Did AI save time after checking and rework?

Start the clock before the draft and stop after acceptance

AI produces the first version in minutes. Then someone checks it, fixes an omission and asks a colleague to review the recommendation.

The draft was faster. Was the work?

For a team deciding whether to keep using an AI-assisted approach, that second question matters. Count the effort needed to reach an output someone can use, including work that moved to another person.

At Altiv, we recommend measuring one recurring workflow before making a broad productivity claim. Define the result, agree what acceptable means, and compare the complete effort required to get there.

Research shows why the result matters more than generation speed

The published study Generative AI at Work examined a phased introduction of an AI assistant across 5,172 customer-support agents. Productivity was measured as issues resolved per hour, and increased by 15% on average, with substantial differences between workers. The outcome was resolved work, not the amount of text the assistant generated. That is evidence from one setting, not a target every team should expect to match. Read the research team's study summary. Read the research team's study summary. ↗

Measurement can also become difficult as work changes. In February 2026, METR explained that selection effects and difficulty recording time for developers using concurrent agents made its follow-up productivity estimates unreliable. Its methodological update is a reason to examine how a result was measured, not to recycle an older headline about AI making everyone faster or slower. Read METR's update. Read METR's update. ↗

For your own team, the practical question is narrower: did this approach help us complete this kind of work to the standard we need?

Agree what counts as finished

Consider a product team comparing two ways to prepare a weekly customer-feedback brief. This is a suggested measurement exercise, not a reported team result.

The brief should identify important themes, preserve contradictory evidence, link findings to the underlying feedback and distinguish a problem worth investigating from a feature commitment.

Both approaches need to meet the same standard. A quick summary that drops a consequential disagreement is not equivalent to a checked brief, even if it looks more polished.

Before the comparison, record:

• The output and who will use it.
• The source material and its size or difficulty.
• What the reviewer needs to accept it.
• The part of the workflow AI will assist.
• The people whose effort will be counted.

Use these conditions to explain differences between runs. A smaller batch of feedback should not quietly become evidence that the AI-assisted approach is faster.

Count the complete effort

Use the same categories for both approaches. The current workflow also includes checking and correction; do not count those costs only against AI.

Record active human minutes. Track elapsed delivery time separately, from the start of the agreed task to acceptance. Waiting for a model or a reviewer can lengthen delivery without occupying someone's attention for the whole interval. Conversely, two people working simultaneously can use more person-minutes than the elapsed clock time.

Keep one-time setup separate too. A reusable template or integration may take effort to prepare, but charging all of that work to every run would exaggerate its ongoing cost. Excluding it entirely would hide the investment.

Calculate a difference only when the outputs are comparable

For accepted outputs of comparable scope:

A positive difference indicates less recorded effort in the AI-assisted run. A negative difference indicates more. If you express it as a percentage, divide by the current-workflow effort, provided that baseline is greater than zero.

Do not fill an unrecorded stage with zero. Zero means you observed no effort for that stage; blank means you do not yet have a complete measurement.

If an output fails the acceptance check, retain the record. Complete the correction and include its effort before comparing accepted outputs. If the approach is abandoned, record the work spent and the failure rather than dropping the run from the evaluation.

For the customer-feedback brief, a result might be “ready for the planning meeting” or “needs another pass because findings cannot be traced to feedback.” The status explains whether the timing supports a useful comparison.

Enlarge image ↗Open original ↗

Close image ×

Repeat with work that reflects the actual job

A single run helps identify where effort goes. It is weak evidence of a dependable improvement.

Compare several representative tasks, including ordinary and more difficult cases. Record who did the work and which tool or model was used. Keep the acceptance criteria stable, and note changes in the process.

If the same person performs the same task twice, the second attempt benefits from familiarity. Use comparable fresh inputs where possible. If feasible, alternate which approach is used first or assign comparable tasks between approaches. Keep remaining differences visible rather than presenting a small workplace comparison as a controlled experiment.

Review the individual records as well as the average. A method that works well on straightforward briefs may require different checking for disputed evidence or missing context.

Time is one kind of value

AI may help a team examine more alternatives, produce a more useful output or attempt work it could not previously do. Those benefits deserve measurement even when the clock does not improve.

Atlassian's AI ROI guidance distinguishes efficiency, quality and innovation measures alongside adoption. That is a useful reminder to choose a measure that fits the intended change. See the AI value measurement guide. See the AI value measurement guide. ↗

Time released is capacity. It becomes a financial saving only when spending changes, or another financial benefit is established. Multiplying minutes by a salary rate estimates the value of capacity; it does not prove that money left the budget.

A financial ROI calculation also needs relevant costs, including tools, setup, training and ongoing operation. Keep those separate from a task-level claim that one run took less effort.

Make a decision about the workflow

Use the results to choose a specific next action:

• Keep the approach where comparable work meets the standard with a useful benefit.
• Change the step that creates extra review or correction effort, then measure again.
• Stop or hold the approach when it cannot meet the required standard or the costs outweigh the benefit being sought.

You may find that AI is useful for organising the feedback but less useful for drafting the recommendation. That is a practical result: the measurement tells you where to use it.

Task and intended output:
Approach and tool/model:
Inputs and difficulty:
Acceptance criteria and result:
Preparation minutes:
Production minutes:
Review minutes:
Rework minutes:
Additional colleague minutes, not counted above:
Total active human effort:
Elapsed time to acceptance:
One-time setup, recorded separately:
Evidence and next action:

Start with the next recurring task and an agreed definition of finished. The useful outcome is a decision you can explain about where AI improves the work.