Cisco's 90,000-Person AI Rollout: Why the Real Innovation Is the Router, Not the Model

Every one of Cisco’s 90,000 employees is getting a personal AI agent. That headline alone would be enough to fill a news cycle — it’s one of the largest enterprise AI rollouts ever attempted, and it lands squarely in the middle of a debate over whether “give everyone an agent” is visionary or reckless. But the part of Cisco’s rollout that actually matters for anyone planning an enterprise AI strategy isn’t the agent employees see. It’s the routing layer they never notice: the system deciding, task by task, whether a request needs an expensive frontier model at all.

That distinction is the difference between an AI initiative that scales and one that quietly bankrupts itself on token bills. Here’s what Cisco built, what its CFO says it actually costs, and what the architecture implies for every company still treating “which model do we use” as a one-time decision instead of an ongoing one.

What MyAgent Actually Is

Cisco calls the system MyAgent, and it began rolling out to the company’s roughly 90,000 employees starting at the end of July 2026, with the deployment still expanding into late August, according to Cisco’s own announcement. It’s built on top of Circuit, described by Cisco as its “secure, governed, multi-model agnostic AI platform,” and it plugs directly into the tools employees already use — Outlook, Webex, Jira, and SharePoint, per reporting from PYMNTS. Each employee’s agent carries persistent memory of their preferences, past interactions, and role context, so it can act on a request rather than just answering a question.

Cisco’s Thimaya Subaiya, Executive Vice President of Operations, framed the internal rollout as a product decision, not just an IT one: “Built by Cisco, for all Cisco employees, we can translate this into the blueprint that helps our customers and partners implement Enterprise AI at scale,” he said in Cisco’s announcement. That’s a notable move — Cisco is explicitly using its own workforce as the reference deployment for a platform it plans to sell.

The scale isn’t an accident, either. Circuit’s usage inside Cisco had already been climbing before MyAgent existed: agentic interactions on the platform grew nearly 350% quarter over quarter in the runup to the rollout, per Cisco’s own numbers. This isn’t a pilot layered on top of a skeptical organization — it’s a rollout riding momentum that was already there. That matters for the onboarding challenge most companies underestimate: getting one AI agent to stick with one team is hard enough. Cisco is trying to do it for an entire company at once, on infrastructure it’s had years to mature.

The Router Is the Product

Ask Cisco CFO Mark Patterson what makes MyAgent different from a chatbot wrapped around GPT, and he doesn’t start with the model. He starts with the routing. “It’s not going to burn a whole bunch of tokens with frontier models,” Patterson told Fortune in July. “It knows which tool is most effective and most efficient.”

That’s the core architectural choice underneath MyAgent: the system doesn’t send every request to the most powerful available model. Instead, it classifies each task and routes it — to a lightweight software automation for a deterministic lookup, to a smaller open-weight model Cisco hosts itself for moderate reasoning, or to an external frontier model only when the task genuinely demands it. “We feel like that’s the most efficient way is to build our own AI stacks,” Patterson said, and Cisco has leaned into that logic by running much of the infrastructure on-premises rather than renting it all from a hyperscaler.

The economics behind that decision are stark once you look at current pricing. Frontier model APIs run anywhere from roughly $2.50 to $15 per million tokens in and out, with top-tier reasoning models charging as much as $25 per million output tokens, according to current LLM pricing comparisons. Open-weight models — Llama, Mistral, DeepSeek-class systems — served through hosted APIs or self-hosted GPUs run $0.07 to $0.90 per million tokens, a 50-90% discount for a large share of enterprise task types. At the scale Cisco is operating — an agent for every one of 90,000 employees, each capable of triggering multi-step workflows — the difference between routing 100% of requests to a frontier model and routing only the genuinely hard ones is not a rounding error. It’s the difference between a program that scales and one that gets killed at the next budget review.

That’s also why the debate over per-token pricing misses half the picture. The unit price of frontier tokens matters less than how many of your organization’s requests actually need to touch a frontier model in the first place. Cisco’s bet is that most don’t.

Where the Payoff Shows Up

The clearest evidence of that bet paying off is inside Cisco’s own finance function. Patterson told Fortune that AI now produces “80%-90% of the first draft, at least” of the Management Discussion and Analysis (MD&A) section of Cisco’s public regulatory filings — a task that used to take analysts days and now takes minutes, with humans reviewing rather than starting from scratch. He’s also building what he calls a “CFO cockpit”: a single dashboard that synthesizes performance data across the business and surfaces recommended actions, rather than requiring analysts to assemble it manually.

None of that requires a frontier reasoning model running at full cost on every query — it requires a well-governed pipeline that knows when a first-draft financial summary is a routine pattern-matching task and when it’s a judgment call that needs deeper reasoning or a human. That’s the same distinction the orchestrator-agent model is built around: a coordinating layer that decides which specialized process handles a given task, rather than one generalist model trying to do everything itself.

Cisco’s broader financial position gives some sense of how seriously the company is betting on this infrastructure. The company raised its fiscal 2026 guidance for AI-related infrastructure orders to $9 billion, up from $2 billion in fiscal 2025, and its stock was up roughly 53% year-to-date as of late June 2026, trading near $117, per Fortune’s reporting. MyAgent isn’t a side project riding on hype — it’s downstream of a company-wide capital bet on AI infrastructure that Cisco is also trying to sell to its own customers.

The Industry Backdrop: Everyone’s Adding Agents, Few Are Controlling Them

Cisco’s rollout lands at a moment when enterprise AI agent adoption is accelerating faster than most companies’ ability to govern it. Gartner forecasts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025 — an eightfold jump in a single year, according to Gartner’s own press release.

That same research firm has also warned that the rush is outpacing discipline: Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, largely because of escalating costs, unclear business value, and inadequate risk controls. Anushree Verma, a Gartner senior director analyst, has attributed much of that gap to companies treating agent deployment as a proof-of-concept exercise rather than confronting the real cost and complexity of running agents at scale.

That’s precisely the gap Cisco’s architecture is designed to close before it opens. A company that hands out agents without a cost-aware routing layer is signing up for exactly the “escalating costs” scenario Gartner describes — and it compounds the problem this publication has covered before: organizations that lose track of how many agents they’re actually running tend to also lose track of what each one costs per task. Deploying 90,000 personal agents without a routing discipline wouldn’t be a governance failure waiting to happen — it would be a budget failure happening in real time.

What This Means If You’re Not Cisco

Most companies don’t have Cisco’s balance sheet, its on-premises GPU footprint, or years of internal platform work to draw on. But the underlying lesson doesn’t require that scale to apply:

• Treat model selection as infrastructure, not a config setting. A routing layer that classifies tasks by complexity and cost-sensitivity, and picks the cheapest model that can reliably do the job, is now a core piece of enterprise AI architecture — not an optimization you bolt on after launch.
• Reserve frontier models for the residual, not the default. Cisco’s approach treats expensive reasoning models as a last resort for genuinely hard tasks, not the default handler for every request. Most enterprise agent workloads — status checks, first-draft summaries, routine lookups — don’t need that level of reasoning.
• Measure cost per task before you scale headcount of agents. The moment you multiply agents across an entire workforce, unrouted token spend multiplies with it. Know that number before rollout, not after the first invoice.
• Use a real deployment, not a pilot, to prove the architecture. Cisco tested its routing and governance model on its own 90,000 employees before packaging it for customers. That internal proof point is worth more than any vendor benchmark.

The agent itself is table stakes at this point — Gartner’s numbers make that clear. The companies that come out ahead won’t be the ones with the flashiest agent interface. They’ll be the ones whose routing layer quietly keeps the majority of requests away from the most expensive model on the menu, the way Cisco’s does.

Frequently Asked Questions

What is Cisco’s MyAgent?

MyAgent is a personalized AI agent Cisco is rolling out to all 90,000 of its employees, built on Cisco’s internal Circuit platform. It integrates with tools like Outlook, Webex, Jira, and SharePoint, retains context about each employee’s role and preferences, and can carry out supervised multi-step workflows rather than just answering questions.

How does Cisco control the cost of running an AI agent for every employee?

Cisco routes each request to whichever model is most cost-efficient for that specific task, rather than defaulting to a frontier model. CFO Mark Patterson has said the system determines “which tool is most effective and most efficient” before processing a request, with much of the infrastructure run on Cisco’s own on-premises servers.

Is MyAgent replacing jobs at Cisco?

Cisco has not framed MyAgent as a replacement for roles; Patterson has described it as augmenting analysts and other staff — for example, having AI produce a first draft of financial reporting sections that humans then review and refine, rather than eliminating the review step itself.

Why does it matter whether a request goes to a frontier model or a smaller one?

Frontier model APIs can cost 10-100 times more per token than open-weight alternatives. At the scale of tens of thousands of employees each capable of triggering multi-step agent workflows, routing the majority of requests to cheaper models is often what separates a program that can scale from one that becomes financially unsustainable.

Is giving every employee an AI agent actually a good idea?

It depends heavily on governance. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs and unclear ROI. Cisco’s approach — a governed routing layer, on-premises infrastructure for cost control, and a phased rollout building on infrastructure it had already been maturing — is closer to the exception than the norm among companies rushing to deploy agents.

What should other companies take from Cisco’s approach?

The specific technology matters less than the discipline: build (or buy) a routing layer that matches task complexity to model cost before scaling agent access broadly, and measure cost per task early rather than discovering it in a budget review after rollout.

Sources

• MyAgent and the Rise of Ambient Intelligence: Cisco’s Next Step in Enterprise AI - Cisco’s official announcement of the MyAgent rollout and the Circuit platform it runs on.
• Cisco is rolling out AI agents to every single one of its 90,000 employees - Fortune’s interview with CFO Mark Patterson on cost strategy, MD&A automation, and Cisco’s AI infrastructure spending.
• Cisco Deploys Custom AI Agent to Entire 90,000-Person Workforce - PYMNTS coverage of MyAgent’s integrations and persistent-memory capabilities.
• Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 - Gartner’s official forecast on enterprise AI agent adoption.
• Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 - Gartner’s warning on agentic AI project cancellations and the drivers behind them.
• LLM API Pricing Comparison 2026: 30+ Models, Every Provider - Comparative pricing data for frontier versus open-weight model inference.