AI Saved the Time. Your Calendar Will Take It Back

A faster task does not create a team dividend by itself. Verify the saving, record its full cost, then give the recovered capacity an authorised destination and a protected calendar slot before routine demand quietly consumes it.

AI can shorten a task without changing the work around it. The draft arrives earlier, then waits for review. The meeting summary takes minutes, then triggers three follow-ups. The individual feels faster, while the team finishes the week with the same backlog.

Do not announce "hours saved" and hope valuable work appears. Use a double-entry time-dividend record. Every verified saving gets an evidence line and an authorised destination, with calendar protection and a review date. Until both entries exist, you have a task-level observation, not recovered team capacity.

This begins after a defined use case has produced evidence. It allocates verified capacity rather than forecasting staffing, setting an operating rhythm or measuring licence adoption. A person retains every work-allocation, workload, staffing and employment decision.

Why does saved time disappear?

Individual tools change actions that one person can alter more easily than work requiring coordination. An NBER working paper by Dillon, Jaffe, Immorlica and Stanton reports a six-month randomised field experiment across 66 firms and 7,137 knowledge workers. Workers were randomly selected to receive access to a generative AI tool embedded in applications used for email, meetings and writing. In the second half of the experiment, the 80 per cent of treated workers who used the tool spent two fewer hours on email each week and reduced work outside regular hours. The researchers did not detect a change in the quantity or composition of tasks from individual-level access (NBER).

The study used telemetry data from workplace applications within a randomised access design, giving stronger evidence about changed work patterns than a recollection survey. It does not show that two hours became transferable capacity, that every user saved the same amount or that an Australian financial-services team could remove two hours from a staffing model. Some authors were employed by Microsoft, the tool maker, during the study. Both NBER papers are working papers circulated for discussion, not peer-reviewed articles.

A second NBER working paper reaches the issue from a different direction. Humlum and Vestergaard linked large-scale adoption surveys to Danish administrative labour-market records. Using a difference-in-differences design, they found no detectable effect on earnings or recorded hours at worker or workplace level and ruled out effects larger than 2 per cent over the two years following ChatGPT's release. They did find changes in work structure, including new tasks involving content generation, AI oversight and integration (NBER).

The Danish results do not prove that AI produces no productivity benefit, and they do not estimate your team's recoverable minutes. They show that self-reported benefits and changing tasks can coexist with no measurable movement in recorded hours or earnings during the study period. Faster components can be absorbed by coordination, new oversight work, additional output or existing demand.

This is why multiplication fails. Ten people each reporting 30 minutes saved does not establish five available team hours. The reports may cover different tasks, periods and baselines. Review or rework may sit elsewhere. The people may already have used the minutes on other work. Never convert individual estimates into headcount, staffing reductions or service commitments without separate workflow and workforce evidence. Capacity planning for AI-assisted work is its own exercise, not an arithmetic shortcut.

What counts as a verified time dividend?

Start with one repeated task and a stable definition of done. A complaint summary is not done when AI produces text. It may be done only after a human checks the source record, corrects material errors, applies the approved format and accepts it for the next step. Measure the whole path across comparable periods.

The evidence line in the record needs six fields:

1. Task boundary: the exact start point, completion point, population and period.
2. Baseline: observed human time before the approved AI-assisted method, including normal variation.
3. Assisted result: observed end-to-end time after adoption, not the generation time shown by the tool.
4. Full human cost: preparation, prompting, verification, corrections, exceptions, escalation and documentation.
5. Quality and control result: the existing acceptance measure, material defects, reopened work and any control failure.
6. Confidence: observed, sampled or self-reported, plus the limitations and evidence owner.

The Digital Transformation Agency's evaluation of the non-randomised Australian Government Microsoft 365 Copilot trial shows why evidence labels matter (Digital Transformation Agency). Of 807 respondents to a post-use question, 41 per cent believed Copilot enabled more time on higher-value or complex tasks (Digital Transformation Agency). The same discipline that grades every claim in an executive AI update applies to a time-saving claim.

That is Australian evidence about perceived reallocation, not a measured causal estimate of capacity. The evaluation's executive summary says the methodology relied on participants self-assessing impacts, which may under or overestimate the benefits, particularly time savings. Its responses may not represent the trial population. Record the result as self-reported, not telemetry, and do not use it to set an expected saving for another workplace.

Use this prompt to structure approved, de-identified evidence. A human process owner and control owner must validate the task boundary, evidence classification and result before any dividend is recognised.

Draft the evidence line for a time-dividend record for [TASK_NAME].

Return:
- task start, completion point, population and measurement period
- observed baseline time and sample
- observed AI-assisted end-to-end time and sample
- preparation, human review, correction, exception and documentation time
- quality, rework and control results
- evidence type: observed, sampled or self-reported
- missing evidence and limitations
- proposed review date

Use only the supplied data. Do not estimate missing minutes, aggregate individual claims, calculate headcount or recommend a staffing outcome. Mark unsupported statements [CHECK].

Approved, de-identified evidence:
[PASTE EVIDENCE]

Reject a dividend when periods do not match, the definition of done changed, review time is missing, volume or case complexity moved materially, or quality fell outside the existing limit. The honest result may be "not yet measurable". That is better than spending capacity that never existed.

How do you stop the calendar taking it back?

The second entry gives an accepted saving a destination. Choose one of four lanes: coaching, complex work, control improvement or skill-building. Name the activity, beneficiary, authorised owner, calendar protection, success evidence and review date. "Higher-value work" is not a destination. It is a category waiting to be crowded out.

Safe Work Australia says good work design should consider the demands, duration, frequency and complexity of tasks, workers' needs, preferences and capacities, and the systems and processes involved in the work, including information technology and business management. The same page also says to involve the people doing the work, engage decision makers and revisit design when workplace changes occur (Safe Work Australia). Safe Work Australia is not a regulator, and the page refers to model WHS laws. Organisations must check the requirements applying in their jurisdiction.

Treat allocation as work design, not as a private calendar preference. Consult the people who perform and review the task. Ask whether routine demand has intensified, whether complex cases are accumulating and whether the proposed destination is usable. AI-driven pace changes can intensify work before anyone decides they should. Do not force employees to manufacture a saving or turn the ledger into individual performance surveillance.

Here is a fictional worked example.

Evidence line: [CLAIMS_SUPPORT_TEAM] samples [SAMPLE_SIZE] low-complexity, de-identified file summaries across [BASELINE_PERIOD] and [ASSISTED_PERIOD]. End-to-end median time falls from [BASELINE_MINUTES] to [ASSISTED_MINUTES] after preparation, verification and correction are included. Acceptance and rework remain within [APPROVED_LIMIT]. [PROCESS_OWNER_ROLE] classifies the result as sampled, not universal.

Destination line: [AUTHORISED_LEADER_ROLE] allocates [VERIFIED_TEAM_MINUTES] each fortnight to a supervised complex-case clinic led by [SENIOR_REVIEWER_ROLE]. The slot is booked for [REVIEW_CYCLE], protected from routine allocation unless [SERVICE_TRIGGER] is met, and reviewed on [REVIEW_DATE]. Success is assessed through cases discussed, decisions clarified and participant feedback. It is not converted to a staffing outcome.

Run this checklist before approving the pair:

• The saving uses end-to-end time and the same definition of done.
• Review, correction, exceptions and documentation are included.
• Quality and control evidence sit beside time evidence.
• The destination is specific, authorised and visible in the team's workload plan.
• Affected workers and reviewers have been consulted through applicable processes.
• A calendar defence names the limited conditions that may displace the activity.
• The review date tests both the saving and whether the destination actually occurred.
• No individual minutes have been translated into headcount or employment consequences.

Use this prompt to draft destination options. The accountable leader, affected workers and relevant risk, people and operational representatives must review the options and make the allocation decision.

Draft three destination options for the verified time dividend below.

Use only these lanes: coaching, complex work, control improvement or skill-building.
For each option provide:
- specific activity and intended beneficiary
- authorised owner and participants
- calendar slot and protection rule
- service trigger that may displace it
- evidence of delivery and value
- review date and stop condition

Preserve existing quality, review and workload controls. Do not rank employees, assign work, alter staffing or infer an employment consequence. Label required human approvals [APPROVAL REQUIRED].

Human-verified evidence line:
[PASTE VERIFIED RECORD]

Calendar defence is deliberately bounded. A critical service event may require the slot to move. Routine busyness should not erase it every week. If displacement repeats, record that outcome. It may show that the claimed dividend was not available at team level or that leaders chose another destination without saying so.

Do this Monday

1. Choose one completed use case. Select a repeated task with an approved AI-assisted method, a stable definition of done and enough recent cases to compare.
2. Build the evidence line. Capture baseline and assisted end-to-end time, full human cost, quality, rework and evidence type. Ask the process and control owners to validate it.
3. Name one destination. Consult the team, then select coaching, complex work, control improvement or skill-building. Write the activity, owner, beneficiary and approval.
4. Defend the calendar. Book the time for a defined review cycle. State the exceptional service trigger that can displace it and record every displacement.
5. Review both entries. On [REVIEW_DATE], test whether the saving remained valid and the destination occurred. Continue, adjust or close the record through a human decision.

Bottom line

A faster task is not automatically recovered capacity. Verify the complete saving, then pair it with an authorised destination and a defended calendar slot. Review both entries before making any broader claim. AI can organise the evidence and draft options, but people decide work allocation, workload, staffing and every employment consequence.

This article is general information and education only. It is not legal, compliance, financial or professional advice. Obligations vary by organisation and circumstance. Verify current requirements against the primary sources cited and seek advice specific to your situation.

References

1. Eleanor W. Dillon, Sonia Jaffe, Nicole Immorlica and Christopher T. Stanton, "Shifting Work Patterns with Generative AI", NBER Working Paper 33795, issued May 2025 and revised November 2025. https://www.nber.org/papers/w33795
2. Anders Humlum and Emilie Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI", NBER Working Paper 33777, issued May 2025 and revised March 2026. https://www.nber.org/papers/w33777
3. Digital Transformation Agency, "Evaluation of the whole-of-government trial of Microsoft 365 Copilot: full report, executive summary and glossary", released 23 October 2024. https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/executive-summary-glossary
4. Digital Transformation Agency, "Evaluation of the whole-of-government trial of Microsoft 365 Copilot: full report, productivity", released 23 October 2024. https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/productivity
5. Safe Work Australia, "Good work design", accessed 31 July 2026. https://www.safeworkaustralia.gov.au/safety-topic/managing-health-and-safety/good-work-design