Cisco started giving a personal AI agent to every single one of its roughly 90,000 employees at the end of July 2026. Not a pilot team. Not one department testing the waters. Everyone. CFO Mark Patterson confirmed the details directly in an interview with Fortune, and Cisco’s own blog laid out how the system actually works.
I’ve written before on this blog about the real difference between AI automation and AI augmentation, and about building an actual adoption framework instead of just handing out access and hoping for the best. Cisco’s rollout is the largest real-world test of exactly those principles I’ve seen documented anywhere. Here’s what they actually did, and what’s genuinely worth learning from it, even if your business has 10 employees, not 90,000.
What MyAgent Actually Does
MyAgent runs on Cisco’s internal platform called Circuit, described as secure, governed, and model-agnostic. Employees can access it from desktop or mobile, and it connects directly into tools they already use daily: Outlook, Webex, Jira, and SharePoint.
The real distinction Cisco makes is important. This isn’t a chatbot you ask questions and get answers from. It’s built for what they call supervised autonomous execution. An employee defines an objective, and the agent coordinates the actual steps needed across those connected systems, with a human still supervising the outcome. That’s a meaningfully different thing than a chat window.
The Detail Almost Nobody Reporting This Story Mentioned
Here’s what I think is the single most important part of this entire rollout, and it’s the part that barely made it into most coverage: MyAgent doesn’t default to the most powerful AI model for every task.
Cisco built cost-based routing directly into the system. A simple policy question gets handled by a lightweight model. A complex financial analysis goes to something more capable. Reporting on the internal architecture describes roughly half of all requests going to cheaper open-weight models, another chunk handled by plain software automation with no AI involved at all, and only a small remainder actually using an expensive frontier model.
Patterson put it directly: the system “knows which tool is most effective and most efficient,” and it’s “not going to burn a whole bunch of tokens with frontier models” for a task that doesn’t need one.
This is exactly backwards from how most businesses I see are approaching AI tools right now. The default instinct is to subscribe to the most powerful, most expensive AI tool available and use it for everything. Cisco is treating this as a real operations and cost problem from day one, not an afterthought to deal with once the bill arrives. If you’re running any AI-assisted workflow in your own business, this is worth copying directly: match the tool to the actual task, not the other way around.
They Didn’t Just Flip a Switch and Walk Away
The second detail worth real attention: Cisco paired this rollout with company-wide training and internal competitions between teams specifically to surface new ways of using the agent. This wasn’t optional extra credit. It was built into the rollout plan from the start, because most large-scale AI rollouts fail exactly at this point. Companies hand out access and assume people will figure out useful applications on their own. Most don’t, and the expensive new tool sits mostly unused.
This matches something I’ve said in my own AI adoption framework on this blog: pilot on one specific process, review the results, and only then expand. Cisco did this at a massive scale, but the underlying discipline is the same one a five-person business should apply before rolling AI into every workflow at once.
A Real, Concrete Example, Not a Hypothetical
Most AI adoption stories stay vague about actual results. Cisco’s finance function gave a specific, checkable one. AI now produces 80 to 90 percent of the first draft of the MD&A section in Cisco’s public regulatory filings, a mandatory narrative section that used to take analysts days to write. Humans still review and refine every draft before it’s final. Nobody’s claiming AI is filing SEC paperwork unsupervised.
This is the augmentation principle in action, at real enterprise stakes. AI handles the repetitive first pass on a well-defined, structured task. A human still makes the actual judgment call before anything goes out the door.
What This Actually Means If You’re Not Cisco
Match your AI tool to the actual task, not habit. You don’t need to build custom cost-routing infrastructure, but you can apply the same thinking. A quick internal question doesn’t need your most expensive AI subscription’s most powerful mode. Save that for the task that actually requires it.
Training and structured adoption matter more than access. If you’ve given your team an AI tool and usage is quietly low, that’s not necessarily a sign the tool doesn’t work. It’s often a sign nobody built real structure around using it. Cisco treated this as a first-class part of the rollout, not an afterthought.
Pick one well-defined, high-volume task to automate the first draft of, the way Cisco did with MD&A sections. A repetitive, structured piece of writing or analysis you do regularly is exactly the kind of task where AI genuinely saves real time, with a human still reviewing before anything ships.
Supervised execution, not full autonomy, is still the real standard even at this scale. Even Cisco, with genuine engineering resources most businesses don’t have, built this around human oversight rather than letting agents run fully unsupervised. That’s a strong signal about where the real, responsible line sits right now.
Bottom Line
Cisco’s 90,000-employee AI agent rollout is genuinely instructive, not because of the headline number, but because of what’s underneath it: matching AI tools to tasks by actual cost and need, building real training and adoption structure instead of just granting access, and keeping humans in the loop even at massive scale. Those are the same principles that apply whether you’re deploying AI across 90,000 employees or just trying to get real, sustained use out of an AI tool in a five-person team.