Stuck Between Experimentation and Execution
Eighty percent of Fortune 500 companies have adopted agentic AI in some form. That number sounds like a success story. It isn’t – not yet. Most of those organizations are still trapped inside isolated pilots, running experiments that haven’t connected to the workflows, data systems, or strategic objectives that actually drive their businesses. The technology is deployed. The scale is not.
Agentic AI – systems where AI models take autonomous actions, make decisions, and interact with external tools and data – has moved fast enough to land inside nearly every major enterprise. Moving it from a proof-of-concept environment into something that changes how a company operates is a different problem entirely, and most organizations are still working through it.
The gap between adoption and meaningful deployment is where the real work begins.
Strategy First, Then the Agents
Arun Chandra, chief operating officer at NiCE, frames the core failure mode plainly: companies are chasing the technology before they’ve defined what they want it to do. “Everybody’s trying to figure out what can we do with this technology?” he said during a webcast produced in partnership with NiCE and MIT Technology Review’s custom content arm. That question isn’t wrong – but it’s incomplete. Without anchoring agent deployment to a defined business objective – whether that’s growing revenue, cutting costs, or a specific operational target – organizations end up with pilots that impress in demos and stall in production.
Chandra’s prescription is to work backward from the strategic goal, then redesign the workflows where agents will operate before deploying them. The order matters. Layering AI onto a broken or outdated process doesn’t fix the process – it accelerates its problems. “The last thing you want to do is to apply AI on an outdated or an inefficient workflow,” he said. That means workflow redesign has to happen before, or at minimum alongside, agent deployment – not after the system goes live and the gaps become expensive.
For agents to actually function at an enterprise level, they need more than a task definition. They need access to the data, the knowledge, and the context that let them make decisions worth acting on. They also need live connections to the back-end systems where those actions produce results. Chandra is direct about what happens without that infrastructure: “The efficacy of these AI agents is purely a function of the context, the knowledge, and the data they can ingest and use.” An agent with a narrow data window makes narrow decisions. An agent with no connection to the system it’s supposed to change does nothing at all.
Fragmentation at Scale Is Still Fragmentation
One of the less obvious risks of scaling agentic AI is that it can reproduce the same organizational fragmentation companies were trying to escape. When individual teams build their own agent systems without coordination, the enterprise ends up with a collection of disconnected tools rather than a functioning AI workforce. The integration problem that existed before the agents arrived simply shifts to a new layer. Governance, privacy, and security concerns – manageable when agents are confined to pilots – become harder to contain once those agents are taking consequential actions inside live business processes.
Chandra’s position is that agents should be evaluated against the same standards applied to human workers. That framing has practical implications. If a human employee needs access controls, accountability structures, and defined decision boundaries, an agent handling the same category of work should have the same. The workforce, in this model, becomes a combination of human and AI agents operating under shared expectations – not a human team with a set of automation tools bolted onto the side. Change management becomes part of the deployment problem, not an afterthought, because the people working alongside agents need to understand what the agents are doing and why.
Looking further ahead, a properly connected agent architecture could enable something more active: agents working proactively, coordinating with other agents, and resolving customer needs without waiting for a human to initiate each step. That capability doesn’t exist in isolated pilots. It requires the connected infrastructure and shared data context that most organizations are still building.
High-Value Use Cases Over Total Coverage
For companies moving from pilots to enterprise scale, Chandra’s advice lands against the instinct to solve everything at once. The phrase he uses is direct: don’t “boil the ocean.” Build a strategy organized around high-value use cases, the workflows tied to them, the workforce changes those workflows require, and measurable outcomes that tell you whether any of it is working. That’s a narrower mandate than most enterprise AI rollouts start with – which is precisely the point. Agentic AI systems that are connected, purposeful, and tied to real business objectives will outperform broader deployments that lack the infrastructure to support them.
The 80% Fortune 500 adoption figure will keep getting cited. What it doesn’t say is how many of those deployments have moved past the stage where someone in a meeting can show an impressive demo. Chandra’s framework – strategy, workflow redesign, data connectivity, governance parity with human workers – suggests that number is considerably smaller than adoption rates imply. The organizations that close that gap won’t do it by deploying more agents. They’ll do it by building the connective tissue those agents need to do anything worth measuring.
The question hanging over most enterprise AI programs right now isn’t whether agents can do the work. It’s whether the organization around them is built to let them.
