Reserve the exception budget before expanding AI intake

Overview

Decision memo for an operations lead

Recommendation: In the hypothetical service business below, do not double automated intake until the team has funded enough capacity to resolve exceptions. Approve a limited intake ceiling and a measured trial of better routing instead. Human exception handling is the constraint.

This memo uses invented operating figures in US dollars. They illustrate a capacity decision, not measured customer results, vendor pricing or a financial forecast.

The proposal on the desk

The team processes 1,000 requests each week. An AI workflow handles routine categorization, but 12% of requests need a person to resolve missing information, contradictory instructions or an unsupported case. An exception takes eight active minutes on average. The team has twenty hours available each week for this work.

Original capacity table: hypothetical averages, excluding waiting for customer replies.

Four spare hours cover only thirty average exceptions. Meetings, staff absence or a more difficult case mix can consume that allowance. Count active handling time separately from elapsed waiting time; otherwise a slow customer reply may be mistaken for an understaffed team.

Now double intake to 2,000 requests without changing the exception rate. Demand becomes 240 exceptions, or thirty-two hours. Existing capacity handles only 150 average cases. Ninety unresolved cases accumulate each week. A cheaper model call cannot remove those twelve missing staff hours.

AWS’s queue-backlog guidance explains how arrivals above processing capacity can create persistent delays and work that finishes too late to be useful. Applying that engineering lesson to a human exception queue is our operational analogy, not an AWS staffing formula. AWS Builders’ Library, published December 5, 2019; verified October 29, 2026.

Three choices for the next budget meeting

Assume $40 per active staff hour. Compare alternatives before expanding intake.

Original option table: rounded arithmetic; improvement scenarios are unproven, and implementation costs are not included.

The apparent zero-cost option runs at full nominal capacity. It leaves no room for variability. Likewise, promising an 8% exception rate is not a budget saving until a trial shows which cases disappeared and whether they were resolved correctly. Quietly passing difficult cases through would improve the percentage while worsening the service.

Prefer a trial that removes one recurring cause, such as asking for an omitted reference before submission. Record engineering effort and ongoing maintenance separately. Judge the trial on fewer unresolved cases and acceptable outcomes, not a lower exception label count.

Approval conditions

For the illustration, cap intake at 1,000 weekly requests while collecting a representative handling-time distribution. Establish who can accept overflow, which commitments require priority, and when new work must receive an honest revised turnaround.

Original decision table: proposed controls, not automatic authority to reject customer work.

Do not delete inconvenient cases to protect the dashboard. Preserve the request, its obligation and its resolution path. Expansion becomes reasonable when observed exception demand fits staffed capacity with deliberate headroom. The budget meeting should end with an intake limit and funded resolution plan, not merely approval for more AI usage.