Time saved at a desk is not capacity released
A shorter individual task does not establish extra organisational capacity. Follow the work through queues, reviews and bottlenecks, and value working-pattern improvements separately from increased output.
Keep development and test groups separate before comparing performance. No measured results are shown.
A productivity estimate becomes a capacity claim only when the organisation can show where the released time goes. In an illustrative office-assistant trial, faster drafting may leave the service queue unchanged. The sponsor needs evidence of throughput, waiting time or additional work, alongside the individual experience.
An individual time saving is a useful outcome. Organisational capacity requires a further observation: work has moved through the system differently without an unacceptable transfer of effort or harm. Name the intended change before estimating its financial value.
What the field experiment measured
In Shifting Work Patterns with Generative AI, researchers randomly assigned access to a generative AI tool across 66 firms and 7,137 knowledge workers. Among treated workers who used it in the latter half of the six-month experiment, they found less time spent on email and less work outside regular hours. They did not detect a change in the quantity or composition of workers’ tasks from individual-level provision.
That is a precise result, not a verdict against the tool. Reducing work outside regular hours may itself matter. A worker may use time for better attention, breaks or quality that the task categories did not capture. What the study does not support is converting a personal time saving directly into an organisation-wide output figure. Its setting also does not establish the result for every role, tool or firm.
The existing article Most pilots return nothing asks whether a pilot has a decision and a credible baseline. This article focuses on the next measurement boundary: once an individual benefit is observed, what evidence shows it became organisational capacity?
Instrument the workflow, not just the user
Choose the workflow and measure arrivals, completions, backlog, review and quality under the existing process. Identify the actual constraint. Saving an analyst’s drafting time may not increase completions if approval is the bottleneck, and it may create checking work elsewhere.
Define the benefit the business intends to buy. For case resolution, measure completed cases and quality. For reduced late work, measure that working pattern. For new demand, record which additional work is accepted. Use a measure that corresponds to the claim rather than converting every benefit into spare staff capacity.
A capacity claim needs completed work
- Baseline
- The existing queue's throughput, backlog, review time and quality before access changes.
- Outcome
- Cases completed or waiting time at comparable demand and service quality.
- Guardrails
- Rework, reviewer load, customer complaints and work outside regular hours.
- Decision rule
- Expand only where the measured workflow improves without shifting unacceptable work or harm to another team.
- Decision owner
- COO or service owner
- Population
- The selected work queue, intended users and observation period.
Representative method for this article, not a measured deployment result.
Baseline: The existing queue's throughput, backlog, review time and quality before access changes. Decision rule: Expand only where the measured workflow improves without shifting unacceptable work or harm to another team.
Reviewed 2026-10-02
Keep a credible comparison through the rollout. Staged access, matched teams or a suitable control can help separate the tool’s effect from staffing, policy or demand changes. Record adoption, but assess the effect of offering the tool as well as selected users’ experience. Active users may differ from others before the intervention.
Suppose the assistant cuts drafting time for claims analysts while senior reviewers still approve every case. If each reviewer already has a full queue, cases accumulate at that review step. The sponsor can measure the analysts’ experience as a benefit, but cannot count those minutes as additional settled claims. To make a capacity claim, the firm would need to change the review process or its staffing, then observe more completed cases at comparable quality and demand.
This is why the cost side belongs in the same measurement. Include licences, integration, support, extra review and the time staff spend correcting outputs. If the new workflow moves ten minutes from an analyst to a reviewer, the local gain may be real while the service gain is smaller or absent. If it reduces late work without raising throughput, call that a working-pattern improvement. It may be worth paying for, but it should be valued as such.
Set a review date before launch. At that point, the owner can choose to redesign the bottleneck, continue for a documented employee benefit or stop the rollout. Leaving the trial running because people report that it feels faster turns a measurement into a habit, while the promised capacity remains untested.
What this does not tell you
The field experiment does not prove that AI cannot create capacity. It shows a case in which detectable task-level reallocation did not follow individual access during the observation period. A queue with a clear bottleneck and an explicit redesign may respond differently. The measures above also require care: faster completions can mask lower quality or deferred work.
The service owner should agree the valued benefit and its evidence before expanding access. Retain employee experience as an outcome where it matters. Where the business case promises capacity, require an operating measure showing the additional work the service can complete at the agreed quality.