Earlier this month I argued that the model is a commodity and governance is the real moat. That most enterprise AI projects fail not because the model can’t reason, but because nobody mapped who approves what, under which policy, with what record.

Cisco just turned that argument into a production case at real scale. Starting at the end of July, when its new fiscal year begins, the company is giving a personalized AI agent to each of its roughly 90,000 employees, CFO Mark Patterson confirmed in an interview with Fortune. This isn’t a pilot inside one finance team. Coverage of the rollout describes it as one of the largest corporate AI-assistant deployments documented so far.

What Cisco actually built

The agent every employee gets does three things: handles tasks, answers questions, and routes each request to whichever model fits the job. Not the most impressive model available. The most efficient one.

Patterson put it plainly: “it’s not going to burn a whole bunch of tokens with frontier models.” The system, in his words, “knows which tool is most effective and most efficient.” A routine question doesn’t trigger the most expensive model in the catalog. Only tasks that genuinely need it do.

Much of that infrastructure runs on-premises, inside Cisco’s own data centers, not fully rented out to a third-party cloud. Patterson gives two reasons: cost control and control over sensitive financial data, which is exactly the kind of data a company this size doesn’t want leaving its perimeter without clear terms.

In finance specifically, AI already produces 80-90% of the first draft of the MD&A (Management Discussion & Analysis) section in Cisco’s SEC filings. Patterson was explicit: “80%-90% of the first draft, at least, is now done by AI.” His team also built a tool that cross-references Cisco’s financial history against competitors’ earnings calls to anticipate what analysts are likely to ask.

On top of that, Patterson’s team is refining an internal dashboard they call the “CFO cockpit”: it synthesizes performance by product, geography, and customer segment, and recommends actions. Patterson himself uses it to compare Cisco’s metrics (revenue growth, EPS, R&D spend) against competitors in minutes, work that used to take his team days.

Cisco didn’t just flip on access and walk away. The rollout comes paired with company-wide training and internal competitions between departments to surface new uses for the agent, with the stated goal that adoption doesn’t depend on each employee stumbling into a use case on their own. Almost none of the coverage of the announcement mentioned that detail, and it matters because most mass AI rollouts fail exactly there: they hand out the license and assume real usage will show up by itself.

Two decisions here are actual governance, not just AI adoption

Two things in this design matter more than the “90,000 agents” headline, and both are exactly the kind of governance decision I argued for in the model is a commodity, governance is the moat.

The first is cost-based routing. Most companies deploying AI at scale make the same mistake: they give every employee default access to the most expensive model available, for every task. Cisco did the opposite. It built a decision layer that evaluates the task before picking the model. That isn’t a minor budget optimization. It’s the “decide” layer of the governance stack working exactly as it should: what’s allowed to spend, on which task, under what criteria.

The second is more subtle and matters more. AI drafts 80-90% of the first pass of the MD&A. It doesn’t approve it. It doesn’t sign it. It doesn’t file it with the SEC. That document still runs through Cisco’s finance and legal teams before it becomes something with real regulatory consequences. That’s the exact line between using AI to speed up work and letting AI act autonomously on something that can trigger a shareholder lawsuit if it goes wrong. The cockpit follows the same rule: it synthesizes and recommends, it doesn’t execute. Cisco automated the draft and kept the decision exactly where it belongs, with a person who has a name and legal accountability attached.

That’s a control tower running in production. Not in a policy document nobody reads. In the actual daily workflow of 90,000 people.

What Cisco didn’t do here is a design decision too. It didn’t claim the agent would replace a financial analyst’s judgment. It didn’t wire the CFO cockpit into any system that executes an action without someone approving it first. The ambition in this rollout is in the coverage (90,000 people) and the efficiency (fewer tokens burned on simple tasks), not in giving AI standing authority to act alone on decisions that matter. That restraint, more than the scale, is the part other companies should copy first.

The trust test Cisco hasn’t passed yet

Now the part I can’t leave out, because skipping it would be selling you half a case as a finished one.

Cisco is rolling this system out the same quarter it’s cutting jobs. The company eliminated roughly 4,000 positions globally in May 2026, and California layoffs begin July 13, just weeks before the agent reaches the remaining 90,000 employees. Cisco isn’t a company in crisis: its fiscal Q3 revenue was $15.8 billion, up 12% year over year. It’s cutting headcount while growing, in a year when US tech companies announced more than 123,000 layoffs between January and May, with AI frequently cited as a driver.

That exact calendar overlap is what led UCToday to headline its coverage as the biggest trust test enterprise AI has faced, not a closed case. An employee who watches AI-linked layoffs at their own company and then, two weeks later, receives a personal AI agent has no obligation to read that as a tool that makes them better at their job. They can just as reasonably read it as the system learning to do their job.

And here’s the point worth stating honestly: that 90,000 people got an agent on day one of the fiscal year, and that the agent works well and earns real, sustained trust over time, are two separate claims. Only the first one is confirmed. The second one has no data yet, and the outlet that broke the trust framing says it better than any spin I could put on it: this is unresolved, and it’s being tested live.

What we do at IQ Source

This is the case I now point to when someone asks what a control tower looks like outside of theory. Cisco didn’t start with the model. It started by deciding where the infrastructure runs, how each task gets paid for, and what AI is allowed to decide on its own versus what always stays with a person who signs their name to it.

That’s the same order we follow in AI Maestro discovery: before we connect an agent to a process, we map where each approval already lives, how sensitive the data it touches is, and which decision can never sit with the AI even when the draft does. And we look at the other side of the Cisco case too, the one almost nobody audits: if your company already has AI adoption without formal governance, handing out official agents on top of that unresolved distrust doesn’t fix it. It formalizes it.

Cisco built the technical part well. The trust part is still being written, one quarter at a time.

Design your control tower before you scale AI agents