Cisco Gives All 90,000 Employees an AI Agent This Month. The Design Choice Worth Copying.

July 27, 2026. Cisco starts its new fiscal year at the end of this month, and with it comes one of the largest AI agent deployments any company has attempted: a personalised AI agent for each of roughly 90,000 employees. CFO Mark Patterson described the rollout to Fortune, and the interesting part is not the headcount. It is the architecture, which is the same one we recommend to businesses running a thousandth of that scale.

What Cisco is deploying

1. Every employee gets a personalised assistant that handles tasks, answers questions and routes each request to the most efficient AI model for the job. Cisco does not separately disclose the cost outside its earnings reporting.
2. Model routing is the central design decision. In Patterson's words, the system is not going to burn a whole bunch of tokens with frontier models, because it dynamically selects the right model per task and knows which tool is most effective and most efficient.
3. Much of the infrastructure is built on premises, which Patterson framed as control over both cost and data. Cisco chose to build its own AI stack that queries different models depending on the use case, rather than routing everything to one vendor.
4. The finance function is already running on this. Patterson said 80% to 90% of the first draft of the MD&A, the mandatory narrative section of a public company filing, is now produced by AI, and the team is refining an investor relations tool that studies competitors' earnings calls to anticipate analyst questions.
5. The rollout is paired with company wide upskilling and internal knowledge sharing, with Patterson expecting internal competition as teams find new applications.

What it means for operators

You cannot copy Cisco's budget. You can copy every design choice in that list, and they are the choices that decide whether an AI programme pays for itself.

Start with routing, because it is the one most small businesses skip. Agents consume dramatically more tokens than chat does. A typical chat exchange might use a few thousand tokens, while a complex agent task that plans, calls tools and processes intermediate steps can consume hundreds of thousands to millions in a single run. Send all of that to a frontier model by default and your unit economics collapse quietly, three months after launch, in a bill nobody reads line by line. Route by task instead: a cheap fast model as the default, a stronger one only on escalation or on steps where quality actually fails a check.

The second copyable idea is scoping data rather than centralising it. Cisco's answer is on premises infrastructure. Yours is more likely to be a scoped connector, a per client workspace, or a rule that certain records never leave your CRM. Same principle, different price tag: decide where the data lives before you decide which model touches it.

The third is that Cisco paired the tools with training. An agent given to an untrained team produces a handful of enthusiasts and a lot of unopened tabs. The measurable wins in the Cisco example are all narrow, repeated, high volume tasks with a human reviewing the output, which is exactly where a first automation should start.

Worth noting honestly: Cisco confirmed roughly 4,000 job cuts globally in May, and the agent rollout lands in the same quarter. Employees notice that sequence. If you are deploying agents into a small team, say plainly what the agent is for and what it is not, because trust is the thing that decides whether anyone actually uses it.

If you want the routing layer, the scoped data design and the measurement built properly the first time, that is the work we do through AI automation and our AI automation agency, and you can hire an AI engineer to run it inside your own stack.