How to check whether generative AI saves you time on an office task

To find out whether a generative artificial intelligence tool can help, you do not need to start with a broad claim about “productivity.” Narrow the question: does it reduce the time needed to prepare a meeting summary, classify requests, or draft an initial version without lowering the acceptable quality level? The answer will apply only to that task, that type of material, and the conditions you tested.

Choose an activity that can be repeated under reasonably comparable circumstances and whose output can be reviewed. Before you begin, define what counts as finished work: for example, a summary that captures the decisions, owners, and deadlines included in a set of notes, without adding information that is not in them. Set the quality criteria before comparing the results so that you do not adjust them later in favor of the version you prefer.

Also note the conditions that could affect the comparison: the length and difficulty of the material, how familiar you are with the task, the tool and version used, and the required format. If the cases differ substantially, the time difference may be due to that variation rather than the use of AI. For an initial check, use test documents or information you are authorized to process.

Run one test without AI and another with the chosen tool, keeping the objective the same and using equivalent material. If you have only one document, do not use the exact first response as the starting point for the second test: already knowing the content could make the repeat easier. It is better to test two similar cases and, if you repeat the exercise, alternate which one you tackle first.

In the AI-assisted test, define what you ask the system to do and what you do yourself. Save the prompt and record whether you needed to rephrase it, provide context, wait for a response, or correct the output. In the test without AI, also count preparation and review time. Compare the entire process, not just the speed of generation: a quick response that requires substantial follow-up work may not reduce the total effort.

A simple protocol helps keep the comparison orderly:

1. Choose two cases of similar difficulty and define the expected output.
2. Measure from the start of the task until the output has been reviewed and is ready.
3. Record separately the time spent on instructions, editing, checking, and corrections.
4. Apply the same quality checklist to both outputs.

A single run is not a conclusive test; it is only an initial indication of whether the exercise is worth repeating.

Record the total time for each case and, if possible, break it down by work phase. Include preparing the material, writing or adjusting instructions, reviewing the response, and correcting errors. Also note how many significant corrections were needed and what problems arose: omissions, incorrect data, unsuitable formatting, or content not supported by the source material.

Then compare the outputs using the same checklist. For a summary, you could check whether it includes all relevant points, whether names and figures match the source, whether facts and uncertainties are distinguished, and whether the requested length has been respected. There is no need to create a sophisticated score: it is enough for the criteria to be observable and applied equally to both versions.

A basic template might include: task and date; tool and version; estimated difficulty; total time and time by phase; number of revisions; significant errors; criteria met; and final decision. If another person reviews the results without knowing which one was produced with AI, some bias may be reduced, although it will not disappear entirely. The aim is not to produce a universal figure, but to find out whether possible time savings outweigh the revisions and the risk of error.

If the AI-assisted version is completed sooner and meets the same quality level, the conclusion should be limited to what was observed: in that specific test, the tool appeared to save time without lowering the evaluated criteria. This does not show that the same will hold for other tasks, people, more complex material, or an entire organization. Nor does it let you attribute any improvement exclusively to AI: practice, the order of the tests, or differences between cases may also have played a part.

If the time goes down but omissions or checking increase, the stopwatch should not be the only measure. If quality improves but the process takes longer, the tool could still be useful, although the benefit would not be time savings. If the result depends heavily on how the prompt is written, record that dependence: learning to use the tool is also part of the cost, at least initially.

Repeat the comparison with several similar cases before changing a workflow. Keep the protocol and note any change in tool or method so that you do not mix results. Analyses of AI and employment address broader phenomena than an individual test and are not a substitute for measuring a specific task. Personal experience can inform a local, revisable decision, but on its own it does not support a general claim about productivity.

Before entering work-related content, check your organization’s rules and the service’s official privacy documentation. Find out what data is processed, for what purposes, how long it might be retained, and what controls are available for the relevant account type. Do not enter personal data, trade secrets, or confidential information without authorization and an appropriate legal basis; removing names is not always enough if the context could identify someone.

If it is unclear whether you may use a real document, consult the person responsible for security, privacy, or compliance. To test the procedure, use a fictional case or public material, provided it does not reproduce protected information. Do not assume that the terms for a personal account are the same as those for a business account: review the terms applicable to the specific tool and settings.

The International Labour Organization published an updated 2025 edition on generative AI and employment, focused on a broader labor framework than the assessment of an individual task. The Spanish Data Protection Agency, meanwhile, discusses prompts as a possible point of exposure for personal data and confidentiality. These references do not replace internal rules or the documentation for the specific service: a performance test does not, by itself, authorize the processing of the data used.

At the end, state the conclusion using four elements: task tested, conditions, observed result, and main limitation. For example: “Across three summaries of fictional notes, AI assistance reduced the recorded time, but every figure had to be checked; no real documents or other tasks were tested.” This is more precise than saying “AI increases productivity,” because it makes clear what was and was not measured.

Repeat the test if the cases were not comparable, there was only one run, or small variations changed the result. It may also make sense to start with low-risk tasks and clear review criteria before considering uses where an error could have serious consequences. If the benefit disappears once editing and checking are counted, the tool may not be worthwhile in that particular workflow, even if it could be useful for another one.

The final decision could be to keep the process manual, introduce the tool with human review, or run a more controlled test. It should not depend on a single figure: the required quality, the cost of correcting errors, data sensitivity, and the repeatability of the time savings all matter. The strongest conclusion is the one that does not go beyond what was actually observed.