Cisco Is Deploying AI Agents to 90,000 Employees. It's Published One Performance Metric.

By the end of July, every one of Cisco's 90,000 employees will have a personalized AI agent.

Not access to an AI tool. Not a shared assistant. A personalized agent — one built on infrastructure the company largely runs on-premises, routed dynamically across frontier models depending on the task at hand. CFO Mark Patterson described the approach to Fortune: "We feel like that's the most efficient way is to build our own AI stacks, which will go out and query the different models based on the particular use case."

This is one of the largest simultaneous enterprise AI deployments in corporate history. And the architecture behind it is more interesting than the headline number suggests.

The Smart Part

Most enterprises that deploy AI agents make one model decision, stick with it, and spend the next twelve months discovering where it breaks. Cisco built something different. Their system is a lightweight benchmarking engine at the point of invocation — routing each query to the model most suited to that class of work, factoring in cost, capability, and context. It's the kind of infrastructure you build when you've accepted a premise most organizations are still resisting: no single model is best at everything, and the cost of getting that wrong compounds across 90,000 people.

Running this stack on-premises isn't just a cost decision. It's a data strategy. Cisco controls what the agents see, how queries are logged, and where performance signals go. For a company that files public disclosures and handles sensitive customer data across a $15.8 billion quarterly revenue base, that's not overcaution — it's architecture that makes the system auditable in principle, even if the audits are internal.

The finance team is the clearest example of what this looks like when it works. AI now produces 80 to 90 percent of the first draft of the MD&A section — the mandatory management narrative in Cisco's quarterly public filings. That's a specific, measurable, consequential use case. Drafting financial disclosures requires accuracy, structure, and fidelity to regulatory expectations. Getting 80-90% of the first draft there isn't a demo. It's a verified outcome that the finance team can point to.

The 89,999 Other Use Cases

Here's the problem.

That MD&A number is the only specific performance metric Cisco has published for any of this. One function, one task class, one outcome figure — for a deployment that covers sales, engineering, operations, legal, support, HR, and every other function across 90,000 people.

Cisco's model routing architecture produces performance signals by design. Every query that routes to one model over another is a data point about which models succeed on which task types. That signal accumulates across the workload in real time. Somewhere inside Cisco's stack, there is a growing body of evidence about how this agent performs across different work — which functions it's materially helping, which ones are seeing completion rates fall short, and where the routing decisions are consistently wrong.

None of that data is visible. Not to the 90,000 employees being asked to trust it. Not to the investors trying to assess whether the restructuring it's enabling is justified. Not to the industry trying to learn from the largest rollout of this kind.

The Timing Problem

The same week Cisco announced the agent rollout, the company filed WARN notices for 471 California terminations — part of a broader cut of roughly 4,000 jobs the company has described as AI-driven restructuring. Those terminations begin July 13. The agent reaches every remaining employee by end of July. Software engineering roles are among those eliminated.

The optics are difficult. But the problem isn't optics — it's information.

If an organization is simultaneously deploying AI agents to every employee and reducing headcount explicitly because of AI, there are two possible interpretations. One: the agents are performing well enough that the same work can be done with fewer people. Two: the cuts are happening at a different layer — capability gaps, role redundancies, structural changes — and the agent deployment is augmentation of the remaining workforce, not displacement of the departing one. Both can be true in different proportions.

Employees being asked to use and trust these agents have no way to know which interpretation is correct. The MD&A metric doesn't answer the question. And without task-level performance data by function — the kind of data Cisco's own routing infrastructure is generating — workers are being asked to trust a narrative rather than a number.

That's a trust architecture problem, not a technical one.

What Scale Creates

A 1,000-person pilot can survive on narrative. Ninety thousand people cannot. At that scale, "the CFO says it's working" is not a trust mechanism. The questions you didn't have to answer in the pilot — how is the agent performing for engineering vs. finance, what's the failure rate on ambiguous queries, which task types are still producing wrong answers — are now questions 90,000 people are encountering daily, without published benchmarks to give the answers context.

Cisco's model routing architecture is genuinely smart. They're building the kind of embedded performance intelligence that most enterprises don't get around to until after their first major agent failure. But that intelligence is only valuable if the signals it generates make it back to the people who need to act on them — the employees whose work depends on the agent being reliable, and the decision-makers calibrating how much of the organization's function to hand over.

The MD&A number is what trust looks like when an organization publishes what its agents actually accomplish. The absence of equivalent numbers for every other function is what the trust vacuum looks like when it doesn't.

One metric isn't a measurement program. It's the first data point of one. For the largest enterprise agent deployment in history, it needs to be followed by the other 89,999.

---