The Short Answer
Cisco announced in July 2026 that its roughly 90,000 employees would each get a personal AI agent, and named it on 27 August: MyAgent, running on Circuit, which Cisco calls a "secure, governed, multi-model agnostic AI platform." Much of the infrastructure stays on-premises. The architecture is the durable lesson, not the agent. With ibl.ai you own all the code and the data.
Every enterprise can buy the same models Cisco uses. Almost none of them can buy what Cisco spent years building underneath.
What did Cisco actually announce about personal AI agents, and when?
Cisco's rollout was first reported on July 1, 2026, in a Fortune interview with CFO Mark Patterson, with the rollout beginning at the end of that month — the start of Cisco's new fiscal year.
That is worth stating plainly, because the item resurfaces periodically as something that happened this week. It did not. It is roughly two months old as of this writing, and the rollout has been underway since August.
Cisco has since published its own account. On 27 August 2026, EVP of Operations Thimaya Subaiya named the agent MyAgent and the platform underneath it Circuit, and confirmed MyAgent is "now rolling out to 90,000 employees."
Cisco describes Circuit as a "secure, governed, multi-model agnostic AI platform," providing access to approved language models, agents and enterprise data.
The Fortune interview is still the earlier source and the one that dates the program. Cisco's own post is what settles what the thing actually is.
Patterson's own example was benchmarking — comparing Cisco against competitors on revenue growth, EPS, R&D spend and capital allocation.
Does Cisco's MyAgent know your role, team and recent work?
Cisco says MyAgent carries memory of your work, and names the applications it acts in.
Cisco's own post attributes "MyAgent's power" to "its persistent memory that retains users' preferences, past interactions, and context over time."
It also describes "supervised autonomous workflows across applications such as Outlook, Webex, Jira, and SharePoint" — four named systems of record, not a chat window.
So the personalization claim is Cisco's own, in Cisco's words — and it is a claim about memory and application reach rather than about the model.
The honest limits are narrower. Cisco has not said the agent reads the HR system or an identity graph, and it has said nothing at all about whether participation is opt-in.
UC Today's coverage, written before Cisco's post, went only as far as saying the agent "adapts to the type of request it receives."
That gap is the interesting part for anyone copying this. "Remembers what you told it" and "wired into every system of record that knows what you do" are different engineering programs separated by years of integration work, and the second one is the expensive part.
Why is a company-wide agent rollout a context-layer problem, not a model problem?
Because an agent that is genuinely useful to a specific employee has to read the systems that describe that employee's work.
Role and team come from the identity provider and the HR system. Recent work comes from the ticketing system, the code host, the document store, the CRM and the calendar.
An agent with none of that is a chat window. An agent with all of it is wired into six systems of record, each with its own permission model.
Cisco's four named integrations — Outlook, Webex, Jira and SharePoint — are that list, partially populated, after years of internal work.
The model is the interchangeable part. Every competitor can license the same frontier models on the same terms, in the same quarter.
What no competitor can license is your identity graph, your access control, your provenance and your audit trail across those systems. That is the context layer, and it is the actual moat.
It is also why workforce-wide rollouts fail quietly rather than loudly. The agent works, and it has nothing specific to say.
What does per-seat AI pricing cost at 90,000 employees?
It costs more than most CIOs model, because per-seat pricing scales with headcount rather than with use.
Microsoft publishes Microsoft 365 Copilot at $30.00 per user/month, paid yearly. Anthropic publishes Claude Team at $20 per seat/month billed annually, with a premium seat at $100. OpenAI does not publish ChatGPT Enterprise list pricing at all.
Run those published numbers at workforce scale:
| Published plan | List price | Annual at 90,000 seats | Annual at 82,400 seats |
|---|---|---|---|
| Claude Team (standard seat, annual)† | $20 / seat / mo | $21,600,000 | $19,776,000 |
| Microsoft 365 Copilot (annual) | $30 / user / mo | $32,400,000 | $29,664,000 |
| Claude Team (premium seat, annual)† | $100 / seat / mo | $108,000,000 | $98,880,000 |
| ibl.ai (self-hosted, usage-based) | no per-seat pricing | tokens consumed | tokens consumed |
† One caveat on the Claude rows. Claude Team is sold "for teams of 2 to 150", so 90,000 seats is not a purchasable configuration; those two rows extrapolate the published list rate to show the shape of the curve, not a quote anyone could sign.
The 82,400 column is not hypothetical. Cisco's own Form 10-K for fiscal 2026 states: "As of July 25, 2026, we had approximately 82,400 employees".
The round 90,000 is the figure both the July reporting and Cisco's own August post use. Cisco announced a restructuring on May 13, 2026 affecting fewer than 4,000 positions, under 5% of its workforce.
That gap is the argument against per-seat in one line. A pricing model indexed to headcount charges you for a number that changes for reasons entirely unrelated to how much AI anyone used.
Per-seat SaaS is not one pricing option among several at this scale. It is the wrong shape: the bill is set before a single token is consumed, and it moves with HR decisions instead of with usage.
Renting enterprise AI costs far more than the invoice suggests for exactly this reason.
What is Cisco's architecture choice, and what does it tell other enterprises?
It tells them that a company with every reason to buy decided to build, and said so on the record.
Patterson's description is specific: "build our own AI stacks, which will go out and query the different models based on the particular use case." The system, in his words, "knows which tool is most effective and most efficient."
Much of the infrastructure is kept on-premises, which Cisco says gives it more control over cost and data security. Cisco's own description of Circuit compresses the same choice into one phrase: "multi-model agnostic."
That is not a hedge about vendor relationships. It is a statement that the model is a component the platform selects, rather than the platform itself.
The cost pressure behind that is real. As BusinessToday reported, agent-based tasks can require hundreds of thousands or even millions of tokens, against a few thousand for a standard chat.
At 82,400 people, a routing layer that sends routine work to a cheap model is not an optimization. It is the difference between a viable program and an unbounded one. Cisco told Fortune it does not disclose those costs separately.
Read the three choices together — own the stack, stay model-agnostic, keep the infrastructure inside your perimeter — and they describe an architecture, not a procurement.
None of it is painless. Cisco's chief customer experience officer, Liz Centoni, called the internal transformation "surgery without the drugs."
How does ibl.ai deploy agents across an entire workforce?
By shipping the architecture Cisco built for itself — a governed, model-agnostic platform with the agents running on top of it — to organizations that do not have Cisco's engineering budget.
With ibl.ai you own all the code and the data.
The platform runs on your own infrastructure with full source code access (perpetual license under codebase transfer). It is model-agnostic across any LLM, so a routing layer can send each task to the cheapest capable model and switch providers without a rewrite.
It is usage-based with no per-seat pricing, so the bill tracks consumption rather than headcount. And it can deploy anywhere: your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network.
The context layer is the part that is bought rather than built: agents read your identity provider, HR system, ticketing, code host and document stores in place, under role-scoped permissions enforced server-side, with every access audited.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
ibl.ai is family-owned and operated from New York, NY.
Related reading: the model is the commodity, the context layer is the moat — the same argument in a customer-service deployment; and renting enterprise AI costs far more than the invoice on how the per-seat curve behaves as adoption scales.
Sources: MyAgent, Circuit, the persistent-memory and integration descriptions, and the "now rolling out to 90,000 employees" confirmation from Cisco's own post by Thimaya Subaiya, 27 August 2026; the July announcement date, the ~90,000 figure, Patterson's quotes and the non-disclosure of costs from Fortune, July 1, 2026; the August start, token comparison and Centoni quote from BusinessToday; the "adapts to the type of request" description from UC Today; the 82,400 headcount from Cisco's FY2026 Form 10-K; the May 13, 2026 restructuring from Human Resources Online; list prices and the 2-to-150 Claude Team team-size limit from Microsoft and Anthropic.