Plan review capacity before expanding AI-assisted output

Count the work that moves downstream

Imagine a synthetic customer-operations team preparing policy-grounded responses. Before AI assistance, a specialist drafts and checks eight complex responses in a shift. A pilot produces 20 candidate responses in the same shift, but each still needs source confirmation, account-context review, and a final owner. If the team can complete only 12 reviews, the queue grows by eight. The apparent drafting gain is not a completed service gain. These numbers are a worked example, not observed Two Prune or customer results.

Start with a small time-and-flow map, not a broad productivity claim. Record how many items arrive, how many are routine, how many require a specialist, what fraction are returned for correction, and how long each review stage takes. Separate time saved in drafting from time added in verification, exception handling, and rework. The ILO's 2026 review of empirical GenAI workplace evidence describes uneven productivity findings and changes in work organization; it does not provide a universal review-capacity multiplier for a particular team. A local measurement is still needed.

The NIST AI Risk Management Framework Core emphasizes mapping the context of use, measuring relevant risks, managing responses, and monitoring over time. This voluntary framework supports examining how the workflow actually behaves, not merely counting generated outputs. Neither source says that adding reviewers always improves quality or that every AI-assisted response must follow the same process.

Sources: [1] [2]

Write the capacity contract in reviewable units

For the fictional team, the contract might have six rows. Intake: 20 AI-assisted drafts per shift, with a named source queue. Triage: a first reviewer distinguishes routine wording edits from claims about entitlement, policy exceptions, or account-specific facts. Evidence check: each consequential claim is matched to the current approved policy and the customer's actual record. Specialist escalation: uncertain or sensitive cases go to a qualified owner. Release: the responsible agent approves the final response and keeps the source version. Stop condition: if unreviewed work exceeds an agreed threshold or critical exceptions are unresolved, the team slows generation rather than sending unchecked drafts.

The contract's capacity calculation is deliberately transparent. Suppose the review team has 360 available minutes in a shift and a routine item averages 12 minutes of verified review. That is a theoretical ceiling of 30 routine reviews before meetings, interruptions, difficult cases, and rework. It is not a sustainable commitment. If five cases need 35 minutes each, those cases consume 175 minutes, leaving 185 minutes, or roughly 15 routine reviews at that assumption. A manager should plan with observed distributions and reserve, not one optimistic average. The example tests whether the review queue is the real constraint.

Pilot the flow and preserve a stop condition

Run a bounded pilot with synthetic or approved non-sensitive cases. Measure complete reviewed-and-released items, age of the oldest unreviewed item, corrections after review, escalations, and which reviewers were available. Sample both easy and adverse cases; a queue can appear healthy if difficult items are quietly deferred. Ask reviewers whether the AI output changed their task from composing to detecting subtle errors, and whether the work still allows sufficient time to inspect original sources. The ILO evidence is a reason to measure local work organization, not a substitute for doing so.

When the queue exceeds the agreed review capacity, reduce the volume of generated drafts, narrow the eligible case types, add qualified review capacity, or change the workflow. Do not silently turn “human review” into a click-through approval because a throughput target was missed. Revisit the contract after a policy, model, source, or customer-mix change. A completed pilot can show how this team handled these cases during this period; it cannot establish that all future outputs will be safe or that workers' performance changed in general.

Two Prune can use a synthetic review-queue scenario to observe how a participant frames the capacity tradeoff, checks evidence, and communicates a bounded recommendation. Its report supplies task-specific assessment evidence for people to interpret. It is not a workforce scheduling system, a quality certification, or an automated employment decision tool.