Cisco Gave Every Employee an AI Agent: The Cost-Routing Blueprint and the Trust Test Behind It

In late July 2026, Cisco began rolling out a personal AI agent called MyAgent to roughly 90,000 employees — one of the largest corporate AI-assistant deployments to date, according to Fortune. Every employee gets an agent that remembers their preferences, pulls context from their role and recent activity, and can run supervised workflows across Outlook, Webex, Jira, and SharePoint. That part of the story is easy to summarize and easy to admire.

The harder, more useful story is underneath the headline: how Cisco built an architecture that keeps 90,000 concurrent AI agents from bankrupting its cloud budget, how it wired governance and human approval into the system from day one rather than bolting it on afterward, and why the rollout landed inside a controversy the company partly created for itself. For any enterprise leader watching agentic AI move from pilot to company-wide deployment, MyAgent is less a case study in what an AI agent can do and more a blueprint — and a cautionary tale — for how to actually ship one at scale.

What MyAgent Actually Does

MyAgent runs on Circuit, Cisco’s internal, multi-model-agnostic AI platform that the company has been building since 2023, when its first internal assistant ran on Azure OpenAI, according to Dawn Liphardt’s breakdown of the rollout. Rather than a single chatbot, MyAgent is a front end to more than 800 backend subagents, each handling a narrower task — drafting a status update, summarizing a Jira backlog, reconciling a spreadsheet — that get orchestrated behind a single conversational interface, as detailed by Lovex’s analysis of the deployment.

The interaction model itself is a step beyond typical copilot tools: an employee states an objective and the desired outcome, and MyAgent determines the sequence of steps needed to get there, coordinating across connected systems rather than waiting for a prompt at each stage, per PYMNTS’ reporting on the deployment. That shift — from “answer my question” to “handle my objective” — is exactly the trajectory Gartner has been tracking across the enterprise software market. Gartner predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025, as vendors and internal platform teams alike move from AI assistants that need hand-holding to agents that act with more autonomy, according to Gartner’s own press release. Cisco just did it at the scale of an entire workforce instead of a single application.

The Real Innovation Isn’t the Agent — It’s the Router

The part of MyAgent that deserves more attention than it’s getting is the routing layer, because it answers the question every CFO eventually asks about agentic AI: what does this cost at scale?

Cisco’s orchestration layer sends 50%-60% of requests to open-weight models Cisco hosts on its own GPU infrastructure, another 20%-30% to conventional software automation that doesn’t touch a model at all, and only a small residual share out to frontier models like Claude, Gemini, or Azure OpenAI, based on the task’s required reliability and latency, per Dawn Liphardt’s reporting. In other words, the majority of the 90,000-person deployment’s daily workload never touches an expensive frontier model at all — it’s absorbed by cheaper in-house models or by automation that was never AI in the first place.

This is the practical answer to a problem we’ve written about before: as the era of flat per-token pricing gives way to a market judging AI spend by useful intelligence per dollar, the winners won’t be the companies with access to the best model — they’ll be the companies with the best triage for deciding which requests need it. Cisco’s tiered router is a working example: reserve the frontier model for genuinely hard reasoning, and let cheaper models and deterministic automation absorb everything else. It’s the same discipline we described in our look at the AI coding agent price war and what “cost per task” means for engineering budgets — except Cisco applied it not to one product category, but to every knowledge-work task across an entire company.

Governance Built In, Not Bolted On

The second thing Cisco got right architecturally is that MyAgent’s autonomy is bounded by design. External actions — anything that changes state outside the employee’s own sandbox — require explicit human approval, and a policy server can block destructive operations or prevent Cisco’s internal data from being used to train third-party models, according to UC Today’s coverage of the rollout. Every one of those 800+ subagents inherits access only through permissioned enterprise connectors tied to the employee’s own identity, rather than a shared service account with broad reach.

That distinction — an agent that proposes and a human who approves — is the same principle we’ve argued belongs at the center of any agentic deployment in our piece on keeping human control over AI with an explicit approve step. It’s also a direct answer to the governance gap we flagged in our analysis of why enterprise AI governance has become the real bottleneck as agent fleets grow: most companies deploying dozens or hundreds of agents don’t have a clean answer for who owns each one, what it’s allowed to touch, or how its actions get audited. Cisco’s answer — map every agent, and every one of its 800 subagents, back to an accountable human owner with tool-level authorization — is exactly the kind of identity discipline we described as missing at scale in our earlier piece on the non-human identity crisis facing enterprises whose AI agents already outnumber their employees. Cisco built that mapping in from the start rather than retrofitting it after agent sprawl had already happened.

The Timing Problem Cisco Created for Itself

None of that architectural discipline fully insulates Cisco from the second half of this story. On May 13, 2026, CEO Chuck Robbins announced a roughly 5% workforce reduction — nearly 4,000 jobs — in the same quarter Cisco reported record revenue of $15.8 billion and $5.3 billion in AI infrastructure orders, with CFO Mark Patterson describing the cuts as a reallocation of resources toward AI and security rather than a cost-cutting measure, according to Tech Times’ coverage of the layoffs. California-based layoffs began taking effect on July 13 — and MyAgent began rolling out to the remaining workforce roughly two weeks later, according to UC Today’s timeline of the two events.

That sequencing is the trust problem Cisco built for itself. When employees can draw a straight line from “we cut staff to invest in AI” to “here is your AI agent,” some will reasonably read the rollout as an invitation to help automate their own function out of existence. The research on how layoff survivors actually respond backs up the concern: in a Leadership IQ survey of more than 4,000 employees who kept their jobs after a company layoff, 74% reported that their own productivity had declined since the cuts, and 77% said they were seeing more errors and mistakes across their teams, according to Leadership IQ’s published findings. Handing that same workforce a powerful new automation tool two weeks later doesn’t reset the anxiety — it can compound it, whatever the tool’s actual technical merits.

What This Means for Your Organization

None of this means the technical architecture is wrong — it means the technology and the rollout communication are two separate problems that have to be solved together, and Cisco solved only one of them well.

If you’re planning an agent-at-scale deployment of your own, the transferable lessons are specific, not general:

• Build the router before you build the agent. Cisco’s real cost discipline came from deciding, per request, whether a frontier model was even necessary — not from picking one model and hoping usage stayed low.
• Map every agent and subagent to an accountable human owner before deployment, not after. Retrofitting identity and access controls onto an agent fleet that’s already sprawling is dramatically harder than designing it in from the first agent.
• Keep humans in the approval loop for anything external-facing. Internal drafting and summarization can run more autonomously; anything that changes a system of record or reaches outside the company should still require a person to say yes.
• Separate your AI investment story from your headcount story, in both substance and timing. Even a technically sound, well-governed rollout can land as a threat if it arrives on the heels of a layoff explicitly attributed to AI.

The build-vs-buy calculus is shifting fast enough that this pattern will keep repeating. McKinsey’s State of AI 2026 survey found that 32% of organizations have already skipped buying an off-the-shelf software product because they could build the equivalent internally with agentic tools, and large enterprises scaling agents in at least one function jumped from 27% to 40% year over year, according to McKinsey’s survey findings as reported by Yahoo Finance. More companies are about to make the same bet Cisco made. The ones that study its router and its governance model — and skip its timing mistake — will get more of the upside with less of the backlash.

Frequently Asked Questions

What is Cisco’s MyAgent?

MyAgent is a personal AI agent Cisco rolled out to its roughly 90,000-person workforce starting in late July 2026. It runs on Cisco’s internal Circuit platform, remembers each employee’s preferences and context, and can run supervised, multi-step workflows across tools like Outlook, Webex, Jira, and SharePoint rather than just answering one-off questions.

How does Cisco keep the cost of running AI agents for 90,000 employees under control?

Cisco built a routing layer that sends the majority of requests — 50% to 60% — to cheaper open-weight models it hosts on its own infrastructure, another 20% to 30% to conventional software automation with no model involved at all, and reserves frontier models like Claude or Gemini for only a small share of requests that genuinely need that level of reasoning.

Does MyAgent act autonomously, or does a human stay involved?

Both. MyAgent can chain together multi-step tasks on its own within an employee’s own systems, but any external action — anything that changes state outside that sandbox — requires explicit human approval, enforced through a policy server that can also block destructive operations.

Why did Cisco’s AI agent rollout draw criticism?

Cisco cut nearly 4,000 jobs in May 2026, describing the move as a reallocation toward AI investment, and then began rolling out MyAgent to its remaining workforce about two weeks after the layoffs took effect. That sequencing has been criticized because it invites employees to read the AI rollout as connected to the job cuts, regardless of the tool’s actual design.

Should other enterprises copy Cisco’s approach?

The architecture is worth copying: tiered model routing for cost control, and identity-and-approval governance built in from the start rather than added later. The rollout timing and communication around it are the part to do differently — pair any large-scale agent deployment with a clear, separate narrative about what it means (and doesn’t mean) for jobs.

What’s the difference between MyAgent and a typical company chatbot or copilot?

A typical copilot answers questions or drafts content when prompted. MyAgent is designed to take an objective and desired outcome from an employee and independently determine and execute the steps needed to reach it, coordinating across more than 800 backend subagents and multiple connected enterprise systems.

Sources

• Cisco is rolling out AI agents to every single one of its 90,000 employees - Fortune - Reports on the scale and start of the MyAgent rollout.
• Cisco Deploys Custom AI Agent to Entire 90,000-Person Workforce - PYMNTS - Details on the objective-based interaction model and cross-system coordination.
• How Cisco Empowered 90,000 Employees with an AI Assistant - Dawn Liphardt - Breakdown of the Circuit platform, subagent architecture, and cost-routing percentages.
• Cisco gave 90,000 workers an AI agent. Who coordinates? - Lovex - Analysis of the 800+ subagent orchestration model.
• Cisco’s 90,000-Employee AI Agent Rollout Could Become Enterprise AI’s Biggest Trust Test - UC Today - Governance details and the timeline connecting the layoffs to the AI rollout.
• Cisco Eliminates 4,000 Jobs on the Same Day It Reports $15.8B Record Revenue, Blaming AI Pivot - Tech Times - Details of the May 2026 layoffs and Cisco’s stated rationale.
• Don’t Expect Layoff Survivors to be Grateful - Leadership IQ - Survey data on productivity and error rates among layoff survivors.
• Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 - Gartner - Industry-wide forecast for task-specific AI agent adoption.
• The Build-vs-Buy Shift: 32% of Enterprises Bet on Agentic Coding Tools - Yahoo Finance - McKinsey’s State of AI 2026 survey findings on build-vs-buy and agent scaling.