One of the better recent business cases for AI adoption is not about flashy consumer copilots. It is about back-office work that most companies would never put on a keynote slide. In a customer story published on May 27, 2026, AWS described how Works Human Intelligence (WHI) worked with the AWS Generative AI Innovation Center to build two operational AI agents for HR support workflows. The headline result was striking: the browser-based agent’s cost per process fell by up to 97% while the system became capable of handling more complex tasks.
That result matters because it points to a more mature stage of enterprise AI. Plenty of companies can get a pilot to work. Far fewer can make an agent economically viable at production scale. WHI’s case is useful precisely because it shows the harder part: taking an agent workflow that functionally works, then redesigning the architecture and prompt economics until the unit cost starts to make business sense.
What WHI Actually Built
WHI develops and supports the integrated HR system COMPANY for major Japanese corporations and public-interest organizations. The AWS case centers on two operational agents tied to that environment.
The first was a Commuting Allowance Agent that automates approval work around employee commuting-allowance requests, a routine but repetitive workflow that surfaces when employees move or need changes to their benefits setup. The second was a Browser Operation Agent that can access the HR system, inspect the current state, prepare proposed changes, ask the human for approval when needed, and then execute those changes while collecting evidence.
That second workflow is the more interesting one from a business perspective. It is not just Q&A or summarization. It is an agent performing a multi-step operational job in a controlled system of record. That makes it much closer to the real enterprise AI opportunity than another internal chatbot.
Why This Case Is Strong
First, the economics are unusually concrete. AWS says the browser agent’s processing cost dropped from $14.5 to $2.1 after enabling prompt caching for user messages, then to $1.0 after optimizing sub-agent behavior, and finally to $0.4 after shifting from Claude Sonnet 4.5 to Haiku 4.5. That sequence is important. It shows that agent ROI did not come from one magic model change. It came from layered optimization across workflow design, prompt structure, and model selection.
Second, WHI and AWS changed the architecture, not just the prompts. The Commuting Allowance Agent started in a more monolithic LangGraph and ECS setup. The teams then broke the system into sub-agents running independently on AgentCore Runtime. That matters because enterprise AI often fails when teams keep forcing new workloads into old application shapes. A modular agent architecture makes future expansion easier, improves observability, and creates cleaner boundaries for authentication, tenancy, and cost tracking.
Third, the use cases are narrow enough to govern. Commuting allowances and browser-driven HR operations are not vague productivity aspirations. They are bounded processes with clear handoff points, explicit data sources, and obvious human-review checkpoints. That is exactly the sort of workflow where AI adoption tends to work better. The business logic is understandable. The exceptions are knowable. And the success criteria are measurable.
The most important metric here is not just 97%. It is that WHI kept making the workflow cheaper while still handling more complex tasks.
Fourth, the case shows that agentic AI is an operations discipline. AWS notes that WHI had already reduced browser-operation tokens by 88% before the newer optimization work, but the earlier system still depended on a more proprietary implementation that made migration and further improvement harder. The newer setup adds managed observability and a clearer runtime foundation. In plain English: the team got beyond “we made it work” and moved toward “we can operate this reliably.” That is the difference between a demo and a business system.
What Business Leaders Should Learn From It
The first lesson is that agent adoption should start where tasks are repetitive, rules-based, and expensive to staff manually. HR support, finance operations, compliance reviews, quote generation, claims handling, and customer-service exception routing all fit this pattern better than wide-open creative work.
The second lesson is that cost discipline has to be designed in early. Too many AI rollouts treat inference cost as something to clean up later. WHI’s case suggests the opposite. Token budgets, context windows, prompt caching, tool outputs, and model routing are part of the product, not an afterthought. If an agent cannot clear an economic threshold per process, adoption will stall no matter how clever the demo looks.
The third lesson is that human approval still belongs in the workflow. The Browser Operation Agent does not blindly make sensitive HR changes without review. It builds a proposed change, presents it to the user, and only acts after approval. That is a practical model for enterprise deployment. In high-value workflows, AI should often compress the work to a higher-quality decision point rather than remove the human completely.
The fourth lesson is that observability is part of ROI. AWS says WHI reduced operational burden by moving from self-hosted logging checks toward AgentCore Observability. That may sound like a technical detail, but it directly affects business value. If every agent requires heavy manual debugging and custom monitoring, cost savings on the front end get eaten by engineering overhead on the back end.
The Caveats
This is still a vendor-led case study. The numbers come from AWS’s account of a joint project, not an independent audit. We do not have a full deployment cost, a labor-savings figure translated into annual dollars, or a complete before-and-after productivity benchmark across WHI’s organization.
There is also a scope warning. This is not yet evidence that every company should deploy autonomous agents everywhere inside HR. It is evidence that specific operational tasks with strong boundaries can be automated economically when the system is designed carefully. That is a narrower, but more useful, conclusion.
And there is a transferability question. WHI is itself an HR-software provider serving large organizations, so it has stronger domain structure than the average business. Companies should copy the pattern, not the exact workflow. The pattern is: pick a repetitive task, break it into controlled steps, measure cost per process, and keep optimizing until the agent becomes cheaper and more trustworthy than the old method.
The Business Takeaway
WHI is a strong AI adoption case because it shows how companies move from pilot enthusiasm to operating leverage. The breakthrough was not simply “we used AI in HR.” The breakthrough was building the surrounding system so the agent could run with lower cost, better visibility, and clear human controls.
If you are evaluating AI agents for your business, do not ask only whether the model can complete the task. Ask whether the unit economics, approval flow, and operating model make the task worth automating at scale. That is where real AI adoption succeeds or fails.