Measuring time with AI: check savings after corrections
Compare similar tasks with and without AI. Measure the work through to acceptance, including source checks, revisions and time spent by everyone involved.
Article language: English

In this article
Imagine a draft made in two minutes. Revising it takes twenty, and checking sources another ten. A demonstration can stop the clock after the first paragraph, but a company needs a finished result. Measure time saved with AI from the start of an agreed task to acceptance at the required quality, including the work of everyone involved.
Separate effort from waiting time. A model can process a file for minutes while a person does something else. It can also answer immediately but require costly correction. One number cannot describe both. A well-designed trial shows where time disappears, where it changes owner and where new value appears.
First decide what “done” means
Choose a repeatable task, such as preparing a draft response to an enquiry. Record the input, required elements and person accepting the result. The result must meet the same criteria with and without AI. Comparing a finished answer with an unchecked draft measures different stages, not the benefit of the tool.
Criteria can include fidelity to the source, complete fields and a correct next step. Set them before the trial and connect them to AI pilot criteria, drawing on the guide to evaluating an AI pilot (in Polish). The evaluator should know which errors need correction and which prevent use of the result.
Do not change the definition of success after seeing an impressive demonstration. If ordinary work includes checking a document, the trial must include it too. The stopwatch should not leave the room before the person accepting the result.
Compare tasks of similar difficulty
Collect typical, difficult and incomplete examples. Group them by features that matter: length, number of sources, missing data or required agreements. If AI receives short questions and manual work gets complex cases, the result will mainly describe that split.
You can assign comparable tasks to two methods or alternate methods across similar cases. Be careful about having the same person repeat exactly the same task: the second time they already know the answer, making the tool's effect harder to assess.
Be cautious with a small sample. List the kinds of cases covered and those omitted. A few good results support a further trial in a defined scope, not a forecast for the entire company.
Measure work after generation too
Keep a simple record of stages: preparing data, working on the result, checking, corrections and handoff. Add retries and clarification of missing information. You do not need to log every cursor movement. You do need to attribute material effort to the task reliably.
Include other people's work. If a specialist checks answers more often than before, that time belongs in the comparison. Moving fifteen minutes from the author to the recipient is not a fifteen-minute saving for the company. It may be a justified change in the division of work, but needs a different description.
The guide to training teams to verify AI (in Polish) develops acceptance rules. In the measurement, note whether the checker knows the tool and criteria. Otherwise you may confuse the difficulty of a new workflow with a lasting feature of the solution.
Separate active work from elapsed time
For an exercise, suppose a manual task requires 40 minutes of work. With AI, preparation takes 8 minutes, the model works for 6, and checking and correction take 17. Human effort is 25 minutes. If the stages run consecutively without further breaks, elapsed time is 31 minutes. These are different results.
If another task was done while the model worked, do not add the same minutes to two active efforts. Record waiting separately. With several agents running in parallel, their combined running time may greatly exceed human working time; that does not automatically mean an equally large saving. For batch processing of records with AI, add the time needed to match each answer to its input and explain unfinished items before treating the batch as complete. Also include handoff of conditions and evidence between agents and the work of combining their results. For these steps, see the article on matching AI batch results to records (in Polish) and the guide to task handoffs between agents (in Polish).
In its February 2026 study-design update, METR describes difficulty measuring time with parallel agents and the effect of participant and task selection. The update concerns experienced developers working on real open-source tasks. Tasks were randomly assigned to allow or disallow AI, and participants recorded completion time. The authors regard the newer data as an unreliable basis for estimating the effect because of the stated limitations. We do not transfer that finding to all companies, occupations or models.
Completing a task faster and creating more value are separate questions
With AI a team may produce more drafts, analyses or comparisons. Check whether they were needed and used. Producing extra documents is not by itself evidence of savings. Sometimes the idea printer works faster than the company can read.
In a May 2026 METR survey, the author separates reported speedup from reported change in the value of work. The survey covered 349 technical workers recruited through a convenience sample from communities and contacts available to the researchers. It was not representative of all workers. These are participants' self-reports, not a direct measurement of savings in any company. The study notes that using AI can change which tasks people choose. Your measurement should likewise distinguish shortening an existing activity from doing a new one that was not previously planned.
If a new result has value, describe it separately. An additional comparison before a decision may be useful even without saving time. Not every benefit needs to fit in the box marked “minutes”.
Classify corrections by cause
Record what required rework: a false fact, omitted condition, wrong format or unclear instruction. The measurement can then help improve the process. Total correction time alone cannot show whether the tool, input or acceptance method needs changing.
For AI-generated sales offers, separate style editing from checking price and scope. For document translation, include checks of names, numbers and meaning. Do not call both activities generic “polishing”. Related examples appear in the audit of AI-generated sales proposals (in Polish) and the article on checking AI-assisted translations (in Polish).
Add resolved errors to AI response regression tests, using the guide to regression testing (in Polish) as a reference. After changing an instruction, check whether effort fell and whether other criteria deteriorated. A faster response does not compensate for losing important information.

Show implementation cost separately from daily work
Preparing instructions, test data and training takes time. Do not mix all of that effort into one case, but do not omit it either. Report startup cost and the work needed to maintain the process.
The total cost of an AI process also includes tools and managing change, as explained in the guide to calculating AI process costs (in Polish). The break-even point depends on the actual number of tasks and the stability of the benefit. Do not multiply the result from one easy case by every full-time position in the company.
After the trial, agree the scope of further use and when to measure again. A new model version or different documents can change the result. Keep the basis for the choice and limits of observation in a project decision register so that, months later, evidence can be distinguished from an optimistic forecast. The guide to project decision registers (in Polish) explains what to record.
A trial log should help measurement, not track every movement
Prepare a short card for each task: input type, difficulty, method, active work time, waiting and acceptance result. Add the number of material corrections and other people's involvement. That record helps explain differences later. Minutes without a description of the tasks soon become a number detached from the conditions that produced it.
Agree the rules with participants before starting. The trial should assess the way the tool is used, not find who looks best in a table. If people fear punishment for corrections, they may not record them. The report may become shorter, but the benefit assessment less reliable. That is a poor trade of data quality for a calm chart.
Do not collect the entire content of messages if a task identifier and error category suffice. Record only material needed to assess and explain the result. Set access, an owner and an end date for retaining the log. A savings trial should not incidentally create another archive of the team's work without a clear purpose.
For selected cases, compare the time record with observation of a controlled task. This checks whether everyone understands start, finish and interruption in the same way. It does not require continuous screen recording. Often walking through an example together and correcting the instructions is enough when one person counts waiting as work and another does not.
Also show the number of unfinished tasks in the summary. Removing them can make the method needing the most rescue look exceptionally good. Record why work stopped and what effort another completion route needs. If that work has not yet been measured, leave a visible gap instead of substituting a convenient zero.
Separate learning a new workflow from stable operation
The first tasks include learning the interface, instructions and checking method. Record that stage. A short onboarding period can precede the actual comparison, but retain its effort: the company really bears it even if it is not charged to every later case.
After learning, repeat the trial on new comparable cases. Do not use only practice cases, because participants already know their pitfalls. Record each person's experience with the tool and task. An expert instruction author's result does not automatically describe a team just receiving access.
Check behaviour with harder input. Can the employee recognise early that further correction will cost more than another method? Agree an allowed route to stop and switch to manual work. Time spent rescuing a poor draft still belongs to the AI method's effort.
Do not confuse improving skill with changing the model. If you change the tool or instruction during the trial, mark the configuration for each task. Otherwise it will be unclear which version produced the benefit. With few cases, the change may call for a new comparison instead of combining all results into one average.
Show the spread and nature of exceptions. The median is the middle value of an ordered set of measurements; with an even count it is the average of the two middle values. It can describe a typical task, but also show the time for difficult cases and the cause of long corrections. Two similar averages can describe very different rhythms: predictable effort or a few very costly surprises.
Check what the company can do with the recovered time
Saved minutes do not always create a free block of work. They may appear in short intervals between tasks and not reduce the number of required shifts. Record whether the benefit is a shorter queue, more accepted cases, lower load or capacity for another needed activity. Do not assign all these effects the same monetary value.
Prepare a simple scenario for using the benefit. If a typical task requires less effort, check what limits the number of finished cases: data availability, an approver's time or waiting for the customer. Removing one part of the work may help a team without shortening the entire elapsed time. Both facts matter to the decision.
Do not forecast savings for tasks that were not studied. Expansion to another department may involve different data, quality criteria and reviewer skills. Keep the result as a reference, then run a limited trial in the new scope. Scaling a calculation is easier than scaling the conditions that produced it.
Record the final decision with its boundaries: which cases showed a benefit, what corrections remain necessary and what is still unknown. That lets the team develop the solution without pretending certainty. An honest “we gained here; this still needs checking” gives the team more than an impressive number without references to its own work.
Before accepting a report, pick one case and reconstruct the whole calculation. Someone outside the trial should be able to identify the author's work, the recipient's work, waiting and the confirmed result. If they cannot see where the saving came from, complete the record. This is a simple check of whether the report describes the task performed or just the first impression of an answer.
Keep the observed result and forecast separate. The first concerns cases performed under known conditions; the second concerns future volume and assumptions. For the forecast, name dependencies that may change. That way, a decision to expand rests on visible premises, and promised minutes do not acquire a life of their own in a budget presentation.
For a discussion of measuring a particular workflow, see Aeneia Audit. Prepare one task, the criterion for a finished result and a description of current checks. Any comparison, including verification and corrections, requires an agreed scope. Learn more about Aeneia Audit.









