Workday · Intake Agent
Enterprise governed agent
People ask the agent in chat. The answer cites the policy it used. Admin sets the route, and a person still approves anything that spends money.
A consumer LLM answers anything. In enterprise spend, each agent response has to be defined and constrained so it doesn’t hallucinate policy. Admin sets the route. Catalog and policy set the bounds. A person still approves anything that spends money.
Form-first worked only when the person already knew the request. Pure chat couldn’t keep the answer in policy or hand off the next step. Governed chat shipped.
The catalog check stays in the thread. A few matches, then the canvas.
Admin defines the rules. The agent can help generate the recipe. Every claim has to show a clear citation and the source it came from. Recipes start read-only, because editing one can knock out the citation.
Multi-item still doesn’t work. First release, checkout is view-only. If the job isn’t this agent’s, it goes to another spend agent.
Conversational input replaced the modal, and the catalog returns the top three before the canvas.
The cart stays view-only, multi-item still doesn’t work, and a wrong-agent ask goes to another spend agent.
Internal support said half their week was repeat “how do I…?” questions. I don’t have a public reduction number to share.