Before turning an AI pilot’s faster task times into a higher delivery target, decide how the team will use any recovered capacity. Extra customer requests, broader responsibilities, tool maintenance and review can absorb it. The pilot should make those choices visible, including whether work is spilling beyond agreed hours.
That does not mean additional output is a bad outcome. A business may deliberately use AI to clear a backlog or improve service. The question is whether the organization has chosen that outcome and funded the work around it, or quietly treated every improvement as permission to add more.
This guide proposes a workload review alongside the usual measurement of accepted work. Its contribution is a capacity allocation record: distinguish the work made faster, the work added, and the decision about what happens next. It is a planning aid, not a validated health assessment or a forecast of staffing needs.
What the research can tell you about a busier team
In a February 2026 Berkeley Haas interview, Xingqi Maggie Ye describes an eight-month ethnography at a 200-person US technology company, including more than 40 interviews. Ye and Aruna Ranganathan observed expanding task scope, work entering former pauses and non-work time, and more simultaneous activity. This was in-progress qualitative research at one company. It identifies mechanisms worth investigating; it does not establish how often they occur across businesses or prove that AI universally increases workload.
A broader study offers a different measurement perspective. Humlum and Vestergaard’s March 2026 working paper links Danish adoption surveys in 11 exposed occupations to administrative records through December 2024. Workers reported new content, oversight and integration tasks, while the authors’ comparisons of adopters and non-adopters found no detectable average changes in recorded hours or earnings. The analysis relies on assumptions about how those groups would otherwise have evolved. It concerns early chatbot adoption, not today’s agent capabilities. Our inference is that a stable hours total can coexist with a changed job; the paper is not a direct measure of strain within those hours.
Anthropic’s 16 October 2026 internal employment report adds a current vendor perspective. Drawing on internal data and employee interviews, it describes attention bottlenecks and growing work to organize and supervise agents. This is the AI supplier studying its own unusually AI-intensive workplace. Adoption was not randomized, and company growth complicates attribution. Its account informs questions about coordination; it does not independently validate a workload benefit for customers.
These sources measure different populations, periods and outcomes. Do not combine them into a single productivity or burnout estimate. Use them to ask what your pilot’s task timing leaves out.
Keep unchanged work and expanded work separate
Start with a recurring service or process whose accepted output is already defined. Preserve its quality standard and identify everyone contributing human effort: the person prompting, the reviewer, the colleague handling exceptions and the person maintaining the automation.
For the unchanged work, compare the full human effort needed to reach acceptance, including corrections. Keep agent runtime and calendar waiting separate. Then record additions: extra cases, new deliverables, new coordination and new maintenance. Check for double counting; reviewing an original case cannot also appear as a separate overhead if its time is already included in that case.
Consider this entirely hypothetical weekly record for a customer-operations team. All figures describe one workflow across the team, not an employee’s whole working week. Assume the original cases remain comparable and meet the same acceptance standard.
| Workflow component | Before pilot | During pilot |
|---|---|---|
| Original 100 accepted cases, including review | 50 hours | 40 hours |
| Ten additional accepted cases | 0 hours | 4 hours |
| New helper maintenance, outside case handling | 0 hours | 2 hours |
| Pilot setup and training, outside case handling | 0 hours | 1 hour |
| Total human effort for this workflow | 50 hours | 47 hours |
The original workload takes ten fewer hours, a hypothetical 20% reduction. Seven hours are then used by additions. Calling the entire result a 20% workload reduction would hide those additions. Calling the ten extra accepted cases worthless would also misrepresent the result.
The remaining three hours are a provisional planning margin in this invented example. They are not evidence of three hours of free time in a real calendar. Other responsibilities, peaks in demand, fragmented availability and uncertainty still matter. Nor can a single week establish that maintenance or training will stay at these levels.
Agree how recovered capacity may be used
For the example, the manager and team might propose keeping the ten additional cases, protecting two hours for learning and leaving one hour uncommitted for variation. That allocation is an editorial illustration, not a recommended universal split. Another team might choose to reduce an existing backlog, strengthen checks or restore pauses displaced by prior overload.
Write the agreement in terms of the work it changes. Name the person who can approve added demand, the existing commitment being reduced or retained, and the date for reconsidering the allocation. Include reviewers and maintainers in that conversation; saved drafting time does not automatically create available reviewer time.
Also decide what should happen if the margin disappears. A workable response might be to defer optional new requests, narrow the pilot or temporarily return a task to the established process. Avoid making extra delivery permanent before checking whether it depends on exceptional effort.
This is particularly important when employees themselves want to try more ambitious work. Enthusiasm can be valuable without becoming an obligation for everyone else to match the busiest pilot week.
Review the work that persists after the demonstration
In a May 2026 practitioner essay, Voys founder Mark Vletter highlights the maintenance obligations of the tools and automations people build. We use that as an attributed practitioner concern, not as proof of the essay’s wider scientific or health claims.
Give each new helper or workflow an owner and an estimate to revisit. Who handles changed source formats, broken connections, instruction updates and exceptions? If nobody has accepted that responsibility, the pilot has created an unresolved commitment.
Record whether this is temporary setup or recurring work. Avoid amortizing setup over an imagined year of successful use when the pilot has not yet established that it will continue. Conversely, one difficult onboarding week should not automatically be treated as the permanent operating cost.
An enterprise can keep this record at the service level and identify transfers between departments. A small business can use a short shared note. In either setting, include the owner who checks outputs after normal delivery work finishes; otherwise the apparent capacity gain may sit on one person’s evening calendar.
Check working boundaries without ranking employees
Agree what participants will report before collecting workload feedback. A brief team discussion can ask whether optional requests have become expected, whether checks interrupt other responsibilities, and whether people feel able to leave non-urgent agent output until their next working period.
Use aggregate workflow observations and voluntary feedback for this review. Do not infer an individual’s effort, health or performance from token counts, active agent sessions or the time of a message. Those signals do not establish how long someone worked or what they experienced. Keep any individual concern in the organization’s appropriate human support process.
Separate unchanged total hours from the pattern inside them. More accepted cases may coexist with fewer interruptions, or with constant switching and less control. Ask about both outcomes. Where review requests are non-urgent, the team can trial agreed review windows and a route for genuine exceptions, then check whether service and working boundaries improve.
Make the next target a recorded decision
At the pilot review, inspect three things together: accepted output and quality, total workflow effort including additions, and participants’ account of how the work fits into agreed responsibilities and hours. Keep uncertainty visible. If the pilot relied on volunteers or unusually strong support, retain the limits on who its results represent.
Expand only the commitment the evidence supports. You might retain a faster process while declining new scope, accept extra demand with dedicated review capacity, or extend observation because maintenance is still changing. A proposed increase in throughput should identify who authorizes it and what gives way if assumptions fail.
For the next pilot, add one sentence to the brief before work begins: “Any capacity recovered will be allocated at the review, with the team, before it becomes a new delivery expectation.” Then name that review’s owner and date. The sentence cannot guarantee a sustainable workload, but it makes the allocation a decision someone must explain.