defend inside a permanent emergency.
A third cause, harder to fix
When AI makes output easier to compare between people, some work longer to stay visible. This is our observation, not a finding, and no policy removes it. What helps is being plain about how people are judged, which is the subject of our next article.
What the evidence leaves open
We could not find a study of leave policy for AI-assisted software teams. Most of what is above is self-reported, comes from volunteers, or comes from other industries. The material on career clarity and planning is mostly survey data. None of it tests the model below. We offer it as a design to pilot, not a finding.
The proposal: declare availability, protect Off
Start with clarity. Every person needs a written definition of what the next level and the next raise require, and a rolling plan of two to four weeks that says what is expected and what is not. Without that, people will use Light and sick days to prove commitment, and the model below fails.
Then replace the question "how many days did you take?" with "what state are you in today?" Three states cover it.
| State | What it means | Who can contact you |
|---|---|---|
| Working | A normal day, from anywhere. Reachable in agreed hours, work flowing. | Anyone, in normal hours |
| Light | You choose it. Reachable on a stated schedule, fewer meetings, mostly asynchronous. For events, travel, a day when you are slow but well. | Only per your stated schedule |
| Off | Nobody calls. No meetings, no reviews waiting on you, no replies expected. | Only a named emergency path |
Seven rules make it work.
1. Off has a floor, not a cap. Each person gets a minimum number of Off days a year, including one long stretch. As a default we propose at least 20 working days Off, or the statutory minimum if it is higher, with one stretch of at least ten working days in a row. Some countries, such as France, have their own rules on the main block. We chose 20 because it matches the legal floor in many places and is easy to defend. GitLab's recommended 25 days is a better target for a team that can afford it. Light days never count toward the floor. For someone on a part-time contract, the floor is counted in their contracted days: a 60 percent contract gets at least 12 Off days a year, never less than the statutory minimum.
2. Off is requested early and rarely refused. You tell the team in advance and name your coverage. A manager may move the dates only for a named delivery reason, and must offer new dates inside the same month.
3. Light is declared, not granted. A person opts in; a manager cannot assign it. Light is for days when work can move, not an expectation that it must. Someone who cannot work from a conference hall or a sickbed is not penalized for it.
4. Sick is its own category. It is never taken from vacation. A sick day is Off. If a person feels well enough to do some work, they record a Working or Light day instead of a sick day; they cannot be both. When someone says they are unwell, the manager's first move is to lower the day's goal. Where a medical certificate or national sick-pay rules apply, those rules decide.
5. The work is the evidence, not the person. State changes and output are visible through commits, reviewed changes and a short written status at an interval the person picks. Our colleague chose hourly. We would default to a daily note. This is not keystroke monitoring. Light is paid working time, and the status note does not replace the legal time record. This also fits our developer scorecard: the scorecard measures output, leave measures availability, and neither should be used as a proxy for the other.
6. AI gains are a decision, not a formula. Set each person's quarterly capacity before adopting a tool, and do not raise it automatically when AI makes people faster. Once a quarter, the team decides explicitly how much of any speed-up goes to more scope and how much to time. Writing the decision down matters more than the ratio. If a team wants a starting point, half and half is an illustration, not a recommendation.
7. Off has one emergency path. A named on-call person, paid or compensated according to local rules, is the only route in. After a night-time page, that person takes the whole next working day Off; chapter 8 of this playbook, on on-call, sets out the rota and the alert budget. Anyone else who calls an Off colleague has broken the policy.
How to test it
Run a 90-day pilot with one team. Track five things: delivery against the plan, defects after release, Off days taken against a pro-rated floor of five per quarter, sick days, and a short monthly check on whether anyone feels pressured to stay in Light. The pilot has failed on its own terms if Off days taken fall below the pro-rated floor, or if anyone spends more than half of their days in Light over a month. Remote work is not one size fits all, and this model should not be either.
Where this breaks
New hires and juniors. They learn from overlap. In the first 90 days we would keep them mostly in Working, while still guaranteeing Off days.
Customer-facing and on-call roles. Coverage is part of the job, so Off needs a rota and a recovery rule, not just a status.
Manager abuse. A manager can keep everyone in Light forever and call it flexibility. The check is Off days taken, reported monthly, to someone other than that manager.
Hourly reports as anxiety. If a person needs to send hourly proof to feel safe, the culture may be the problem. The report is the person's tool, never a requirement. Pair it with the habits in leading engineers without reading code.
The legal floor
This model sits on top of the law, not instead of it. Five points need local review before anyone adopts it.
- Light is paid working time and is recorded as working time under the EU ruling above. The status note does not replace the time record.
- Evenings, weekends and public holidays are Off by default. Daily and weekly rest rules still apply.
- Waiting on call can legally count as working time. "Paid or compensated" needs local advice.
- The model is for employees. For contractors, a firm that controls their availability and leave can look like disguised employment, in Belgium and elsewhere. Contractor terms belong in the contract.
- Off days are paid leave and follow national rules on accrual, carry-over and payout.
This is not legal advice.
What to do tomorrow
- Write one sentence that defines Off for your team, such as "nobody calls", and publish it where people will see it.
- Count last year's Off days that each person actually took. Anyone below the floor you would pick is a signal, not a performance problem.
- Split sick time from vacation in writing, and state that a sick day is Off.
- Write the next four weeks of expectations for each person on one page, and write down what the next level and the next raise require.
Next: pay in the AI era
If output no longer tracks hours, pay should not rest on hours alone. In the next article we take this one and the developer scorecard and ask what a salary and raise structure looks like when AI changes what one person can deliver.
Frequently Asked Questions
Is this the same as unlimited vacation? No. Unlimited policies remove the cap and, in the survey data we found, can lead to people taking less. This model keeps a protected floor and defines what Off means, which most policies leave out.
Can employees work while sick? Not as a sick day. If they feel well enough to work, they record a Working or Light day. The research we found suggests that working while ill, especially from home, tends to hurt performance the next day, so the safe default is that a sick day is Off.
Does this work with EU labor law? It is meant to sit on top of it, not replace it. EU rules on working-time records, rest and annual leave still apply, and national laws differ. This is not legal advice, so have local counsel review any change.
How do you know people are really working? By their output: reviewed and committed work, delivery against the plan, defects after release. For a small team this is more reliable than hours. If you cannot tell from the work whether someone is contributing, that is a visibility problem to fix whether or not you change leave.
Rethinking how your engineering team works, with or without AI? Book a diagnostic call or see how our Fractional CTO service can help.