Dear CIO,

We have officially moved past the question of whether AI is real in the enterprise. It is real, it is in production, and in many cases, it is further along than most people assume. In the first session of our AI Lunch & Learn series, I focused on what I believe is the real shift underway, not just better models or faster code generation, but a shift in operational authority. We are moving from assistance to agents, and that transition changes everything. Many organizations are already beyond experimentation. 

Best Regards,
John, Your Enterprise AI Advisor

What’s Real? What's Risky?

I have worked with large enterprises, including manufacturing firms, major banks, and other global institutions, where AI is embedded into development workflows at scale. At the same time, we see conflicting headlines: AI is failing in the enterprise, yet AI is dramatically increasing coding velocity. Both can be true. What we are witnessing is the classic J-curve: early instability paired with increased throughput, without the system discipline to absorb the speed.

One of the most important points I raised is that vocabulary matters. We are using terms like assistant, agent, autonomous, copilot, and GenAI interchangeably, which creates confusion at exactly the moment we need clarity. Assistance is reactive and prompt-driven. It suggests a human decides. Agents are different. They are goal-directed, persistent, capable of tool use, and often possess write authority. Once something is goal-oriented, it will optimize toward that objective, sometimes in ways that conflict with policy or intent. This is an operational reality. This is why autonomy cannot be treated as binary. Rather than it being strictly on or off, it is on a gradient. From read-only informational systems, to advisory recommendations, to supervised execution, to policy-bound automation, and eventually to cross-domain autonomy, each level introduces increasing decision latitude and risk. Most enterprises today operate at the lower end of that spectrum, while open source ecosystems are moving much faster. That mismatch creates tension.

We are also seeing that throughput is not the same as flow. Recent DORA data shows coding velocity and deployment frequency increasing, but reliability and change success rates declining. AI amplifies what’s already there. In healthy, generative systems, it accelerates value. In bureaucratic or fragile systems, it magnifies dysfunction. As Goldratt taught us, improving anything other than the bottleneck is an illusion. If AI speeds up development without strengthening observability, governance, and runtime discipline, the downstream chaos becomes overwhelming. Security boundaries are also shifting. Agents perform multi-step reasoning, invoke tools, and operate across domains. Static defenses are no longer enough. We’re moving from “human in the loop” to “human on the loop,” which requires mature observability, policy clarity, kill switches, and continuous visibility.

Perhaps the most underappreciated gap is in observability itself. Today, we must think in three layers: infrastructure observability, application observability, and generative AI evaluations. The first two are familiar. The third is essential and often missing. When you deploy probabilistic systems, you must measure drift, hallucination rates, and behavioral consistency over time. Failures like the Air Canada chatbot incident weren’t model failures as much as evaluation failures. If you are moving beyond experimentation, evaluation tooling is no longer optional. Ultimately, the question is whether AI is amplifying your value or amplifying your dysfunction. Before increasing autonomy, organizations need to ask where added velocity might destabilize the system, whether bottlenecks are understood, whether policies are explicit and machine-interpretable, and whether autonomy is being treated as a gradient rather than a switch. 

This conversation is only beginning, and in our upcoming Lunch & Learn sessions on March 2 and March 18, we will go deeper into real-world agent failures, autonomy boundaries, platform thinking, and what mature enterprise AI governance truly looks like.

Save your spot here.

Monday, March 2nd:

https://event.webinarjam.com/462og/register/2w601amr

Wednesday March 18th:

https://event.webinarjam.com/462og/register/3p6gkbl2