AI Agents in Production: What Holds Enterprises Back

The debate over whether to bring AI agents into engineering and cloud operations is over. What remains open is a harder question: how much autonomy to give an agent, who approves each action, and what happens when it gets something wrong.

That is the core finding of a survey published on August 6, 2026 by Caylent, an Amazon Web Services and Anthropic partner, conducted by Censuswide among 200 senior leaders at organizations with more than 1,000 employees in the United States and Canada. The numbers are striking: 100% of respondents report active engagement with agentic AI, and 59.5% of senior leaders say their organization already has agents running autonomously in production, not in sandboxes or pilots.

Adoption is settled, the work isn't

For the past two years, much of the enterprise conversation around agentic AI centered on whether it was worth the investment. That phase is over. According to Caylent's survey, none of the 200 leaders surveyed said they were outside of active exploration of agents in engineering or operations.

The focus has shifted to execution. Organizations are already piloting, evaluating, or deploying agents across specific tasks:

• Automated testing: 67.5% of organizations.
• Automated incident response: 60.5%.
• Agents writing and committing code autonomously: 43%.
• 23.5% of leaders report agents are already broadly deployed across engineering and operations workflows, beyond initial pilots.

That last figure is the one any CTO or CDO still in pilot mode should pay attention to: nearly one in four direct competitors has already moved past that stage.

The barrier isn't technology anymore, it's governance

The most important finding in the study isn't how many companies are adopting agents, it's under what conditions. 98% of surveyed leaders said there are specific conditions under which they would allow an AI agent to autonomously execute changes in production. Only 2% ruled out any scenario of full autonomy, which shows how close the market already is to operating with fully autonomous AI, provided certain conditions are met.

What are those conditions? The study identifies guardrails, meaning the limits, approvals, and controls that govern what an agent can and cannot do, as the factor that weighs most heavily on the decision: 83% of respondents place guardrails on equal or higher footing than model intelligence when it comes to accelerating adoption. Randall Hunt, CTO at Caylent, put it plainly in the original release: model accuracy isn't the hardest part of the problem anymore; it's how much latitude to give an agent, how every action gets approved, and what happens when it's wrong.

At the same time, 93.5% of respondents find autonomous execution acceptable in production environments, as long as those controls exist. The takeaway is clear: the willingness to adopt is there, the governance architecture mostly isn't, at least not yet in most organizations.

What this means for US and Latin American companies

One clarification on the data before drawing broader conclusions: Caylent's sample is US and Canada based, with organizations of 1,000+ employees. There is no LATAM specific data in this particular study.

That said, the pattern it describes (widespread active adoption, with governance as the real bottleneck) is consistent with what shows up in companies across the Americas that have already run generative AI pilots: the technology works fine in controlled tests, but scaling it to production without an audit framework, cost controls, and human approval in the loop is where most projects stall.

For a company in CPG, banking, logistics, or manufacturing that already has two or three generative AI pilots running, the question this survey raises isn't whether to add agents, it's what governance structure to put in place before extending them into processes that touch customers, inventory, or financial decisions.

From isolated pilots to agents with clear rules

The challenge isn't purely technical. Before scaling any agent to production, it requires defining:

• Which decisions the agent can make without human intervention, and which require approval.
• How every action the agent takes gets audited, so it's possible to reconstruct what happened if something goes wrong.
• How cost per token and per execution is controlled, so automation doesn't turn into an unpredictable expense.
• Who is operationally responsible for the agent once it's in production, not just during development.

None of these points gets solved by picking a better model. They get solved through architecture, process, and a team that understands both the business and the underlying infrastructure.

Turning this into a concrete roadmap

At Greencode Software, we work on this exact problem through our Agentic AI service, structured around two pillars: custom agent development and enterprise AI platform implementation. The governance and observability component (quality metrics, decision auditing, and cost control per token) is precisely the kind of structure that, according to Caylent's survey, now determines whether an organization can scale agents with confidence or stay stuck in pilot mode.

The natural entry point for a company that recognizes this problem but doesn't know where to start is an Agentic Readiness Assessment: a two to three week evaluation that identifies which processes are automatable, what data enables them, and what level of governance each case needs before moving to production. It's not a long term promise, it's a concrete first step with a clear deliverable at the end.

If your organization already has agents running in pilots but no formal structure for approval, auditing, and cost control, that's exactly the gap separating companies already in the 23.5% that has scaled from everyone else still debating where to start.