At World of Workato (WOW) 2026, the annual user conference for enterprise orchestration and AI, three Workato Embedded customers compared how they decide what an agent can touch.
Their finding: An AI agent doesn’t have to be malicious to cost you money. In some cases, it only has to offer a discount nobody approved.
In the session, Governing Agentic AI at Scale, moderator Justin Paul, Director, Solutions Consulting at Workato, sat down with:
- Steve Micklewright, Integration Services Principal, LogicGate: a GRC platform for third-party risk, regulatory compliance, and cyber risk.
- Robert Harris, Director of Product Operations, Nextiva: a customer experience and unified communications platform.
- Anand Narasimhan, CTO, S-Docs: document generation and e-signature for public sector and regulated enterprises.
Each company is at a different stage in its agentic AI journey, bringing a different perspective to four foundational questions.
Question 1: What can the agent see?
Answer: Ask what your agent inherits, not just what you grant it.
LogicGate already tied permissions to roles for human users. A role decides which applications, data, and workflows someone sees. When Steve’s team built agents, they were assigned roles from the same model, so governing them needed no new build. According to Steve, this process had worked for years for their end users, so why reinvent it now?
LogicGate’s internal agents read GitHub, Confluence, and Zendesk’s knowledge bases, but they aren’t connected to Salesforce or the HR system, keeping those systems out of reach.
At Nextiva, Robert described a narrower form of access. Its support agent sees only part of Salesforce and returns only information that belongs to the customer asking.
“I wonder how many people out there have implementations where they’re probably giving access to too much information and they don’t even notice.”
— Robert Harris, Nextiva
Anand pushed the question past the agent. If a connected system has weak role-based access (RBAC), the agent inherits that weakness, however well the orchestration layer is governed. An agent’s access is only as well governed as the systems it connects to.
Question 2: What can it decide?
Answer: Give agents clear limits, and bring people in when the stakes are high.
Robert’s examples from Nextiva’s customers show how those limits vary by situation. A shipping company might let callers track an order but require a person to start a return. For Nextiva’s emergency-vet customers, an urgent situation calls for the agent to recognize what’s happening and hand the call to a person.
LogicGate sets a decision boundary in vendor onboarding. An agent reviews questionnaires and vendor analytics, then recommends a decision. A person makes the final call. Steve said the human review works because the agent shows its reasoning:
“We don’t give them an AI summary per se. We show how the AI made the decision on why what they’re doing is a safe decision.”
— Steve Micklewright, LogicGate
For Anand, contracts make the stakes clear. S-Docs generates proposals, NDAs, and MSAs, and a stray clause can have serious consequences. An agent that quietly adds a 20% discount for a missed SLA, he said, would send any legal team into a meltdown. S-Docs aims to become a governed, secure agentic platform for enterprise documents that’s simple enough for nontechnical users.
The right boundary depends on the risk: an agent may handle routine requests, while urgent situations and consequential decisions call for a human handoff or review.
Question 3: How will you know if the agent is wrong?
Answer: Test against the cost of being wrong, then keep learning from real use.
The panelists’ answers differed because the cost of an error depends on the use case. Anand put evaluations, role-based controls at every level, and guardrails on his short list.
“If you’re doing agents without evals, you’re kind of shooting in the dark. Hope it works.”
— Anand Narasimhan, S-Docs
Steve started with the cost of an error. In GRC, he said, an answer can’t be a “maybe.” If an AI told an international bank it might be compliant with a regulation, the bank could face a $100 million fine the next day.
For LogicGate, governance was the easy part; training the AI took time. Once testing produced good-to-great answers, the team released the agent to as many users as it could, with guardrails in place and without giving it full access. Users prompt in different ways, so feedback started coming in the next day and shaped the next version.
Robert added something formal evaluations might miss: how the end customer feels. Nextiva’s customers put agents in front of their own customers, so he advises staying involved and watching sentiment instead of assuming people will love the experience.
Together, the examples point to a continuing process: evaluate before release, keep access bounded, and use real-world feedback to guide what improves next.
Question 4: How do you do it again, faster?
Answer: Reuse the architecture and controls that already work.
At LogicGate, employees talk to an internal agent in Slack. Workato connects Slack to AI APIs and memory stores, and recipes copy knowledge bases into those stores. Workato also queues jobs so many employees can use the agent at once, without LogicGate having to design that scaling layer itself.
Building the first agent for the integration services team took two to three months. The second shipped within 48 hours because the recipes were ready to reuse. The following week, the product management team changed one line in a recipe and had its own agent the next day.
“As long as [the agent] has the training, the architecture remains the same. It’s just rinse and repeat. Clone a recipe, point it elsewhere.”
— Steve Micklewright, LogicGate
The reusable parts – connections, recipes, and the underlying architecture – let each team focus on what its agent needed to know and do.
Anand extended the idea to governance. Teams can build a control plane themselves, he said, but a platform that already provides one gives them a foundation to build on.
The speed comes from making the repeatable parts reusable, from the agent’s architecture to the governance around it.
Build governance you can reuse with Workato
The four questions point to a practical way to scale agents: define what they can access, decide where people need to step in, test against the cost of mistakes, and reuse controls as new agents come online. That gives product teams a repeatable foundation as agents move into more customer workflows.
Workato ONE for SaaS Platforms gives product teams one governed platform for every workflow, agent, and integration they ship, with verified user access, human-in-the-loop approvals, and audit trails built in.
Explore Workato ONE for SaaS Platforms
