Where does the team's day go? Work sampling answers a different question from a stopwatch
Use transparent work sampling to examine activity structure, distinguishing observed shares from task duration and case waiting time.
Article language: English

In this article
Suppose a team still ends the day with a backlog after introducing an AI tool. The conversation quickly becomes a contest of memories. One person remembers correcting output, another meetings, a third looking for materials. Each may be describing a real part of the work. Those fragments do not yet show the whole.
Work sampling observes the type of activity being performed at selected moments. It helps identify the structure of work rather than starting a stopwatch for every task. In a company, agree a transparent purpose from the outset: we are looking for questions about the organisation of team work. The result will not be a ranking of people or a way to guess their engagement from a screen image.
Imagine an office preparing product information. Some material is produced with AI, employees check sources and pass texts for approval. The manager wants to know whether the problem is the share of corrections, completeness of data or cases waiting for a decision. These questions must be separated. One method does not automatically answer all three.
The share of observations is not the duration of one task
An NPTEL lecture on work sampling describes momentary observations at random times. The observed share of an activity is used to estimate its share of the time covered by the study. The material distinguishes the method from continuous measurement and notes that it does not directly provide the times of individual task elements. Selection and number of observations affect the basis for a conclusion.
A numerical example of our own can illustrate the difference. Suppose twenty out of one hundred recorded observations concern checking AI output. We can say they made up one fifth of those observations. We cannot infer the average duration of checking one text. Nor do we know whether a particular item waited an hour or several days for approval. This is an invented set illustrating the calculation, not a result from a real team.
When measuring work time with AI (in Polish), we ask about the effort needed for a specific task. In sampling, the question might be: how often at the observed moments do checking, data preparation or discussion of material occur? The wait of a particular case must instead be reconstructed from its history. The shared word “time” does not turn these three measures into one piece of information.
First agree what the team will identify
In our office “working on text” would be too broad. It would hide both preparing information and checking a tool's response. The team might propose separate working categories: preparing input, developing content, checking material, clarification with another person and another task. This is a proposal for the example and needs testing with actual employees.
Do not call all checking “correcting AI errors” if some of it concerns ordinary data approval. Do not put a discussion with an expert in “break” when it is needed to prepare the content. A category should describe an identifiable activity. Assessment of whether it can be reduced comes later and requires understanding why it is performed.
In research by Madushanka, Fiirgaard and Pérez described by Aarhus University, the authors separated classes of activities and work locations on a Danish construction site. The 2025 case covered three construction trades. Our office needs different activities, but the example shows why work type and location can be described separately. One label need not explain everything visible at a moment.
Before the main study, check several descriptions with two people. If one classifies the same situation as content development and another as checking, clarify the boundary. Add a short example that fits and one that does not. Do not expand the list until selecting a category becomes another major task for the team.
Do not observe only when you remember
Regularly looking in after the morning briefing may mainly show what happens after that briefing. To ask about a broader period, the plan must actually cover it. Random selection of moments is part of the described approach, not decoration in a spreadsheet. Convenient notes from a few moments can suggest questions, but should not be presented as a precise structure of an entire month.
NPTEL also points to informing participants in advance and preparing the observation method. In our example the team should know the purpose, categories, scope of information recorded and use of the aggregate result. Appropriate people handle organisational agreements and any assessment of legal requirements. A method described in a lecture is not automatic permission for arbitrary monitoring.
There is no need to record private message contents or document every click. Select the information scope to match the transparent question. If the company wants the share of material checking in a particular AI workflow, that workflow sets the boundary. An overly detailed record may create a new problem instead of explaining the existing one.
Record an unusual period separately. Our office may have been launching a catalogue or covering for an absent colleague. More spreadsheet rows do not remove that context. The result should identify the period and type of work covered. The team can later examine another period, but should not silently combine the two as identical conditions.
An open application does not name the person's activity
In the fictional office an employee has an AI response on screen but is discussing a variant label with a colleague. Another person reads a paper product sheet while preparing data for the next prompt. If activities were named only after the open application, the first record would be counted as AI work and the second could disappear from it. That would show a tool on screen, not the task the team intended to study.

When preparing the trial, establish what can actually be identified and how uncertainty is marked. An observer should not guess the content of a conversation or an employee's intention. If the selected category cannot be distinguished in the agreed way, improve its description or change the question. Filling every cell with an arbitrary answer does not make a spreadsheet more reliable.
An unclear record can remain unidentified and prompt a return to the definitions during method review. There is no need to add details of the document or the name of the person consulted. The team wants to know which type of work remains invisible, not reconstruct every private exchange. If the method needs more information, agree the new scope transparently before collecting it.
After clarification, a category might describe preparing sources for content development regardless of the application used. The company then retains the question about a specific activity. It may separately identify tools the team uses. The two sets can later be compared while keeping the activity and tool information distinct.
What a short record will not show
An employee waiting for a response on one item may be checking another at the same time. The observation will show checking. It will not automatically show that the first case has been stalled for a long time. To understand delays, the manager must connect the question about people's work to a separate question about the material's path. Otherwise she may try to reduce an activity that lets the team use waiting time productively.
A short observation will not explain the reason for a conversation either. Someone may be consulting about a product term, correcting missing data or answering a new colleague. Further investigation may require transparent observation with an expert (in Polish), with a separate purpose. The aggregate picture points to a question; looking at the whole task helps explain its course.
Analysis of team interruptions (in Polish) adds another perspective. A “conversation” record does not say whether the employee is responding to a needed consultation or returning after an accidental interruption. Without that distinction, limiting all conversations may hinder the task the company wanted to improve.
A rare event may not appear in the observations. Its absence from the record does not prove it never occurs. In our example one difficult correction may matter greatly to a catalogue deadline. Assessing it requires the history of that case, not only a category's share in a sample.
From a result to one thing to test
Suppose preparation of AI input appears frequently in a fictional summary. The first conclusion should not be “employees take too long to type prompts”. Find out what the category includes. Perhaps the team combines information from several documents, looks for an approved variant or clarifies a product name. A text-generation tool will not automatically remove that work.
The manager selects one item and asks about sources of the necessary data. If an inconsistent variant description emerges, the next check may concern input preparation rather than typing speed. The team records a hypothesis and chooses a limited change to test. It does not attribute the whole category share to one cause merely because that cause is easy to put in a presentation.
When investigating causes using evidence (in Polish), retain the difference between observation and explanation. “Data preparation appeared often in the sample” describes a result. “The cause is scattered sources” remains a hypothesis until the relevant cases have been reconstructed. A good question can come from the sample even though answering it needs another kind of material.
If the company repeats the study after a change, it should preserve category definitions or explicitly describe changes to them. Otherwise a lower share of checking may result from moving records into a new column. Comparison also needs the period and scope of work. Two similar-looking charts do not by themselves establish that observations are comparable.
The report should lead to a decision about work
Present the result with the number of observations, period, categories and gaps. If some records remained unclear, show them rather than distributing them by intuition. The team should recognise its work in the description. Ask which situations were not captured and whether any label combines activities needing different responses.
Do not turn it into a ranking of people. Measures that produce unwanted behaviour (in Polish) illustrate the problem of assessment based on visibility alone. If employees see checking as a category that lowers their rating, the result will cease to serve the question about information flow. The company needs a reliable picture of tasks, including work required to identify an error.
At Aeneia Audit, this material can be one input to assessing actual AI use: where input is prepared, who checks output and which tasks remain with people. Start with one tool and a defined workflow. A category's share cannot replace reviewing sources, checks and the needs of people using the result.
A work sample can end the contest of memories without immediately identifying one cause. A good result may be a more specific question: which data do people repeatedly reconstruct, and why? Only then do you know what to investigate. A stopwatch can still help with another task; it need not pretend to be a map of the whole day.






