AI Drove More Than a Quarter of Workday's New Sales. Its CEO Credited the Guardrails.
Workday says the controls around its agents are part of why customers trust them. Gartner predicts governance failures will force many companies to pull autonomous agents back from production. The common issue: companies need to know what an agent can do, who authorized it, and what gets recorded afterward.
Chris Hart, Chief Executive Officer
||9 min read|
tl;dr
Workday says AI drove more than a quarter of its new sales, and its CEO credited the controls around the agents, not the models. Gartner predicts 40 percent of enterprises will pull autonomous agents back by 2027 over governance gaps found in production. Three controls decide whether an agent ships: what it may do, who approved that, and what gets recorded. Ask any vendor to show you a blocked action.
Workday says the controls around its agents are part of why customers trust them. Gartner now predicts that governance failures will force many companies to pull autonomous agents back from production. The common issue is straightforward: companies need to know what an agent can do, who authorized it, and what gets recorded afterward.
Workday reported its fiscal 2027 second quarter on August 27. AI drove more than 25 percent of new annual contract value, and more than 5,500 customers are now using at least one of Workday's own agents.
Then Aneel Bhusri, the co-founder and chief executive, explained the numbers this way: "Because of Workday's deterministic rails, customers can trust our agents with the work that matters, and you're seeing that in the numbers." (Workday, August 27, 2026)
The important part is what Bhusri chose to credit.
Workday had strong AI numbers to point to and could have emphasized model quality or benchmarks. Instead, Bhusri pointed to the controls around the agents and the trust those controls create.
That does not prove why every customer bought. But it is a useful signal about what Workday believes is helping move agents into real use.
Why the wording matters
An earnings-call quote is not independent evidence of product quality. Workday has every reason to frame its results favorably. But the framing itself is useful: the explanation a CEO gives investors tells you what the company believes is resonating with buyers.
That matters because weak controls are also one of the reasons agents stall before production.
The bottleneck is getting agents into production
Gartner predicted in May that by 2027, 40 percent of enterprises will demote or decommission autonomous AI agents because governance gaps are discovered only after production incidents (Gartner, May 26, 2026). That is a more specific warning than the usual concern that agent projects may simply fail: the problem can emerge after companies have already decided the technology works.
Camunda's 2026 survey found that 71 percent of organizations say they use AI agents, while only 11 percent of agentic use cases reached production in the last year, and 73 percent report a gap between their agentic AI vision and current reality (Camunda, January 14, 2026). Camunda has an obvious commercial interest in orchestration, so I would not treat the survey as neutral. Even with that caveat, the gap between experimentation and production is striking.
Together, the numbers describe a market that is experimenting much faster than it is putting agents into production.
Model capability is only part of the production decision. Security and risk teams want limits. Budget owners want cost controls. Operators want to know what happens after a bad action. The recurring questions are simple: What can the agent do? Who approved that authority? What gets recorded afterward?
BRYN byCivicLabor Day offer ⬩ through September 17
Save Your Labor (Day)
Bryn watches your site, scores the account, runs the Play, and files the run. A free month of it, on any tier.
Timesheet ⬩ arbor.devPunched ⬩ Tue 2:02 PM
2:02:08 PMWatched a return to pricing, then the comparison page
2:02:09 PMScored the account 86
2:02:10 PMRan the pricing.follow-up Play into Slack and the CRM
In ordinary usage, deterministic means the same input produces the same output. A generative model's judgment should not be treated that way. You can constrain the model, but you should still expect some variation in how it interprets a situation.
The useful promise behind Workday's "deterministic rails" is narrower: the model's judgment may vary while the controls around what it is allowed to do remain fixed.
Three controls matter:
Boundary. Which systems and accounts the agent may touch, which actions it may take, and how much it is allowed to do.
Approval. Who approved the workflow, when they approved it, and what authority that approval covered.
Audit log. What the agent saw, what decision it made, and what action it took, recorded so another person can reconstruct the run afterward.
Buyers do not need every model output to be identical. They need clear limits on what the agent may do, recorded approval for that authority, and a usable record of what happened. Companies already manage people this way: a salesperson has a territory, discount limits, and a CRM record. The person still exercises judgment inside those rules.
What can you control when an AI agent acts?
Open each step to see whether it depends on model judgment, a fixed rule, or both.
Model judgment
The model weighs the signals, so similar situations can receive different scores.
Fixed rule
The workflow checks the account, channel, and volume against rules set in advance.
Fixed before the run
Approval is recorded before the workflow is allowed to execute.
Mixed
Model judgment can affect whether the trigger is met; the permitted action and channel come from the approved workflow.
Fixed rule
Every run records the same required fields, even though the values inside those fields change.
Steps and terms as used in this article. The phrase "deterministic rails" is Aneel Bhusri's, from Workday's fiscal 2027 second quarter results announcement, August 27, 2026.
Ask the vendor to show you a blocked action
Most vendors can show you an impressive happy-path demo. A more useful production test is to ask the system not to do something.
Show me an account the agent was not allowed to touch. Show me the rule that blocked the action. Then show me the record explaining what happened.
If those controls are built into the product, a vendor should be able to demonstrate them quickly. If the answer is mainly a policy document, a roadmap item, or a settings screen, the control may still be a promise rather than an enforced mechanism.
Then ask what happens after a bad action. Who is alerted? What gets logged? Can the operator reconstruct what happened from one place? A team that cannot answer those questions will have a hard time trusting the agent with consequential work.
How Bryn handles these controls
Bryn is Civic's GTM agent for founders and growth teams. It watches signals from your product, site, and CRM, scores them against the customer profile you define, and runs an approved workflow when the conditions you set are met. Bryn calls that workflow a Play.
Boundary. The workflow defines which channels and accounts Bryn can touch, the volume allowed, and any actions that are off-limits. Those limits live with the workflow itself.
Approval. A person approves the workflow before it runs. Bryn can then execute that approved workflow when the trigger occurs without asking for another approval each time. That preserves speed while concentrating attention on the important decision: what authority you grant up front.
Audit log. Bryn records the signals it saw, how it scored them, and the action it took. The log can be read by the teams responsible for the workflow and exported when needed.
There is an important difference between Workday and Bryn. Workday can enforce controls inside systems it owns end to end. Bryn coordinates actions across third-party tools. We can control what Bryn is allowed to initiate, but we do not control what an external system does after it receives that action. What we can provide is a clear record of what Bryn initiated and why. Buyers should understand that distinction.
Three questions to ask before putting an agent into production
If you are evaluating an agent, ask three questions.
What is the agent not allowed to do? Who approved its authority, and where is that recorded? Can you show me an action the system blocked and the log explaining why?
If those answers are visible in the product, you have evidence that the controls are real. If they live mainly in documents or roadmap slides, there is still work to do before production.
The useful promise behind "deterministic rails" is not a deterministic model. It is an agent whose authority is bounded, whose approvals are explicit, and whose actions can be reconstructed. That is a practical standard companies can build for and buyers can evaluate.
If you are evaluating an agent and want to compare how these controls work in practice, I'm glad to talk.
Chris Hart is the CEO at Civic; he brings together decades of experience across technology, finance, and identity to help businesses navigate the shift to agentic AI. His Silicon Valley career spans more than 25 years, from running infrastructure at early internet and fintech startups to leading finance and operations teams at high-growth technology companies.
Beyond Civic, Chris has championed veteran leadership as Vice Chair of the Pat Tillman Foundation since 2006. When he isn't thinking about the future of identity and AI, you'll probably find him surfing or hanging out with his Dalmatian.
