When an attendance record has a start and no end, the hours worked that day are not data any more. They are somebody's decision. In our own production data that happens on 6,110 of 40,970 records, which is 14.9 per cent, or about one shift in seven.
The number that makes it interesting is the other one. Records with a clock-out and no clock-in: zero.
Why the asymmetry matters
If terminals were dropping punches at random you would see both kinds in roughly equal quantities. Some shifts would lose their opening punch, some their closing one. That is not what the data shows. It shows every single incomplete record missing the same end.
So this is not a hardware problem. It is a behavioural one, and once you see it that way the cause is obvious.
Clocking in is load-bearing. It is how you are recorded as present, how the shift starts, and in many workplaces how the door opens. Skip it and something immediately does not work.
Clocking out is the last act of a shift, performed by somebody who is already leaving, and nothing visible happens if they skip it. The consequence lands two weeks later on a payslip, on somebody else's desk, in a queue.
Every incentive points one way. One shift in seven is what that looks like at scale.
What it costs
The hours become an opinion. Somebody in payroll decides what time that person left. They will be reasonable about it, and they will use the roster, and they will usually be close. But the number on the payslip was produced by a judgement, not a measurement, and the employee has no way to check it.
It concentrates in the wrong places. Missing clock-outs are not evenly distributed. They cluster on shifts that end at awkward times, at sites with a single exit point that gets congested, and among people whose shift end varies. Which means the same individuals are repeatedly having their hours estimated.
It undermines everything downstream. Overtime is computed from hours worked. If one shift in seven has estimated hours, then one shift in seven of your overtime is estimated too. In our data 15,651 overtime hours were recorded across the period, at real cost, and a share of that rests on inference.
The part that matters in a dispute
This is where an incomplete record stops being an inconvenience.
If an employee disputes their pay, the employer's position rests on the attendance record. The strength of that position depends almost entirely on one question: how hard would it be for the employer to have changed this?
A record captured by a biometric terminal, where manual entries are visibly marked as manual, is a strong document. In our data 140,764 of 140,775 punches came from a terminal and 11 were entered by hand. That ratio is itself the argument. Hand-entry exists, because terminals fail and people forget, and it is rare enough and visible enough that its presence is meaningful.
The failure mode to avoid is the one that feels most helpful: automatically closing an open shift and writing the inferred time into the record as though it had been captured. It cleans up the report beautifully and it destroys the evidential value of every record in the system, because from that point on nobody can tell which times were measured and which were assumed. A tribunal does not need to prove your specific record was wrong. It only needs to establish that your records generally cannot be distinguished from estimates.
So: infer freely, and never disguise an inference as a measurement.
What actually reduces it
Make the last punch unavoidable, if the site allows it. A door that requires a badge to exit produces complete records without asking anybody to remember anything. This is only available at some sites and it is by far the most effective option where it is.
Prompt at the right moment. Not the next morning. A notification twenty minutes after a scheduled shift end reaches somebody who is still nearby and can still fix it themselves.
Give supervisors a fast close. If the fix takes four clicks and a reason field, it gets done. If it takes a form and an approval, the record stays open and eventually somebody guesses in payroll instead.
Watch the concentration, not the total. The total tells you the size of the problem. The concentration tells you where it is. If a quarter of your missing clock-outs come from two sites, those two sites have an exit problem and you can solve it physically.
Where the gaps actually cluster
The total tells you the size of the problem. Where it sits tells you what to do about it, and in our data the concentration is not random.
Shifts that end at a handover. When one group leaves as another arrives, the exit is congested at exactly the moment everybody wants to use it. Our busiest single hour across 274 days is 18:00, carrying 11.1 per cent of all punches, and a queue at a terminal at the end of a shift is the most reliable way to make somebody decide it is not worth waiting.
Sites with one door. A single entry point that also serves as the exit handles double the traffic at both ends of a shift. The busiest terminal in our fleet has recorded 11,805 punches while the quietest has recorded two, and the top five carry 27.2 per cent of everything. Those five are where an exit problem becomes a payroll problem.
Shifts whose end is not fixed. Anybody who finishes when the work finishes rather than when the clock says so has no cue to punch out. There is no five-o'clock feeling to attach the habit to.
None of those three is solved by reminding people. Two of them are solved by changing the physical layout or adding a second terminal, and the third is solved by accepting the gap and making it cheap to close properly.
The rule we would give a supervisor
Close the record the same day, and write down why.
Same day, because the information decays fast. Today the supervisor remembers that the shift ran late because a delivery arrived at six. In a fortnight, when payroll asks, nobody remembers anything and the roster gets used as a substitute for what happened.
Write down why, because the reason is what makes the closure defensible. "Closed at 18:30 per roster, no exception reported" is a position you can hold a year later. A time with no explanation is indistinguishable from a guess, and a year later that is what it will be treated as.
The honest position
We have not solved this. Our own data is the 14.9 per cent quoted at the top, and it is the strongest argument we can make for the design principle underneath it.
The principle is that a system should be very good at recording what it observed and very explicit about what it assumed, and should never let the two look the same. That is an unglamorous property. It costs you a tidy dashboard. It is also the only thing that makes an attendance record worth anything on the day somebody disagrees with it.
That principle is built into Wurxa attendance. Punches arrive live from the ZKTeco terminals you already own, a day with an arrival and no departure is raised as a missed punch-out the next morning while people still remember, and a punch added by hand is stored as manual with its reason, so a captured punch never looks the same as an entered one, so your records hold up on the day somebody disagrees with them. It is free for up to 5 employees on one site, and every new workspace gets every feature for 14 days with no card. Start your free trial, or read the full dataset.
Put the workflow into practice.
Start with every Wurxa feature for 14 days. No credit card required.
