Severity-adjusted error impact modeling
Error impact must account for asymmetry and reversibility
Original excerpt
Impact(error) = s × C, where 0 < s ≤ 1
In An Agentic World Where Automation Gets Cheap, Which Work Is Worth Routing to a Human?
Scale Blog ·
A source examining how to allocate human review in automated financial workflows using dollar-denominated thresholds, risk-adjusted confidence boundaries, expected-loss calculations, severity-adjusted error impact modeling, calibrated confidence scoring, capacity economics of human versus agent labor, and a historical analogy to Jevons’ Paradox. Read 6 viewpoints with supporting evidence and source links.
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki
Lines show the reading structure. Select an idea to read its explanation and evidence.
Synthesis
Finance workflows enable dollar-based human review thresholds
In finance—especially credit recovery—the value of a decision and the cost of human time are both denominated in dollars, enabling quantitative determination of when human review is economically justified. Regulatory or policy-driven decisions are handled as fixed rules, not priced.
Supporting evidence 1
Original excerpt
Both sides are already denominated in dollars, so the boundary at which human review becomes worthwhile can be set directly.
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 6
Read in source context →Context
Finance is often such a domain, particularly credit recovery. A credit is either recovered or it is not, and an analyst’s hour is either spent on one statement or another. Not every decision works this way. A CFO certifies quarterly statements because the law requires it, not because their review is the cheapest way to catch an error, and some decisions inside this workflow are similarly governed by regulation, internal policy, or the vendor relationship. Those are handled as rules rather than priced.
Continue exploring
Dollar-denominated review thresholds in finance →Human review allocation should be risk-adjusted and prioritized
Maximizing automation share is no longer the objective; instead, limited human-review capacity should be allocated to decisions where intervention yields the greatest expected value, using a risk-adjusted boundary that increases required confidence as potential error impact rises.
Supporting evidence 1
Original excerpt
The resulting boundary is risk-adjusted rather than fixed: as the potential impact of an error increases, the confidence required for autonomous action should increase with it.
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 8
Read in source context →Context
Maximizing the share of work that is automated is no longer the binding objective. The question is which slice of an expanded universe of work deserves human attention. We argue that in finance workflows this boundary can be located quantitatively, by pricing human review against the expected value of the decisions that review improves. We develop this framework using an accounts-payable credit-recovery workflow as the motivating case and show how it can be used to allocate limited human-review capacity toward the decisions where intervention has the greatest expected value.
Continue exploring
Risk-adjusted human review allocation →Review is warranted when expected loss exceeds review cost
Human review is economically justified for a decision when the product of error probability and error impact exceeds the cost of review. Work failing this test is a candidate for automation—or exclusion from the pipeline.
Supporting evidence 1
Original excerpt
P(error) × Impact(error) > Cost(review)
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 16
Read in source context →Continue exploring
Expected loss versus review cost formula →Error impact must account for asymmetry and reversibility
In credit recovery, false negatives incur near-full loss while false positives cost only minutes of analyst time. Error impact must be modeled as severity-adjusted, where severity reflects reversibility and downstream detection likelihood.
Supporting evidence 1
Original excerpt
Impact(error) = s × C, where 0 < s ≤ 1
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 22
Read in source context →Continue exploring
Severity-adjusted error impact modeling →Route human review by calibrated confidence and error cost
The boundary between agent and human work should be determined not by accuracy alone, but by calibrated confidence evaluated against the financial consequences of error—converted into a dollar threshold for escalation to human review.
Supporting evidence 1
Original excerpt
The framework developed here locates the boundary based on calibrated confidence, which is evaluated against the financial consequences of error and converted into a dollar threshold at which the agent would escalate.
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 43
Read in source context →Context
The reduction in the cost of knowledge work produced by agentic systems expands the volume of work worth performing and shifts the strategic question from the extent of automation to the placement of human judgment within it. This boundary cannot be determined by accuracy alone, because the value of human review depends on both the probability of an error and the consequence of that error. The result is a system that directs human attention toward the decisions in which it creates the greatest value, rather than applying a uniform review threshold.
Continue exploring
human-in-the-loop routing →Design human teams for high-severity decisions; agents for volume and variance
Because human review capacity scales slowly—requiring hiring, training, and fixed costs—while agent capacity is elastic and low marginal cost, an economically sound routing framework sizes human teams for highest-severity decisions and assigns agents to both volume and variability.
Supporting evidence 1
Original excerpt
Human review capacity is slow to adjust in either direction. It is added one hire at a time, requires months of training and accumulated institutional knowledge, and is paid for in quiet periods as well as busy ones. Agent capacity is elastic, scaling from a handful of documents to many thousands without a corresponding change in cost structure. A framework that prices review correctly will therefore tend to recommend a human team aimed at the highest-severity decisions, with the agent absorbing the volume and the variance that a fixed team cannot economically carry.
Benjamin Chen, Duncan McKeen, Manik Mukherjee, Sara Bolouki · Paragraph 42
Read in source context →Context
One further property generalizes.
Continue exploring
capacity economics →
Key passages6
Attributed passages with the context to verify them. Open the original text to check the source.
human-in-the-loop routing
Route human review by calibrated confidence and error cost
Original excerpt
The framework developed here locates the boundary based on calibrated confidence, which is evaluated against the financial consequences of error and converted into a dollar threshold at which the agent would escalate.
Context
The reduction in the cost of knowledge work produced by agentic systems expands the volume of work worth performing and shifts the strategic question from the extent of automation to the placement of human judgment within it. This boundary cannot be determined by accuracy alone, because the value of human review depends on both the probability of an error and the consequence of that error. The result is a system that directs human attention toward the decisions in which it creates the greatest value, rather than applying a uniform review threshold.
Dollar-denominated review thresholds in finance
Finance workflows enable dollar-based human review thresholds
Original excerpt
Both sides are already denominated in dollars, so the boundary at which human review becomes worthwhile can be set directly.
Context
Finance is often such a domain, particularly credit recovery. A credit is either recovered or it is not, and an analyst’s hour is either spent on one statement or another. Not every decision works this way. A CFO certifies quarterly statements because the law requires it, not because their review is the cheapest way to catch an error, and some decisions inside this workflow are similarly governed by regulation, internal policy, or the vendor relationship. Those are handled as rules rather than priced.
Risk-adjusted human review allocation
Human review allocation should be risk-adjusted and prioritized
Original excerpt
The resulting boundary is risk-adjusted rather than fixed: as the potential impact of an error increases, the confidence required for autonomous action should increase with it.
Context
Maximizing the share of work that is automated is no longer the binding objective. The question is which slice of an expanded universe of work deserves human attention. We argue that in finance workflows this boundary can be located quantitatively, by pricing human review against the expected value of the decisions that review improves. We develop this framework using an accounts-payable credit-recovery workflow as the motivating case and show how it can be used to allocate limited human-review capacity toward the decisions where intervention has the greatest expected value.
capacity economics
Design human teams for high-severity decisions; agents for volume and variance
Original excerpt
Human review capacity is slow to adjust in either direction. It is added one hire at a time, requires months of training and accumulated institutional knowledge, and is paid for in quiet periods as well as busy ones. Agent capacity is elastic, scaling from a handful of documents to many thousands without a corresponding change in cost structure. A framework that prices review correctly will therefore tend to recommend a human team aimed at the highest-severity decisions, with the agent absorbing the volume and the variance that a fixed team cannot economically carry.
Context
One further property generalizes.
Expected loss versus review cost formula
Review is warranted when expected loss exceeds review cost
Original excerpt
P(error) × Impact(error) > Cost(review)
Source & methodology
These viewpoints are linked to their original sources. Paraphrases are labeled and are not verbatim quotes.
Open transcript or source material (opens in a new tab)Report an issue