Artificial Intelligence
Lawful by Design: CIO Lessons From Rising
How to build AI agents that follow the rules, scale safely, and stay on budget.
Julie Colwell
Principal Strategist
Workday
"When was the last time 'winning' actually meant something to your business, and not just a news headline about 'the AI race'?"
That's how Workday Chief Technology Officer Gabe Monroy opened his keynote at Rising in Las Vegas this week. Most conversations about AI centre on who's ahead: the fastest models, the most agents deployed. "I get the excitement," Monroy said, "but for anyone who's running a business, it's just a distraction."
For CIOs, the priority is putting AI to work safely, at scale, inside their organisations today.
For CIOs, the priority is putting AI to work safely, at scale, inside their organisations today.
Headless execution is now table stakes. But for autonomous agents to add incremental value, not risk, CIOs must decide where the reasoning runs and where the experience lives. Get it wrong and a model hallucination becomes a security leak, or complex decisions stall because they are forced into a chat bubble.
Lawful AI was a primary focus of conversations at Workday Rising. Here’s what rose to the top.
Guide
The CIO Playbook: Top 3 Outcomes of Agentic IT
Agents Need Data Governed Your Way
The largest enterprises are building their own agents, meeting employees and, more recently, customers inside the tools they already use. Every Fortune 50 company Monroy talks with is building this. "It's happening right now, at scale, everywhere," he said.
Those agents need two things from core systems:
Data governed the way you already govern it. Zero-copy federation lets analytics and AI platforms query live data without stale copies or fragile extraction pipelines, while the source system's security model remains in control.
The ability to take real actions, scoped to identity. Execution is where access control matters most. Rather than handing an agent broad database keys, give it narrowly defined tools tied to the logged-in user, delivered over open protocols like the Model Context Protocol (MCP), with human confirmation for every write and an audit trail behind it.
Intermountain Health shows what the data side can deliver. The healthcare company once ran hundreds of custom reports to feed its analytics platform, and they sank more than 300 developer hours into pipelines that still delivered stale data. Moving to live, governed access saved more than 100 developer hours per report, and non-technical leaders can now ask operational questions about workforce, finance, and clinical data in plain English.
Intermountain Health saves more than 100 developer hours per report by replacing custom reporting pipelines with live, governed data access.
Without Permissions, AI Is a Capable Liability
An agent pursues its goal with whatever access it's given. This foundation decides whether it stays inside your rules or runs past them.
A lawless agent connects to core systems through unmanaged API keys. It can complete a task its user has no authority to perform, report success, and never throw an error. "The raw APIs don't care about your org hierarchy or security policies," Monroy explained.
"The raw APIs don't care about your org hierarchy or security policies."
—Gabe Monroy, Chief Technology Officer, Workday
Workday CEO Aneel Bhusri made the same point in his keynote, saying: "An agent that's not grounded in business rules might move fast, but it is essentially lawless."
Jay Wieczorkowski, general manager of Workday's developer platform, explained why. "An agent will do exactly the task it's asked,” he said. “The problem is when nobody tells it who's allowed to do what."
A lawful agent works through scoped tools tied to the logged-in user's identity, so it inherits the permissions, approval chains, and audit requirements already in place. An authorised change moves through approvals and into the audit trail. In a product demo at Rising, Yasmeen Ahmad, Workday senior vice president and general manager for Data Cloud, showed what that looks like in practice.
“We didn’t just understand the data; we analysed it, we decided, and we wrote the decision back to the system of record with every security protocol and every workflow still intact,” she said. “You can only do that if you own the entire decision loop.”
For CIOs, the choice comes down to accountability. When reasoning runs outside the system that enforces your rules, developers have to write, test, and maintain the logic that keeps things safe. The result, Monroy noted, is "massive engineering and security debt sitting on your books."
“We didn’t just understand the data; we analysed it, we decided, and we wrote the decision back to the system of record with every security protocol and every workflow still intact. You can only do that if you own the entire decision loop.”
—Yasmeen Ahmad, SVP and GM, Data Cloud, Workday
Build Outside, Govern Inside
Industry analyst Jon Reed and Workday’s Ben Pierce, general manager of Adaptive Planning, framed the problem from the buyer's side: why would a customer trust a frontier model?
"People are scared," Pierce said. "They want something they can trust."
Reed's caution was blunt: "Today's AI will behave like a sociopath if there isn't a human in the lead." Enterprises want a deterministic baseline, with context and harnesses that help models produce precise results from customer data that is never stored by the model, plus auditability that shows who did what and why. Reed also urged precision when evaluating vendor claims: today's models are better described as guided by a system of record than fully grounded in one.
The principle CIOs can take to any architecture review is "build outside, govern inside." Let teams build in the tools they prefer, and enforce security, lifecycle, compliance testing, and ownership where data and rules live. Every agent should be registered, scoped to the permissions it needs, red-teamed before production, assigned an owner, and measured for value.
"Today's AI will behave like a sociopath if there isn't a human in the lead."
—Jon Reed, Co-Founder, diginomica
When Anyone Can Build, The Stakes Are Higher
AI-assisted development lets people describe what they need in a single plain-language prompt and review the resulting plan before anything deploys. Not only does this keep developers in control, but it also means anyone who understands the business can become a builder.
While that expands capacity, it also raises the stakes for governance. If anyone who can describe an outcome can build an agent, controls have to apply no matter who builds it or where it lives and functions.
Keep AI Models and Costs Predictable
"Moving agents from development to production is hard," said Dean Arnold, vice president of AI platform at Workday. Agents have to scale to your workforce and meet your security team's requirements without consuming the entire token budget.
Pierce argued that cost discipline starts with the model. "You don't need frontier models for most tasks that enterprises are doing," he said. With AI pricing still unsettled across the industry, it falls on enterprise vendors to bring sanity to consumption and help customers match models to tasks.
The following practical approach can help track AI spend:
Plan: Estimate consumption by agent and usage driver before you commit, with the math visible to your finance partners.
Monitor: Track usage by service and by agent, including credits nearing expiration.
Control: Route alerts at set thresholds to the right owners, set per-agent budgets, and enforce hard limits.
Arnold summed it up as "predictability before you commit, visibility while you're running, and enforcement when you need it.” In other words, leaders should apply the same governance they’d expect for cloud spend or any other line item.
"You don't need frontier models for most tasks that enterprises are doing."
—Ben Pierce, GM, Adaptive Planning, Workday
Choose Your Path, Then Lead
Moving forward, CIOs can build their own agents, deploy a vendor's agents in their own tools, or adopt a native AI workspace. Monroy said each can work with "one set of governed rails underneath all of them."
This division of labour is a useful starting point: People for the what and why. Agents for the how and do. Here’s how to do it:
Map where reasoning runs for the agents and integrations you have today, and flag any operating on broad API keys.
Pick one workflow where employees already want help from an agent, such as internal mobility or self-service.
Right-size models to tasks before you scale.
Set budgets and alerts before agents go into production.
Governed rails solve the technical problem, but that’s only half the battle. The organisational challenge in putting AI to work safely, at scale, falls to leadership. Pierce was direct about it: "AI can't fix your silos, but you can."
Guide
The CIO Playbook: Top 3 Outcomes of Agentic IT
A gap between effort and impact is the real ROI problem. Discover the solution.