AWS: 'hours saved' overstates agent ROI; price exceptions and oversight too

AWS published a post on 7 October 2026 arguing that the standard business case for automation, hours saved times labour cost minus build cost, was designed for rule-based robotic process automation (RPA) and undercounts both the value and the cost of agents. The post, by AWS authors Manish Ballal and Sumit Wasuja, is on the AWS Machine Learning Blog and promotes Amazon Quick Automate, so it is vendor-published. The worked numbers are illustrative, not measured.

Why the old formula breaks

The post says the RPA-era model assumes stable processes, rule-based tasks, and that the automated work is the whole job. That leaves no line for maintenance when processes change, none for exceptions, and none for human oversight. It also treats a saved hour as banked, when AWS says freed capacity often refills with backlog and never reaches the profit and loss statement.

It cites McKinsey's 2026 "1:3:5 pattern": for every dollar spent on agentic technology, successful organisations spend three on process redesign and five on capability building and adoption, while the post quotes McKinsey as saying most companies invert this.

Four value pools, one condition

• Time savings: still counted, now over a larger base of work than fixed rules could handle.
• Exception handling: AWS planning ranges put the cost of correcting an error at 1.5 to 4 times the original transaction, and human error at 2 to 15 percent of operational cost.
• Decision quality: consistent policy application with a logged rationale, though AWS notes error rates need continuous measurement.
• Change resilience and maintenance: brittle scripts avoided, offset by evaluation, prompt, monitoring and model-operations costs that agents add rather than remove.

The condition on all four is value realisation: each benefit needs a mechanism that turns an operational gain into money, and an accountable owner.

The worked example

AWS uses a claims-triage process with illustrative figures: 200,000 claims a year, about 12 minutes each and $45 an hour fully loaded, a base cost of roughly $1.8 million. Automating the routine 70 percent releases about 28,000 hours, or roughly $1.26 million of capacity, before the oversight those claims still need. If attrition and lower overtime capture half, the case should carry about $630,000.

On top, 8 percent of claims (16,000) need correction. At 3.5 times the roughly $9 handling cost, AWS puts annual correction exposure near $504,000. A 40 percent reduction, adjusted by a 75 percent realisation factor, gives a modelled benefit near $151,000. The post's net-value line then subtracts implementation, runtime, integration, evaluation, oversight, governance and change-management costs, plus losses from new errors the agent introduces.

What the post leaves out

It gives no measured agent runtime or oversight cost, so the cost side of the formula is a list of headings, not numbers. The customer figures are self-reported by the customers and cover only some pools: Kitsa reports 91 percent cost savings and 96 percent faster data acquisition; dLocal reports automating up to 75 percent of merchant-compliance reviews in controlled evaluations; Genpact reports cutting disruption-impact analysis from 2-3 days to minutes. The post itself says none of them proves all four pools.

For prioritisation it proposes a two-by-two of task complexity against decision risk: keep low-complexity, low-risk work on RPA, automate high-complexity, low-risk work for throughput, and keep a human in the loop where risk is high.