By Poonacha Kongetira, CEO and Co-founder, Classie
If 2026 had a headline in enterprise AI, it was rogue AI. Agents reached into data they had no reason to touch. Employees fed sensitive files into tools their companies never approved. AI bills arrived that nobody could fully explain. Most of these stories shared a root cause: companies put agents to work before they had a way to see what those agents were doing.
That gap is closing, partly because it has to. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of this year, up from less than 5% in 2025. Yet only 13% of organizations told Gartner they have the right agent governance in place. Here are three shifts I expect in 2027.
1. Security and governance move to the front
For the past two years, most companies treated AI as an experiment. Teams tried tools, ran pilots and planned to sort out security later. In 2027 that order flips.
Boards, auditors and customers are already asking questions most AI programs can’t answer. Which agents are running? Who started them? What data did they use, and what did they do with it? Companies will respond by writing their own rules first, and many will set internal guardrails stricter than any law requires.
Regulation will follow, and Europe is likely to lead. The EU AI Act is already in force, with more of its obligations phasing in, and it gives other governments a model to copy or react to.
A guardrail written into a policy document stops nothing on its own. The organizations that handle this well will enforce their rules while agents run and keep a record of each session that holds up when someone asks what happened. Supervision will become a standard part of putting AI into production, the same way access controls are for any other system.
2. Companies start measuring what AI returns
Companies are making large workforce decisions based on the savings they expect to achieve. Hiring plans are shrinking and roles are being redesigned around agents. Few of these organizations can show what they are actually getting back.
That will have to change, because managing a mix of people and agents takes real numbers. Leaders need to know what each agent costs, who uses it, what work it finished and whether that work was any good. Token spend alone is a weak measure. A team can spend heavily on a model that returns slow, mediocre answers, or spend very little on an agent that handles hundreds of routine tasks a day.
In 2027, expect finance and IT to treat AI spend the way they now treat cloud spend, attributed to the team, the user and the workload behind it. The record that answers a security question (what did this agent do?) also answers the business one (was it worth it?). Companies that can connect cost to results will make better staffing decisions. Everyone else will be guessing with very large numbers.
3. The question shifts to ‘How do we make models safer by design?’
Models keep breaking the rules. They ignore instructions, take shortcuts and find their way around the limits set for them. The usual response is to add another layer of guardrails on top. That helps, but each layer adds cost and delay, and running large models is already expensive. The industry is heading toward data centers measured in hundreds of gigawatts, which is hard to sustain.
In 2027 the question shifts from how to contain unsafe models to how to build models that behave better from the start. Expect more research and investment aimed at models that follow an organization’s rules because of how they were built and trained, and at smaller, more efficient models that cost less to run and are easier to check.
Safer models won’t remove the need for oversight. A well-trained employee still works within policies and still gets audited. But better-behaved models, watched closely, will mean fewer incidents and a lower bill for keeping them in line.
Where to start
All three predictions point the same direction. Enterprises are moving from asking what AI can do to asking what it did, what it cost and whether they can trust it next time.
Leaders getting ready for 2027 can begin with a few practical steps. Build an inventory of every agent in use, approved or not. Decide who owns the rules those agents follow and how those rules are enforced. Start tracking cost and output per agent now, while the numbers are still small enough to understand.
Organizations that can answer those questions will have the confidence to put more agents into production.
##
ABOUT THE AUTHOR

Poonacha Kongetira is co-founder and CEO of Classie, an enterprise AI supervision company.






