The enterprise AI conversation has a consensus problem. Boards want more AI. CEOs want faster agentic deployments. Vendors are selling autonomy as a competitive advantage. And somewhere in the middle of all that momentum, the harder question has started to disappear: how much autonomy actually creates value? The default assumption is that more autonomy automatically creates more value. In practice, the reality is more complicated.
After years of working with enterprise teams and speaking regularly with IT leaders navigating AI adoption at scale, I’ve come to believe that one of the most important skills in enterprise AI right now isn’tknowing how to deploy an agent. It’s knowing when not to.
The Tax Nobody Wants to Talk About
There is a productivity illusion running through enterprise AI right now. Executives see workflows deployed and tokens consumed and conclude that value is being created. Token consumption is not a business metric.
I recently spoke with the IT leadership team at an e-commerce company. They had been early and aggressive adopters of AI-assisted workflows. What they found surprised them. The processes that ran on deterministic logic — order routing, inventory syncing, payment reconciliation — were fast, reliable, and cheap. The moment they introduced agents into those same workflows to handle edge cases, costs rose, latency increased, and auditability became a headache. The AI wasn’t wrong, exactly. It was just unnecessary.
This is not an isolated story. I hear versions of it regularly, across retail, financial services, and manufacturing. The pattern is consistent. Organizations that default to agents end up paying a premium for outcomes that rule-based systems could have handled more reliably at a fraction of the cost.
Gartner predicts that 40% of agentic AI projects will be canceled due to escalating costs, unclear business value, and inadequate risk controls. That number should be higher.
What Governance Failure Looks Like
An HR team at a large enterprise deployed agents to assist with resume screening at scale. No cost controls. No governance model. No visibility into what the agents were doing with sensitive candidate data. By mid-year, the spend was unsustainable and the audit trail was essentially nonexistent.
This is what happens when autonomy is treated as a goal rather than a tool. A Gartner survey found that only 13% of IT leaders strongly agree they have the right governance structures for AI agents. That figure is alarming given how quickly agent deployment is accelerating ahead of the infrastructure needed to manage it responsibly.
The IT leaders who will have a good story to tell their boards in 18 months are the ones building governance infrastructure now, not the ones maximizing agent deployment and cleaning up the mess later.
Least Agency: A Framework for Deliberate Autonomy
The concept I keep coming back to is what I call “least agency.” The idea is simple: just because a process can be handled by an AI agent doesn’t mean it should be. Organizations should apply only as much autonomy as necessary to accomplish the task at hand, rather than defaulting to agent-driven workflows simply because the technology exists.
Think of it as a dial. On one end, you have fully deterministic processes: step A produces step B, always, predictably, at scale. These are mission-critical workflows, often on the critical path to revenue, where consistency and auditability are non-negotiable. On the other end, you have genuinely ambiguous tasks — interpreting unstructured data, handling unique and new customer situations, synthesizing information across multiple sources — where AI earns its place because no deterministic rule set could do the job.
Most enterprise workflows live somewhere in the middle. A process that runs headlessly and correctly 90% of the time doesn’t necessarily need an agent for all of it. Maybe agentic capabilities are only needed for that 10% of cases that genuinely require judgment. The principle of least agency means deliberately determining where autonomy adds value and limiting it everywhere else, rather than defaulting to full autonomy because it feels more modern.
Finding Agentic Equilibrium
The goal isn’t to resist AI. It’s to find what I think of as agentic equilibrium. This is the right balance between deterministic execution and autonomous decision-making for each specific workflow, continuously refined as you learn more about where AI adds value and where it introduces unnecessary cost or risk.
This is an ongoing process, not a one-time architecture decision. As agents run in production, patterns emerge. Repetitive agent decisions that always produce the same outcome are candidates for refactoring into deterministic logic, which are likely cheaper, faster, easier to test, and more reliable. Unpredictable situations that genuinely require reasoning stay with the agent. Over time, the balance may shift, and the system gets more efficient.
The organizations I see getting this right share a few traits. They define success in terms their CFO can verify before they deploy anything. Not “we shipped an agent,” but something like “we reduced order exception handling time by 40%” or “we cut support escalation rates by a third.” They instrument AI workflows from day one. And they treat autonomy as something that must be justified by task complexity, not assumed by default.
What IT Leaders Should Do Now
For IT teams trying to navigate this in practice, the framework is straightforward even if the execution isn’t.
Start by mapping your workflows against two axes: how often does this process require genuine judgment, and what is the cost of a wrong or slow outcome? High-judgment, low-stakes processes are good candidates for agentic approaches. Low-judgment, high-stakes processes — anything on the critical path to revenue, anything touching regulated data or anything where a 1% error rate is unacceptable — should remain deterministic until you have strong evidence that an agent can match that reliability bar.
Build governance infrastructure before you need it. Audit trails, cost controls, identity and access management for AI agents are not afterthoughts. They are the difference between a production-ready system and a liability.
Measure outcomes, not activity. An IBM CEO survey found that 79% of executives perceive AI productivity gains but only 29% can actually measure ROI with confidence. Token consumption, agent count, and workflow creation speed are not business metrics. Hold AI implementations to the same standards you hold everything else.
The enterprise AI era is not going to be won by the organizations that deployed the most agents the fastest. It will be won by the organizations that figured out, deliberately and systematically, where autonomy creates value and where it gets in the way.
Knowing when to apply least agency is the most important capability an IT leader can build right now. The organizations that internalize that idea early will outperform the ones still chasing deployment velocity long after the costs of that approach have become impossible to ignore.
