A new CNBC report published today pulls back the curtain on something enterprise technology teams have been experiencing quietly for months: AI agents that perform impressively in demos and controlled environments fail to hold up under the demands of real enterprise operations.

The reporting focuses on the gap between the promise of AI agent platforms and the reality of deploying them inside complex organizations. The core problems are not exotic. They are the same problems that have derailed enterprise software deployments for decades, now showing up in a new context.

What Is Actually Breaking

Technical teams from major technology companies told CNBC that creating and operating AI agents is far harder than the marketing suggests. The specific friction points include memory management, agent coordination, and handling interdependencies between systems.

One executive quoted in the piece put it plainly: while AI agents work well for personal or isolated use cases, they “definitely cannot reach the enterprise level” without solving critical problems in memory, agent management, and inter-agent communication. Another warned against giving “all of your tokens and all of your money” to an AI bot that will waste resources without producing results.

The phrase one executive used to describe the current state of enterprise AI agent systems was “chaotic.” That is an honest assessment. Multi-agent systems touch many parts of an organization simultaneously. The interdependencies are what make them powerful in theory and brittle in practice. When one agent fails, or misunderstands context, or loses track of state, the failure can cascade.

The Token Problem Is Real

Token waste is a cost and reliability issue that is easy to underestimate until you are paying the bills. AI agents that operate inefficiently, making redundant calls, holding unnecessary context, or retrying operations that should not have been attempted, generate real costs at scale.

For businesses evaluating AI agents, this is worth understanding before signing contracts. The cost of running an AI agent system is not the cost of the software license. It is the compute cost of every token processed, every API call made, every retrieval operation executed. Poorly designed agent systems can consume orders of magnitude more compute than necessary to accomplish the same task.

This is not a theoretical concern. Companies running AI agents in production are already discovering this.

Why Demos Always Look Better

The demo-to-production gap in AI agents is wider than in most enterprise software categories for a specific reason: demos are designed around the cases the system handles well. The curated prompt, the clean data input, the well-defined task with a clear success state. Those conditions almost never exist in actual enterprise environments.

Real enterprise environments have inconsistent data, ambiguous instructions, legacy systems that do not return clean outputs, and edge cases that no demo scenario anticipated. An AI agent that navigates a clean demo workflow smoothly can fail or behave unexpectedly the moment it encounters anything outside that designed path.

The teams at major technology companies who spoke to CNBC are not struggling because AI agents are fundamentally broken. They are struggling because deploying software that handles real-world complexity inside real organizations has always been hard, and AI agents inherit all of that complexity while adding new challenges specific to how language models behave.

The Right Way to Think About This

None of this means AI agents are not valuable. It means the difference between a working AI agent deployment and a failed one is the same thing it has always been in enterprise software: thoughtful design, realistic scoping, and implementation expertise.

The companies getting value from AI agents in 2026 share a few common traits. They start with narrowly defined use cases rather than broad mandates. They invest in understanding their data and systems before automating against them. They build in human oversight at the points where errors would be costly. They test against real-world edge cases, not just the happy path.

The companies struggling are usually the ones who believed the demo was a reasonable approximation of what production would look like, and who underestimated how much integration and tuning work was required.

What This Means for Business

If you are evaluating AI agents for your organization, the CNBC report is a useful reality check. Here is how to use it:

Treat demos as existence proofs, not performance guarantees. A demo shows you what the system can do under ideal conditions. Ask vendors to demonstrate the system failing and recovering. Ask about token costs at scale. Ask about memory and state management in multi-step workflows.

Define success before you start. Vague mandates like “automate our operations with AI agents” create vague outcomes. Specific mandates like “reduce the time to process a customer support escalation from 4 hours to 30 minutes” create measurable benchmarks and focused design decisions.

Work with people who have done this before. The implementation gap between AI agent theory and AI agent practice is exactly where experienced operators add value. The technical pieces are available to everyone. The judgment about what to build, how to scope it, and how to avoid the failure modes takes experience with real deployments.

Enterprise DNA’s Omni Ops practice exists precisely because this gap is real. We have built AI agent workflows for businesses across industries and have direct experience with what breaks in production and how to design around it. If you are planning an AI agent deployment and want to avoid the scenarios described in this CNBC report, that is the conversation to have before you start, not after you have already spent money on a system that does not hold up.

The hype around AI agents is not wrong about the potential. It is wrong about the difficulty. Those are two different things, and treating them as the same is what gets organizations into trouble.