Workday's chief executive says the company is still learning how to price AI agents

Bhusri was direct about pricing. "We're still learning. I'll be honest; it's a new muscle for us," he said ^[3]. His stated worry is customer overspend. "What I want to make sure is that customers spend the amount of money they want to spend and not more than that," he said. "That's the danger with tokens; you get carried away." ^[4] He said he had done it himself with Sana, asking more and more questions until he realised "I probably just ran up a huge bill" ^[5].

The controls shown on stage address that worry from the customer's side. Alongside the budget alerts in the demo ^[6], the autopilot agents release lets customers "set how often agents run, so they can manage both activity and cost" ^[7]. Both depend on the customer setting them.

Liability is a different matter, and the record there is thinner. The writer who asked the question described autopilot agents that will soon run payroll, accounting and compliance in the background and, with C1, have their own identities and access rights ^[8]. Bhusri called the question "a very deep question" and began from the principle that "there's always a human in the loop to make sure that the right things are being done, but the human doesn't have to do all the mundane tasking" ^[10]. The account available to us ends partway into that answer and does not include Workday's rates.

So the record supports an admission on pricing. On liability it supports a design principle from Bhusri, and whether Workday has contract terms behind that principle is not known. A principle of human oversight describes how the system is meant to run. The question put to him asked how a loss would be divided among three parties, which is a separate thing.

There is also a tension inside the two descriptions. The agents are described as working in the background ^[8], while Bhusri's principle keeps a human in the loop but relieves that person of routine tasks ^[10]. In our view, where that person sits in the approval chain is what a contract would have to settle before anyone can say whose error a wrong journal entry is.

Kazmaier supplied the nearest thing to accountability machinery. By his account, the largest obstacle CIOs return to again and again is securing AI and being confident it acts within the law ^[11]. "That's why the acquisition of C1 is so important for us," he said, because Workday needs "a completely new security model" for agents ^[12]. He added that "the CIO holds the keys to the kingdom for data access, privilege management, and security" ^[13].

Identity and access rights are what any later allocation of fault would rest on: which agent acted, and what it was permitted to touch. For the first bills, the protection on record is the budget alerts. For the first errors, it is the identity model Kazmaier described and the principle Bhusri stated.