Las Vegas, October 13. Workday opened its Rising conference by announcing 'Autopilot' agents that work across hiring, payroll, accounting and compliance — and, in the same breath, a definitive agreement to acquire C1.ai, an identity platform whose whole job is to give each person and each agent the right access at the right time and remove it the moment the work ends. Read those two announcements together and the message is not 'our agents are smart.' It is 'our agents are controllable.' Workday's own release puts it in four words: agents that are 'lawful by design,' inheriting each company's policies, permissions and approval rules from day one. That is the sentence every team shopping for AI this quarter should pin above their desk — because the vendors who cannot say it are the ones Gartner expects to get cancelled.
1. The Market Just Admitted What the Bottleneck Is
For two years the pitch for AI agents has been capability: look what it can do. Today the largest enterprise HR vendor spent its flagship keynote on the opposite axis — what the agent is allowed to do, who approved it, and how fast its access disappears when the job is done. The C1.ai acquisition is the tell. You do not buy an identity-governance company for your agents unless your customers have told you, loudly, that an agent without a permission model is a liability they will not sign for. Deloitte's 2026 State of AI in the Enterprise survey of 3,235 leaders found close to three-quarters of companies plan to deploy agentic AI within two years — and only 21% have a mature governance model for it. That 53-point gap is the market Workday is pricing.
2. Why 'Lawful by Design' Beats 'Smart'
An agent that is merely smart is a very fast new hire with every door key in the building. It will do brilliant work right up to the moment it emails a candidate you have not cleared, pays an invoice nobody approved, or deletes the thing it was told to clean up. Gartner's June 2025 forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027 names three killers — escalating costs, unclear business value and inadequate risk controls — and two of the three are governance problems, not intelligence problems. The agent that survives procurement is the one whose every action can be traced to a policy a human wrote, scoped to a role a human defined, and stopped by an approval a human still holds.
3. This Is Exactly How Kavanah Was Built
We did not get here by watching a keynote; the permission model is the product. Every AI Employee in Kavanah runs inside capability scopes you assign, like a job description it cannot exceed. A safe mode strips external sends, payments and deletes from the run entirely. An action-policy engine lets you allow, budget-cap, park-for-approval or block whole categories of action — and this month we added the ability for a guardrail to block one specific tool for one specific AI Employee, so the CFO agent can read the books without ever touching the send button. Every action lands in a ledger you can audit, and the ones that can be reversed carry an undo. 'Lawful by design' is not a slogan here; it is the shape of the system.
4. The Identity Problem, Solved at the Source
Workday is buying C1.ai because, in most companies, nobody can answer 'what can this agent reach right now?' In Kavanah, the question has one answer per agent, and it is visible. An AI Employee's access to a project or a task runs through the same per-project and per-task access gate a human member's does, so a non-member cannot drive a write into a project it cannot see — and an agent cannot either. Connect a provider and the agent gains exactly those tools; disconnect it and they vanish from the agent's hands the same minute. There is no separate identity product to bolt on, because the workspace is the identity boundary.
5. What to Ask Every AI Vendor This Week
Workday has handed you the procurement script. Ask three questions of anything agentic you are evaluating this quarter. Can I see, per agent, the complete set of actions it can take? Can I put an approval in front of a category of action — not a single prompt, a category — and will the agent stop and wait rather than route around it? And when the agent does act, is there a ledger a human can audit afterward? A vendor that answers all three has built lawful by design. One that answers with a demo of how clever the model is has built the thing Gartner is counting toward 40%.
The AI agent market grew up today. The winners will not be the agents that can do the most — they will be the ones you can say yes to without holding your breath.
