Workday's chief AI officer lets models pick the workflow while approval steps stay fixed
Workday's Joel Hellermark says its agents pair reasoning models with deterministic guardrails, itwire reported. Sana, bought for US$1.1 billion, now orchestrates every Workday agent, so the model-driven half of that split runs in one acquired product.
The Board Room · Leadership desk
For an IT ticket, approval steps are deterministic; reasoning picks which workflows to call and in what order. Workday bought Sana for US$1.1 billion; it orchestrates every Workday agent. An itwire question to Hellermark says Sana is going into other companies' chat tools.
- constraint IT ticket approval steps Deterministic: the steps and how they run are fixed, in Hellermark's example, claim 3
- decision Reasoning models Work out which workflows to call and in what order, per Hellermark, claim 3
- capability Sana The orchestration layer for every Workday agent and the new Workday Home, claim 7
- decision Workday Bought Sana for US$1.1 billion and finished acquiring it on 4 November 2025, claim 6
- exposure Other companies' chat tools Workday is putting Sana into them, per an itwire question to Hellermark, claim 13
| Who | How | Kind | Claim |
|---|---|---|---|
| IT ticket approval steps | Deterministic: the steps and how they run are fixed, in Hellermark's example | constraint | 3 |
| Reasoning models | Work out which workflows to call and in what order, per Hellermark | decision | 3 |
| Sana | The orchestration layer for every Workday agent and the new Workday Home | capability | 7 |
| Workday | Bought Sana for US$1.1 billion and finished acquiring it on 4 November 2025 | decision | 6 |
| Other companies' chat tools | Workday is putting Sana into them, per an itwire question to Hellermark | exposure | 13 |
What happened
- In Hellermark's IT ticket example, approval steps and how they run are deterministic, while choosing which workflows to call, and in what order, is left to probabilistic reasoning.
- He said the split avoids "rogue agents that even have a two percent probability of going off the rails. It's not going to work for our customers."
- CEO Aneel Bhusri made the same argument on stage the next morning, saying: "You can't be 90 percent accurate on the payroll."
- Jev, the TypeSafe AI model behind Sana's new interface, takes fields and allowed values defined up front and returns typed choices with probabilities attached.
Why it matters
- constraint Reasoning is confined to choosing and ordering workflows, so on Workday's account a model error changes what runs but not how an approval executes.
- exposure Workflow selection, the probabilistic half, runs in the layer Workday bought, so wrong-workflow errors would originate in one orchestrator for every agent.
- decision A buyer weighing Workday agents can ask, workflow by workflow, which steps are fixed rules and which the model selects; the IT ticket is the example Hellermark gave.
- precedent Once Sana takes requests from third-party chat tools, the guardrail claim has to hold at entry points Workday does not design.
Hellermark, Workday's chief AI officer, described the pairing as the company's AI thesis in a 30-minute media Q&A held the day before the Innovation Keynote [1]. "The core of our AI thesis is to combine reasoning models with deterministic guardrails, and that's really the heart of Workday," he said [2]. That is a position about Workday's own product. Payroll, the case CEO Aneel Bhusri cited, is a task where a near-miss counts as a failure [5]. We think the split follows that logic: rules where a near-miss is a failure, models where several orderings of steps could reach the same result.
The orchestrating layer is young. Workday finished buying Sana, which Hellermark co-founded and ran, on 4 November 2025 [6], and itwire says that less than a year later Sana is the orchestration layer for every Workday agent [7]. Jev, the model Sana's release names as powering its new interface, entered early access on 15 September [9]. Hellermark said Sana could already generate interface components before Jev existed, but "we couldn't make it instant" [16].
On Jev, itwire's reporter asked what it does inside Sana and whether it decides when Sana acts on its own [12]. Hellermark's reported answer concerned the interface. "With every keystroke, there's a decision model that selects the right UI components for that," he said [14]. TypeSafe pitches Jev as fast and cheap enough for routing and guardrail decisions [10]. Itwire notes that the speed and cost figures are TypeSafe's own and have not been independently reproduced [11]. Hellermark said he believes Workday is "the first company to put this into production in close partnership with TypeSafe" [15].
Sequencing matters here. Workday is making itself reachable headless, over MCP and APIs, and is building Sana into chat tools owned by other companies [13]. A buyer approving agents now would meet the rule layer's real test when that access arrives, because requests will then come from front ends Workday did not build. In our view the question is whether approval steps stay fixed when a request enters through a third-party chat tool.
What to watch
- Whether Workday says if Jev or any other probabilistic model sits at its guardrail decisions or only in interface selection.
- How approval steps are enforced when Sana is called through MCP, APIs or other companies' chat tools.
- Whether anyone independently reproduces TypeSafe's speed and cost figures for Jev.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption20
- Hype gap+15
- Incentives70
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Joel Hellermark, Workday's Chief AI Officer, took part in a 30-minute media Q&A at Workday Rising 2026, held the day before the Innovation Keynote.
- [2]
Hellermark said: "The core of our AI thesis is to combine reasoning models with deterministic guardrails, and that's really the heart of Workday."
- [3]
Using an IT ticket as the example, Hellermark said the approval steps, and how they run, are deterministic, while working out which workflows to call and in what order is where probabilistic reasoning comes in.
- [4]
Hellermark said the split is how Workday avoids building "rogue agents that even have a two percent probability of going off the rails. It's not going to work for our customers."
- [5]
Workday CEO Aneel Bhusri made the same argument on stage the next morning: "You can't be 90 percent accurate on the payroll."
- [6]
Hellermark co-founded and ran Stockholm-based Sana, which Workday bought for US$1.1 billion and finished acquiring on 4 November 2025.
- [7]
Less than a year after the acquisition closed, Sana is the new Workday Home and the orchestration layer for every Workday agent.
- [8]
Jev is a model from TypeSafe AI: users define a set of fields and allowed values up front, and Jev returns typed choices with probabilities attached.
- [9]
The Sana Home release says the new dynamic interface is powered by Jev, which launched in early access on 15 September.
- [10]
TypeSafe pitches Jev as fast and cheap enough for routing and guardrail decisions.
- [11]
TypeSafe's speed and cost figures for Jev are its own and have not been independently reproduced.
- [12]
itwire's reporter asked Hellermark what Jev is doing inside Sana and whether it decides when Sana acts on its own.
- [13]
Workday is opening headless access through MCP and APIs, and putting Sana into other companies' chat tools.
- [14]
Hellermark said of Sana's interface: "With every keystroke, there's a decision model that selects the right UI components for that."
- [15]
Hellermark believes Workday is "the first company to put this into production in close partnership with TypeSafe."
- [16]
Hellermark said Sana could already generate interface components before Jev, for example when asked to "update Salesforce on my behalf," but "we couldn't make it instant."
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- itwire.comWorkday's Joel Hellermark on Jev, "meta work" and why AI agents need deterministic guardrails
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.