The “hours × rate” formula was made for RPA and undervalues AI agents. What to measure instead and how to budget costs in the business case.
AWS published a framework for AI centers of excellence: how to build a business case for agent automation.
The authors assume that the standard formula, “hours saved × labor cost - development cost,”
I agree with the diagnosis and take it further: evaluate an agent project by its impact on a business metric, while hours remain a secondary indicator.
An RPA robot repeats a person's clicks according to a fixed script.
A replaced click is easy to calculate: the number of monthly operations, the minutes per operation, and the employee's rate.
The process outcome remains the same, while the cost of execution falls. This formula worked for RPA. An agent works differently.
It reasons through the task, selects tools, and handles exceptions that the workflow did not anticipate.
It takes on work that was never automated before: analyzing an unusual email, checking a contract against regulations, or preparing a customer response based on three sources. There are no previously spent hours for this work: it was either not done or done selectively.
The formula counts zero here and undervalues the project.
AWS authors point to the value agents create beyond time savings. I group it into four categories, each with its own metric. Cycle speed. A request that previously waited two days in a queue between departments is processed in minutes.
Time-to-result metrics work here: approval time, time to product launch, and customer response time.
For us, this is TTU, time to use: a tool's value is measured by how quickly it delivers results.
An impressive architecture is not part of this assessment. Quality and errors.
The agent checks every item, while a person checks a sample under high workload.
The cost of a missed error in a contract, item record, or report can exceed an operator's annual salary.
It is measured by the defect rate and the cost of an incident. Coverage. Tasks that previously went unattended start getting done.
This is new revenue or mitigated risk, and the hours calculator has no line for either. Scale without hiring.
Doubling the workload does not require doubling the headcount.
It is measured by the cost of processing one unit of work as volume grows.
In an RPA case, costs ended with robot development and maintenance. Agents add easily overlooked items: token and model-call costs, answer-quality evaluation, access control, observability, and human review for disputed cases. If the business case includes development only, the project may look unprofitable after six months simply because it was calculated incorrectly.
For value to be measurable, an agent must leave an audit trail.
1. The baseline before launch. Cycle time, error rate, and unit cost are recorded for one or two months before implementation.
Without this, there is nothing to compare against. 2. A unified integration layer.
In AI-native integration, this layer is built on MCP servers and queues such as Apache Kafka, and every agent action is logged. 3. A gateway to models.
All requests pass through the LLM & Security Gateway: tokens and costs are counted for each process, personal data is removed, and access policies are applied.
These figures make up an expense line in the business case. 4. Answer grounding. The agent answers with a link to
. Qlik Answers on Amazon Bedrock is built around this requirement: employees need an answer they can verify, and in a regulated environment, an answer without a source is usually not accepted for use.
We address this with RAG, where every agent conclusion is linked to a document.
We often build orchestration for simple processes with n8n, while complex integrations use Python and enterprise service buses.
The choice depends on the number of systems in the process and audit requirements.
How to validate the business case before allocating a budget: - choose one metric the CFO can see: time, defects, or cost per unit of work; - record its current value; - agree on what value after one quarter counts as success; - include model costs, quality evaluation, and access control in expenses; - decide in advance at what result the project will be closed. The last point is the most uncomfortable. Without a closure threshold, a team can showcase a pilot at conferences for years while the metric remains unchanged.
Hours of labor saved suited robots that repeated clicks.
Agents change the process itself, so they should be evaluated by the process outcome.
The client pays for results, while an AI project that does not move a metric becomes theater for which those who launched it are ultimately held accountable.
A simple interface where an employee asks a question and receives a verifiable answer relies on complex engineering: integrations, a gateway, and an action log.
This engineering is what makes the result measurable.
AWS Machine Learning Blog, “Beyond hours saved: Building the business case for agentic automation”: https://aws.amazon.com/blogs/machine-learning/beyond-hours-saved-building-the-business-case-for-agentic-automation/. Additionally: “How Qlik built grounded, enterprise-scale AI with Amazon Bedrock” on the same blog.
