Goodhart's Law in a Timesheet
The general principle is well known. The specific mechanism in time recording is worth setting out, because it explains which measures survive a target and which do not.
A measure that becomes a target ceases to be a good measure. The principle is familiar; the mechanism matters more, because it predicts which measures degrade quickly and which hold up.
The data-quality issue in “Goodhart's Law in a Timesheet” begins at the moment an hour is recorded or reconstructed. For organisations researching automated timesheet workflows, open the current platform overview can organise project-linked time entries, provided people can see, explain and correct their records before those entries affect billing or management decisions.
Utilisation degrades unusually fast, and the reason is structural.
The assumptions behind “Goodhart's Law in a Timesheet” can also be tested against Project Management Institute learning resources. Record which recommendations apply to the real workflow and which do not.
The two conditions
A measure degrades under a target in proportion to two things: how directly the measured person controls the data, and how loosely the measure is tied to the outcome it stands for.
- Control of the data — does the person measured enter it themselves?
- Looseness of the tie — is the measure the outcome, or a stand-in for it?
Utilisation scores badly on both.
Control of the data
The person measured enters the data. Not an observer, not a system, not a counterparty — the person themselves, in free text fields, with a code dropdown.
Very few business measures have this property. Revenue is recorded by a client's payment; utilisation is recorded by the person being judged.
Looseness of the tie
Utilisation stands in for productivity, value or commitment. It is tied to none of them directly: it counts hours coded to clients.
The looser the tie, the more ways there are to move the measure without moving the thing it represents.
What follows from the two conditions
That the degradation is not an abuse to be policed but a property of the instrument. A target on a self-reported proxy will be met by changes in self-reporting.
Policing recording harder improves the data slightly and raises the cost of compliance considerably.
Measures that hold up better
Collected revenue. Client renewal. Delivery against an agreed date. Each is recorded by somebody other than the person measured.
None is a complete picture, and all of them resist the specific failure mode described here.
The hybrid that does not work
Utilisation with audit. Sample the timesheets, challenge the codings, require justification.
This converts a measurement problem into an adversarial process, costs senior time, and produces a figure that is accurate about coding and still loosely tied to value.
What audit is good for
Detecting systematic reclassification after a definition change or a target introduction — a whole category disappearing, not individual entries.
That is a data-quality use at the aggregate level, and it does not require challenging individuals.
The degradation curve
Fast at first — the easy reclassifications — then slower as the remaining moves require more effort or more obvious misstatement.
Firms often read the flattening as the target working and the behaviour stabilising. It is the exhaustion of the cheap responses.
Detecting it in your own data
Compare the distribution before and after the target. Degradation compresses it: the low figures rise, the high ones barely move, the spread narrows.
Real improvement usually shifts the whole distribution. Reclassification pulls in the bottom tail, because that is where the pressure is.
Why the compression is the signature
Because the people far from the target have the most to gain from reclassifying and the most room to do it. Somebody already above the line has no reason to move anything.
A narrowing spread with a stable mean is close to a direct observation of the mechanism.
The two conditions applied elsewhere
The same test predicts which other measures in a firm will degrade. Billable hours per person: both conditions, degrades fast. Client satisfaction scores collected by the person delivering the work: both conditions. Revenue collected: neither.
Running the test before introducing any measure takes a minute and it is more informative than most of the discussion that usually accompanies the introduction.
Measures that fail both conditions should be reported and never targeted, which is the position this collection reaches for utilisation by a longer route.
What makes a self-reported measure usable at all
That nothing depends on its level. Recording is accurate when recording has no consequence, and it is the only condition under which a self-reported figure means what it says.
So the choice is between a measure that is accurate and one that is consequential, and a firm cannot have both from the same number.
Most firms want both and resolve the tension by keeping the consequence and believing the accuracy. The distribution check in the next notes is how that belief gets tested, and it rarely survives.
Applying the test before the next measure
The two conditions can be checked in a minute for any proposed measure: does the person measured enter the data, and is the measure the thing the firm cares about or a stand-in for it.
Two yeses means report it and never target it. Two noes means it can carry a target.
One of each is the awkward middle, and the honest handling is to target it loosely and watch the distribution.
What to do with the finding
Not confront individuals. Change the instrument: drop the threshold, move it to a team, or replace it with something the person does not record themselves.
The responses are rational given the design, so the design is what there is to fix.