Agents in the Wild

Works Human Intelligence cut its HR browser agent's cost per process by up to 97%

Japanese HR software maker Works Human Intelligence built two support agents on Amazon Bedrock AgentCore. One handles commuting allowance requests; one operates its HR system in a browser and changes data only after user approval.

Organization
Works Human Intelligence
Industry
HR software
When
May 2026
Evidence
Vendor case study
Agents and tools
Amazon Bedrock AgentCoreStrands AgentsLangGraphPlaywright MCPClaude Haiku 4.5Claude Sonnet 4.5

Reported outcome

AWS reports the browser operation agent's cost per process fell by up to 97% after prompt caching, shorter sub-agent prompts and a switch from Claude Sonnet 4.5 to Haiku 4.5.

  • down by up to 97% (AWS blog)
    Cost per process
  • $14.5 → $2.1 → $1.0 → $0.4 (AWS blog)
    Cost steps (no units given)
  • 88% reduction, before AWS joined (AWS blog)
    Browser tokens, earlier work

Caveats

Published by the company that sells the agent. Treat numbers as marketing until confirmed elsewhere.

  • Written on the AWS blog, partly by AWS staff who worked on the project. AWS sells the platform.
  • The result is about cost, not quality. The post gives no accuracy, error or failure rates.
  • The cost figures have no units or volume, so you cannot tell the total saving.
  • The post does not describe a human review step for the commuting allowance agent.

What the agents do

Works Human Intelligence (WHI) makes an integrated HR system for large companies. With the AWS Generative AI Innovation Center, it built two agents for customer support work.

  • Commuting Allowance Agent: automates approval of commuting allowance applications, a routine task that spikes when employees move. Slack is the entry point. Sub-agents run separately on AgentCore Runtime.
  • Browser Operation Agent: opens the HR system in a browser for a customer. It finds an operation template in a knowledge base, fills it in to make an operation manual, checks the current data, and drafts a change proposal. It makes the change only after the user approves. It asks questions when instructions are unclear.

What happened

Most of the post is about cost. The team added prompt caching for user messages, shortened sub-agent prompts, and moved the browser agent from Claude Sonnet 4.5 to Claude Haiku 4.5. AWS reports the cost per process fell by up to 97% overall. Before AWS joined, WHI had already cut browser operation tokens by 88%.

What to copy

  • Make the agent write a change proposal and wait for approval before it touches customer data.
  • Try a smaller model once the workflow is stable. Measure quality before and after.
  • Use prompt caching for long, repeated instructions.

Starter prompt

Written by us, not by WHI or AWS.

You operate our HR system for a customer. Task: {task}
Operation template: {template}

1. Fill in the template to make step-by-step instructions.
2. Read the current values and list them.
3. Write a change proposal: field, old value, new value.
4. Stop and ask the user to approve. Do not change anything before approval.
If any instruction is unclear, ask one question first.

Sources

  1. Building AI agents for business support using Amazon Bedrock AgentCore — AWS Machine Learning Blog, May 27, 2026

Added Nov 2, 2026 · Sources last checked Nov 2, 2026. Spotted an error? Send a correction.

Get new use cases weekly

One short email. Sources included.

Related use cases