Managing AI Agents in Enterprise: The New Standard

The Moment Pilots Became Standard Issue

Until recently, an AI agent at a large company meant a pilot: roughly 1,000 people at Uber testing an assistant while everyone else watched from the sidelines. Now that same assistant reaches more than 30,000 employees, Cisco is handing a personal agent to about 90,000 people, and KPMG is wiring agents and a governance layer into 276,000 professionals across 138 countries. Managing AI agents in enterprise has stopped being a lab exercise. It is an operating discipline with its own budget line.

The reason this matters this quarter, not next year, hides in what these companies bought alongside the agents. It was not more models. A handful of the biggest names in security and identity spent record sums on one specific problem, and what that problem turns out to be changes how even a 50-person company should start. The numbers and a full case are below.

The scale is no longer a rumor. According to Fortune's reporting, Cisco began rolling out agents to roughly 90,000 employees at the start of its fiscal year in late July 2026, with each person getting a personalized assistant that handles tasks, answers questions and routes requests to the most suitable AI model. In a post on August 27, Cisco's EVP of Operations Thimaya Subaiya named the product MyAgent, built on the company's internal Circuit platform, which Cisco describes as a secure, governed, multi-model agnostic AI platform. It performs supervised autonomous workflows across Outlook, Webex, Jira, SharePoint and other applications.

KPMG moved on the same front. On June 9, 2026, it announced with Microsoft that Microsoft 365 Copilot would reach its entire workforce of more than 276,000 professionals, and that Microsoft Agent 365 would be used to manage, monitor and secure agents across the firm and its client environments. Microsoft has also described its own internal IT rollout using a tiered model: personal agents for individual work, team agents for known low-risk workflows, and enterprise agents for organization-wide processes.

Analysts see the same curve. Industry forecasts point to a sharp rise in the share of enterprise applications embedding task-specific agents by the end of 2026 compared with 2025. When a technology moves that fast from a small fraction of applications to a large one in a single year, it is no longer an experiment. It is plumbing.

What "standard" actually looks like

Notice what these deployments have in common. None of them is "give everyone a chatbot." Each pairs wide access with a central layer that decides who may launch an agent, what it can touch, and how it gets switched off. That pairing, not the model underneath, is what separates a rollout from a liability.

Why Seat Counts Hide the Real Problem

A seat count is the number executives remember: 30,000 licenses, 90,000 employees. It is also the wrong number to manage by.

One trade analysis of the Uber rollout took a conservative assumption of 45 machine identities per seat and arrived at roughly 1.35 million identities for 30,000 employees. Applied to Cisco's 90,000 employees, the same ratio produces about 4 million. Treat these as an analyst's estimate, not an audited figure. But the direction is hard to argue with: every agent that logs into a system, calls an API or spawns a helper becomes something that holds credentials.

Agents pick up permissions while they work

A human employee gets access once, on day one, and a manager can roughly picture what they can reach. An agent accumulates permissions at runtime, and multi-agent workflows spawn short-lived identities that exist for seconds. The same analysis points out that the potential damage from one misconfigured agent grows as the agent works. A procurement agent that can read supplier contracts and also approve payments is a fraud scenario with a friendly name.

Who owns the agent when the pilot ends

The second gap is organizational. Pilots have sponsors; production fleets have nobody. When the champion changes teams, the agent keeps running with the credentials it had. This is the part of a rollout that no one signs for, and it is the part that auditors and regulators will ask about first. For a closer look at how credentials for agents can be issued and revoked safely, see our piece on secure access for AI agents.

A pilot is judged by what the agent can do. A fleet is judged by what it was never allowed to do.

Managing AI Agents in Enterprise Became a Security Race

Follow the money and the priorities become obvious. In late July 2026, Okta agreed to acquire Permiso Security in a deal reportedly valued at just under $200 million, and Cyera agreed to buy Oasis Security for $1 billion, both aimed at agent and non-human identity. Cyera announced the completion of the Oasis deal on September 3.

It was not an isolated week. Trackers of agent-security deals count dozens of acquisitions since 2024, many of them in 2026. SailPoint bought Entro Security, and Cisco bought Astrix. Large identity and security platforms have been buying the startups that secure AI agents and the non-human identities they run on.

Venture funding has followed: agentic AI funding in 2026 has already far exceeded the total for all of 2025.

What the buyers are really purchasing

Read the list of targets and a pattern appears: identity, observability, pre-launch testing, data access. Nobody is paying billions to make agents smarter. They are paying to answer four boring questions: who is this agent, what can it reach, what did it do, and can we stop it. The market has quietly decided that control is the scarce resource, and the winners will be companies that can prove it.

For a business owner, this has a practical reading. If the largest vendors are folding agent controls into the platforms you already pay for, you will likely buy governance as a feature of your identity, security or productivity suite rather than as a separate product. That makes vendor selection a governance decision from day one; our checklist for choosing an AI agent vendor covers the questions to ask.

Case: What Two Giants Built Around Their Agents

Numbers make the contrast clearer. Here is how four large rollouts compare, based on what the companies and the press have disclosed.

Company	Scale	Control approach	What you can borrow
KPMG	276,000+ professionals, 138 countries	Microsoft Agent 365 as a central layer to manage, monitor and secure agents; combined with its "Trusted AI" framework	Buy governance together with the agents, not after them
Cisco	About 90,000 employees	Personal MyAgent on the internal Circuit platform, model routing, supervised task execution	One entry point per employee, with a router deciding which model does the work
Microsoft (internal IT)	Large internal rollout to employees and vendors	Three tiers: personal, team and enterprise agents	Start with low-risk workflows, escalate only as controls mature
Uber	From a 1,000-person pilot to 30,000+ employees	Seat growth of roughly 30x; analysts estimate over a million machine identities	Count identities, not just licenses

Two details in the KPMG case are worth a CEO's attention. First, the firm already ran Copilot for about two years before this expansion, so the June announcement was a second step, not a first leap. Second, KPMG operates in regulated work like audit and tax, and the announcement frames agent governance as something clients will also buy from it. Control is not just defense; it has become a product.

Cisco's case adds the human dimension. Reporting says the company told staff in May 2026 that it would cut fewer than 4,000 jobs, under 5% of its workforce, while investing in AI. Commentators argued that handing agents to the remaining staff in the same period creates a trust problem that no technical control can fix. Cisco's CFO, Mark Patterson, said the company is pairing the rollout with company-wide upskilling and internal knowledge sharing. The lesson: an agent fleet needs a communication plan as much as an access policy.

The first hundred agents are a technology project. The first ten thousand are a management problem.

A Playbook for Mid-Size and Small Businesses

You do not have 90,000 employees. You also do not need a $1 billion security acquisition. But the same logic scales down, and the sooner you apply it, the cheaper it gets. This is a practical order of operations for managing AI agents in enterprise settings of any size.

Step 1: Take inventory before you add more

List every agent, bot and automation that touches company data: the sales chatbot, the invoice reader, the scheduling assistant, the coding tool someone installed last month. For each, write down an owner, what systems it connects to, and what it can change. Most owners discover agents they forgot approving.

Step 2: Tier by risk, as Microsoft did

Separate agents that only read and summarize from those that send, approve or pay. The first group can move quickly. The second needs a named human owner, spending limits and a log. If you are weighing building your own agents against buying them, our breakdown of what custom AI agent development really costs shows where the hidden spend sits, and monitoring is usually on that list.

Step 3: Give each agent its own credentials

Shared logins make it impossible to tell what an agent did or to shut it off without breaking a human's access. One agent, one identity, the narrowest permissions that let it work, and an expiry date.

Step 4: Demand a stop button and a trail

Before any agent touches procurement, compliance or approvals, you should be able to answer three questions in minutes: what did it do yesterday, why, and how do we pause it? Vendors who cannot answer should not run your critical processes. The reasons agent behavior can drift beyond a team's intent are laid out in our analysis of why even AI labs struggle to control agents.

Step 5: Measure in business outcomes

Cisco's approach, as described in coverage of the rollout, is to start with a real problem and measure results in business terms. Track approval cycle time, hours of manager attention spent on routine sign-offs, and cost per processed document. If a number does not move in 90 days, the agent is a toy.

Do this, and the daily experience of running the company changes. You stop waiting for the incident report and start opening a dashboard that shows what every agent did overnight. That is what control feels like: calm in the places where a single wrong approval used to cost a night's sleep, and confidence that adding the next hundred automations will not multiply the chaos.

It also changes how you are seen. A CEO who can show the board an inventory of agents, an owner for each, and a log of every consequential action stops being the executive who "is experimenting with AI." Investors and directors read that as a leader who turns a risky technology into predictable margin, and your own managers read it as someone who builds systems instead of improvising around them.

Frequently Asked Questions

What does managing AI agents in a company actually involve? It covers five things: knowing which agents exist, giving each a distinct identity and limited permissions, logging what they do, testing them before launch, and being able to pause or retire them quickly. Large vendors now sell this as a "control plane" over the agent fleet.

How many employees can AI agents realistically be rolled out to? Tens of thousands, and beyond. Cisco is deploying agents to about 90,000 employees and KPMG to more than 276,000 professionals, and Microsoft has described a large internal rollout of its own.

What is non-human identity, and why does it matter for agents? It is the set of credentials that software, including agents, uses to log into systems and call services. Agents multiply these identities quickly, which is why identity vendors such as Okta, SailPoint and Cyera have spent heavily to acquire agent-focused startups.

Does a small business need agent governance? Yes, in a lighter form. An inventory, one owner per agent, separate credentials and a log of actions cost little at 10 agents and a great deal to retrofit at 100.

Should we buy governance separately or get it from our existing vendors? Increasingly it will come bundled, since identity, security and productivity suites are acquiring agent-control startups. Ask your current vendors what they already offer before adding a new product.

The companies that rolled agents out to 90,000 and 276,000 people did not win by moving first. They won by building the brakes before they pushed the accelerator, and the security market's spending spree is the proof that everyone else has noticed. Your fleet will be smaller, but the rule holds at any size: an agent you cannot see, name and stop is not an asset, it is an unsigned liability.

Whoever controls the agents controls the company's speed. Decide now whether that is you or an unlogged script.