AI Agent Guardrails for CIOs: How Cisco, Intuit, Workday and AT&T Govern Autonomous Agents
As reports of AI agents misbehaving in lab tests grow, CIOs are building guardrails around agents already in production. What Cisco, Intuit, Workday and AT&T are doing, what Deloitte's survey shows about adoption, and a practical guardrail checklist.
Most enterprises did not wait for AI agent governance to mature before deploying agents. Assistants with access to email, documents and business systems, and agent features built into platforms such as Salesforce, are already taking actions on employees’ behalf. Meanwhile, AI labs have published test cases where agents found ways around monitoring or exploited system weaknesses. In September 2026, Fortune reported on how technology leaders are responding. This article summarises those approaches, the adoption picture behind them, and a practical checklist for putting guardrails around agents that are already running.
How widely agents are used
According to CIO.com, a Deloitte survey of more than 3,200 executives found that 75% of organisations use at least some agentic AI, and 95% expect to within two years. Usage is mostly light today: 20% say they use agents moderately, 2% extensively and 1% as a fully integrated part of operations. Within two years, 46% expect moderate use, 23% extensive use and 5% full integration.
CIO.com points to Anthropic’s release of the Model Context Protocol (MCP) in November 2024 as a turning point, because it let models interact directly with data sources and tools. That is also why governance has become urgent: an agent that can act through tools carries more risk than a chatbot that only answers questions.
What leading companies are doing
Cisco: one platform, no third-party agents
Cisco built an internal agent platform called MyAgent, launched in August, that centralises all approved models, agents and enterprise data. Thimaya Subaiya, executive vice president of operations, told Fortune the company would “cannibalize and kill every other AI assistant within the company” to avoid “agent sprawl”, and that he will not authorise agents sold by third-party vendors. About 90,000 employees have access, Cisco reported 50% daily adoption within two weeks, and around 700 employee-built agents have been approved by a central team.
Intuit: security designed in from the start
Intuit CTO Alex Balazs said that when the company sketched the first architecture for its generative AI operating system, GenOS, it also drew “GenSRF” for security, risk and fraud. Every AI request is tracked and every response recorded. “You don’t want to try to retrofit the ability to enforce security and responsible AI foundations after the fact,” he said, adding that relying on models to “intrinsically do the right thing” expects too much of the labs.
Workday: a system of record for agents
Workday CTO Gabe Monroy described an “agent system of record” that manages the non-human identities of its digital workforce, used internally and sold to customers. He also emphasised training so that engineers and other users understand an agent’s risk profile.
AT&T: own the platform, stay model-agnostic
CIO.com describes AT&T’s approach. Chief data and AI officer Andy Markus said the company built a model-agnostic platform where it controls the knowledge used, grades accuracy and monitors governance, and that it has more than 1,000 agentic workflows built or in production, designed to be multi-purpose to keep the total number of agents down. He reported a five-times cash return on AI investment the previous year; that is a company-reported figure.
Accountability stays with people
PwC’s global chief AI officer Joe Atkinson put the principle bluntly: “‘The agent made me do it’ is not going to be a defense from a moral or legal perspective.” He argued that technology and security leaders must enforce guardrails and track agent tasks, but that business department heads also need to own the agents in their areas. Lumen’s Jim Fowler argued for “secure acceleration, not slowing down”, since attackers and competitors will not pause.
Platforms are connecting agents across vendors
Governance is getting harder as agent platforms connect. At Dreamforce 2026, Salesforce and Google Cloud announced that agents on either platform can reason and act on the same data without custom integration, using Salesforce’s MCP-based headless architecture. Salesforce will begin migrating select US customers to Hyperforce on Google Cloud in the fourth quarter of 2026, and Tableau analytics can be queried from Gemini Enterprise with row-level security enforced.
For CIOs, cross-platform agents mean that permissions and audit trails need to work across vendor boundaries. An agent in one system acting on data in another must be visible to both governance processes.
A guardrail checklist
Based on these approaches, a practical set of controls for agents already in production:
- Inventory every agent. Include agents built into SaaS platforms, employee-built agents and pilots. Record owner, purpose, data access and actions allowed.
- Give agents their own identities. Use separate, non-human identities with least-privilege permissions instead of borrowing employee credentials.
- Centralise approval. Like Cisco, require a central review before new agents go live, with a fast path for low-risk use cases.
- Log every request and action. Like Intuit, record inputs, outputs and tool calls so incidents can be investigated.
- Classify actions by consequence. Allow reversible, low-impact actions freely; require human approval for payments, deletions, external communications and decisions affecting people’s rights.
- Assign business owners. Each agent should have a named person in the relevant department responsible for its behaviour.
- Monitor for drift. Track error rates, overrides and unusual activity, and review agents after model or prompt changes.
- Train users. Make sure staff understand what agents can do and when to escalate.
- Plan for shutdown. Be able to disable an agent quickly without breaking dependent workflows.
We cover how to decide where agents should be allowed to act in a companion guide, and why many agentic AI projects are cancelled over governance gaps. For the services side, see our analysis of forward deployed engineers for Gemini Enterprise.
Our analysis
The common thread among the companies Fortune and CIO.com describe is control of the platform: a single place where agents are approved, given identities, logged and monitored. Companies that let agents arrive piecemeal through many vendors will find governance much harder. The good news is that most enterprises are still early, with only a small share using agents extensively. That leaves time to put inventory, identity, logging and approval in place before agent use scales, rather than trying to retrofit them after an incident.