Employee-built agents and tacit-knowledge capture can spread implementation ability inside organizations while transferring know-how into reusable machine systems. The distinction is who controls that system and its later use, not the number of agents created.
UKG, Wipro, Valon and Caterpillar illustrate different ways workers create or supervise AI-supported workflows. Creating an agent does not establish ownership, a binding say in deployment, compensation or portable skills. Expanded capability and later staffing compression both remain possible.
UKG reports workforce-built agents and applications, multi-provider use, avoided long contracts and autonomous handling of 27% of customer inquiries. The case suggests distributed workflow design without established worker control of deployment.
Wipro reports substantial employee-built agents within a Microsoft-mediated stack; its CTO separately estimates 20,000-worker-equivalent capacity and says it is redeployed. These are one correlated firm/programme case.
Valon delays generative-AI access for new non-engineering employees until managers judge them able to identify errors. It treats expertise formation as an implementation prerequisite.
Caterpillar describes experienced operators training autonomy and a developing path to supervising multiple machines. Tacit knowledge and supervisory leverage can complement labor while reducing labor per machine.
VideoAmp’s August restructuring cut about 20% of staff, combined product/engineering control and reportedly did not refill the standalone CTO role, with an explicit agentic-AI strategy.
An October 8 OpenAI case study describes Oracle employees building AI-supported recruitment, analytics and engineering workflows. Oracle reports approximately 130,000 active ChatGPT users and more than 95,000 active Codex users.
A recruitment tool uses a job description to research comparable roles, pay and local talent pools. For business analytics, internal knowledge about data and its relationships is encoded so that a plain-English request can become a database query. The case reports a 98% reduction in recruitment-research time, but it is a vendor/customer account without an inspected independent evaluation of that estimate. Vendor/customer case study ↗.
The mechanism is more useful than a generic adoption percentage: people define the workflow and encode domain knowledge, while the model supplies part of the execution. That can broaden who performs analysis, while leaving significant dependency on the provider.
Unclear for workers: building agents may increase useful skill without giving control over the encoded expertise or later deployment.
What would change this assessment
Worker participation, ownership of work products, portable skills, compensation and the ability to change providers would show whether new responsibilities bring greater control.