An agent can release 28,000 hours and save half the money promised. Hours are capacity, not value

A steering committee approves an agent for claims intake. The business case says it will release 28,000 hours a year, worth about $1.26 million. A year later the agent works as designed and the hours were released. Finance finds about half of the money.

Nobody failed. The other half went where unassigned time always goes: into slightly slower days, slightly more meetings, slightly better work that nobody measures.

The numbers in that scenario are not ours. They come from an illustrative claims example in a framework AWS published on October 7, and the authors, Manish Ballal and Sumit Wasuja, are blunt about the lesson: “freeing hours isn’t the same as saving money.” In their example, 200,000 claims a year at twelve minutes each, 70% automated, release about 28,000 hours. Only around $630,000 of the $1.26 million is realized, through attrition and reduced overtime.

Hours are an input, not a result

The hours-saved model was designed for rule-based automation. AWS lists what it misses for agents: maintenance as processes change, exceptions, the cost of human oversight, and the assumption that saved hours turn into profit and loss on their own.

Our view, and it is an argument rather than a measured result, is stronger. Hours saved are capacity. Capacity becomes value only through an operating decision: a cost line that shrinks, or a named outcome with its own owner and its own measure. Without that decision, the agent can work perfectly and the business case still fails.

1. Hours released by the agent
2. Destination named before go-live
3. Cost cut or outcome measured
4. Minus run costs: tokens, evaluation, oversight
5. Net value finance can trace

• The agent: Hours released by the agent
• Management decides: Destination named before go-live · Cost cut or outcome measured
• Finance measures: Minus run costs: tokens, evaluation, oversight · Net value finance can trace

Produced by the systemAn operating decision

Where the larger value often hides

The same framework points to value that hours miss entirely.

Exceptions. AWS puts the cost of correcting a transaction at 1.5 to 4 times the cost of handling it once. In its claims example, 8% of claims needing correction represent about $504,000 a year of exposure. An agent that cuts errors may be worth more than one that cuts minutes, which is the redesign we described in Chatham Financial’s evidence-first trade validation.

Decision quality. Consistency and accuracy at scale are a separate case, justified on their own terms rather than as cost reduction.

Change. Rule-based scripts break when the process changes. The framework nets the maintenance an agent avoids against what it costs to run.

Where the costs hide

The same week brought two reminders that the cost side moves too. Anthropic cut the price of Claude Haiku 5.5 by 90% below 100,000 tokens, which we discussed as an architecture decision about the price line. Cheaper tokens help, but tokens are one line in AWS’s cost formula, next to implementation, integration, evaluation, oversight, governance and change management. In regulated work, human review is a requirement, not an option, so oversight is a fixed cost no price cut removes.

Google, meanwhile, extended pay-as-you-go usage above pooled quotas to more Gemini Enterprise editions. A license is becoming the floor of the bill, not the ceiling, which is why agent budgets belong next to agent permissions.

And time metrics need their whole process. Cornerstone OnDemand reports database diagnosis falling from about 45 to 10 minutes. That is a real gain in one segment of an incident. The value to customers depends on whether resolution time falls with it.

A business case that survives its second year

1. Name the destination of every released hour before go-live: a cost that shrinks, or an outcome with an owner and a metric.
2. Count each benefit once. AWS’s rule is worth copying: count redeployed value or a cash cost reduction, not both.
3. Put exceptions in the base case. Measure the correction rate before the agent and after.
4. Price the run, not the token: evaluation, oversight, governance and metered usage beyond the license.
5. Measure the whole process, not the step the agent touches.

None of this argues against agents. It argues against approving them on a number that measures activity instead of value. The second year is when finance looks for the money, and the decision about where the hours go has to be made in the first.