The conversation about AI agents in the enterprise has split into two camps, and both of them are wrong. One camp says agents are failing. The pilots looked good in the boardroom demo, went nowhere in production, and the whole thing is being oversold.
The other camp says transformation is imminent, that agents will automate knowledge work at scale, and organizations that aren’t moving aggressively are already falling behind. Neither camp can answer the question that actually matters: which specific workflows should I deploy an agent against, and why?
That gap is showing up in the data. A March 2026 survey of 650 enterprise technology leaders found 78% have at least one agent pilot running. Only 14% have successfully scaled to production. That is not a technology problem. The models are capable. The tooling has improved. The variable isn’t the technology. It’s workflow selection, and that’s a problem no one is talking about clearly enough.
The Agents that Work are Boring on Purpose
The deployments that hold up in production share a consistent profile: narrow scope, repetitive execution, fully documented process, no judgment required. Just a sequence to complete. Think about the kind of call that floods a healthcare operations center: a physician’s office calling to ask why a claims payment receipt didn’t arrive, even though the payment actually went through. The answer involves pulling structured data from a defined set of systems and returning it in a predictable format. No ambiguity. No reasoning required. That’s agent-ready.
What’s not agent-ready is any workflow that requires judgment calls an enterprise hasn’t actually mapped yet. The failure pattern I keep seeing is organizations pointing agents at complex, ambiguous processes because those are the ones they most want to automate. Those are also exactly the ones agents are least equipped to handle. The harder a workflow looks to automate manually, the harder it is for an agent to do reliably. The irony is that the processes most worth automating first are the ones that feel too simple to bother with. High volume, low variance, fully legible process. Start there.
Building Agents One at a Time is Not An AI Strategy
At Availity, we process billions of healthcare transactions. We’ve been featured in three AWS case studies and Amazon Q Developer now generates roughly a third of all code written by our engineers who use it. Hundreds of thousands of lines of AI-generated production code, running in a regulated environment. We know what it takes to get this out of the demo stage. What I’ve watched happen across the industry is enterprises treating agents like projects: one team, one use case, one framework, one deployment. Ship it, move on, build the next one. Eighteen months in, they have fifteen agents, twelve frameworks, zero shared governance, and an audit committee that has no idea what any of it is doing.
The insight I keep coming back to, reinforced by a recent conversation with Anton Kornienko of NLP Logix who runs production-grade agentic AI: agents are not applications. They are stateful, autonomous, consequential. They don’t sit still between requests. They plan, they act, and they produce side effects in systems that log things for regulators. Treating them like applications is how you accumulate risk faster than capability. The companies getting this right are not the ones with the best individual agents. They’re the ones that built the platform to run any agent safely. The agent is not the product. The platform is the product.
That platform has non-negotiables: policy enforced at runtime rather than reviewed in a committee; validation and control loops built into execution, not bolted on after; observability across decisions, actions, and outcomes; lifecycle management with the same rigor you’d apply to any production asset; and shared infrastructure so the tenth agent costs a fraction of the first. None of that ships with a demo. All of it is the difference between a pilot that impresses a boardroom and a deployment that survives regulatory contact.
The Risk that Doesn’t Fail Loudly
When AI-generated output accumulates faster than human review can catch up, you have a new category of operational risk. In a regulated environment like healthcare IT, that means audit exposure, compliance gaps, and errors embedded in systems that are hard to untangle after the fact.
The organizations most at risk are not the ones moving too fast. They’re the ones moving fast in the wrong direction. An agent deployed against the wrong workflow, without the infrastructure to observe and govern it, doesn’t fail loudly. It fails quietly, over time, in ways that are hard to detect until they become expensive to fix.
This is also why the “wait for governance” posture I see from some enterprise leadership teams doesn’t protect you. It just shifts who owns the problem. Somewhere in your organization, an engineer with a problem and a laptop is already running agents. The governance question isn’t whether agents are running. It’s whether you have any visibility into what they’re doing.
How The Standoff Breaks
The practical path forward is less dramatic than either camp suggests. Start with local agents, work automation within the security envelope your engineers already operate in. Agents refactoring code, writing tests, searching logs, drafting documentation. The same human owns every outcome. The risk profile is closer to a smarter IDE than to an autonomous service.
That move doesn’t require organizational approval, committee sign-off, or a new governance framework. It produces real usage patterns, real risk surface data, and credibility you spend later when you do need approval for the centralized platform. At Availity, we had engineers demo the agentic workflows they built internally at town halls, from managing tickets to running full workflows locally on their workstations. The kind of demos where the room leans forward. That’s how you earn the platform conversation: with working code instead of strategy slides.
The debate about whether agents are delivering or failing is not the right question for enterprise leaders to be spending time on. The right question is more specific: have you mapped the workflows in your organization that are narrow enough, repetitive enough, and well-documented enough to be agent-ready right now? If you haven’t done that work, you don’t yet have the information you need to deploy successfully or to evaluate whether the technology is working.
The gap between 78% running pilots and 14% reaching production is not a story about AI. It’s a story about workflow selection and platform readiness. Fix that, and the technology takes care of itself.
Michael Privat is Chief Data and Engineering Officer at Availity, the nation’s largest health information network. He has 25 years of experience in healthcare IT and writes about AI, engineering leadership, and organizational change on Substack.
