# Staff Before You Escalate

Source: https://dinand.com/research/staff-before-you-escalate/
Author: Dinand Tinholt
Published: 2026-09-05

> Routing more work to review only helps if the reviewers grow with it: escalating with headcount fixed scores −0.59, and with staffing scaled +1.15.

Staff Before You Escalate

Executive brief · follows "Pricing Agent Autonomy"

# Staff Before You Escalate

We re-ran the delegation study with reviewers who tire and queues that back up. The finding for operators fits in one sentence: routing more work to review only helps if the people doing the review grow with it, and sustained utilisation of 0.8 is the line to manage to.

90,568 simulated configurations · verified against the raw ledger · September 2026

Dinand Tinholt · the full paper is [The Oversight Yield Curve](https://dinand.com/research/oversight-yield-curve/) · earlier study: [the Delegation Cliff campaign](https://dinand.com/research/campaigns/delegation-cliff/)

Our published study told organisations that reviewer capacity is the largest positive lever in agentic-AI operations, worth +1.46 on the delegation margin. A fair objection existed: reviewers in that model never tired, so the lever could never saturate. We rebuilt the model with a live reviewer pool. Escalations queue, service slows as sustained load rises, and backlogs compound. Then we swept 90,568 configurations and measured what happens to the lever.

Configurations, zero error units

90,568

Escalating more, headcount fixed

−0.59

Escalating more, staffing scaled with it

+1.15

Capacity lever below utilisation 0.8

+0.93

## Staff before you escalate

Sending a larger share of agent output to a review team of fixed size made outcomes worse in every region of the space we measured, at a slope of −0.59 across the swept range. The identical routing decision paired with proportional staffing paid +1.15. The intervention most organisations reach for first, tightening review policy, carries a negative sign on its own. It earns its positive sign the moment headcount moves with it.

## The sign of "more oversight" depends on whether capacity scales

A circulating research claim says oversight follows an inverted-U, so adding review can make systems less safe. In our space the picture is simpler and harsher. More oversight against fixed capacity was harmful from the first increment onward, with no initial benefit. Once staffing scaled alongside it, the same sweep turned into a monotone gain. Before your next governance decision, ask one question of the proposal: does review headcount move with review load? The answer determines which curve you are on.

## Utilisation 0.8 is where the geometry turns

Operating margin vs sustained reviewer utilisation

*41,024 fully-swept configurations, binned by the review team's sustained (7-day average) utilisation. The margin crosses zero between 0.8 and 1.0 and keeps falling through deep overload, reaching −0.93 in the worst bins. Source: oversight-yield-curve campaign ledger, 2026-09-05.*

Below 0.8 sustained utilisation, every unit of review capacity you add returns +0.93 of margin, and it remains the best organisational lever available, ahead of buying a better model. Above 0.8 the organisation crosses into a regime where the margin goes negative and slides. The practical instrumentation is cheap: track your review team's rolling utilisation the way you track service uptime, and treat 0.8 as a paging threshold. Teams living above it are past the point where dashboards or policy help; the buy-back at that point is headcount.

## Method, in one paragraph

This is a simulation study. We extended our published 12-dimension model of agentic-AI operations with a reviewer pool that fatigues under sustained load, swept 15 organisational dimensions across 90,568 configurations, and scored each on the margin between delivered quality and a 95% quality floor. A slice of the design with fatigue disabled reproduces the published campaign's coefficients to the third decimal, so the changes reported here come from endogenous reviewer saturation and from nothing else. All findings were reproduced from the raw ledger by an independent verification pass before this brief was written.

---
Site guide for AI agents: https://dinand.com/llms.txt