Ninety Thousand New Employees

Know what's running.

3 SIGNALS · 1 LESSON · 1 ACTION · ABOUT 5 MINUTES

SIGNAL 01

What happened. Starting with its new fiscal year at the end of July, Cisco began rolling out personalized AI agents to all 90,000 employees. Not a pilot, the whole company. CFO Mark Patterson discussed it in an interview with Fortune, calling AI "the most significant technology transition that we've seen in probably our lifetime." The design is deliberate: Cisco built its own AI stack that routes each request to the most efficient model for the task. "It's not going to burn a whole bunch of tokens with frontier models," Patterson said. The rollout ships with company-wide upskilling. The part that should stop you mid-scroll is already live in his own shop: AI now produces "80%-90% of the first draft, at least" of the Management's Discussion and Analysis section in Cisco's public filings. His team is also refining a "CFO cockpit," a planned AI dashboard intended to synthesize performance across products, geographies, and segments and recommend actions. All of this lands in the same year Cisco announced roughly 4,000 job cuts as it realigns toward AI.

Why it matters. Three leadership details are doing more work here than the headline.

First, look at who is discussing the rollout. Not only the technology or security leadership. The CFO. Patterson's prominent role signals that Cisco treats AI as a business and financial initiative, not only an IT project. The architecture reflects that discipline: model routing and an in-house stack instead of a pile of unmanaged subscriptions.

Second, the filing detail is the whole accountability model in miniature. AI drafts up to 90% of the MD&A, but the certification on that filing still carries a human signature, and everyone involved knows it. Drafting was delegated. Ownership wasn't. That's the pattern that scales: the work moves to the agent, the signature stays with a person.

Third, notice what the announcement doesn't discuss. The interview covers model routing, token costs, cockpits, and upskilling, and says nothing about how 90,000 agents get reviewed, monitored, or corrected. Maybe Cisco has that layer built. But the silence is the tell for everyone else: vendors and executives will hand you the capability story; the accountability story is the part leadership has to write itself. And with the rollout landing in the same window as job cuts, the trust question ("is my agent my assistant, or my replacement's training data?") is one more thing leaders own, whether they address it or not.

What to do. If agents are anywhere on your roadmap, borrow two things visible in Cisco's announcement: financial leadership involved in the initiative, and training that ships with the tools rather than after them. Then add the thing it didn't mention. For each workflow an agent enters, write one sentence: this output still requires a human signature, and here is whose. If you can't finish the sentence, the workflow isn't ready for the agent.

Sources: Fortune, Mark Patterson interview · Yahoo Finance

SIGNAL 02

What happened. Microsoft's David Weston took the Black Hat keynote stage with a talk titled "The End of Rare: Defending When Offense Is Cheap," and it comes down to one uncomfortable ledger. Security's entire architecture assumes that breaking a boundary (kernel, hypervisor, network) is rare and expensive; as Weston put it, "we said we're safe because this is rare and it costs a lot. Unfortunately, AI's changed this." The numbers he shared: Microsoft's Security Response Center is doubling the vulnerabilities it patches every six weeks, nine times the volume it processed in March. A single internal discovery tool has found, since April 1, 66% as many critical and important issues as every source combined found all of last year. And the sharpest figure of the conference: of roughly 200 Linux kernel vulnerabilities Microsoft's agentic system found in its internal Azure Linux distribution, an automated triage module converted 182 into crash-level proofs of concept. Weston said many were fully working exploits, including root exploits. The average cost of generating each proof of concept was $3.61 in tokens and 21 minutes. Earlier in the keynote, Weston described the system finding serious remote TCP/IP vulnerabilities of a kind he once spent a year crafting by hand. Those findings, he said, were being "spit out at industrial speed."

Why it matters. The $3.61 figure does not price a real-world intrusion. It measures one step: converting an already-found kernel vulnerability into a crash-level proof of concept in Microsoft's internal experiment. But it shows how sharply the cost of a traditionally expensive stage of exploit development is falling. Weston's conclusion was blunt: "you will lose the detection battle against AI agents, even with the help of AI." His prescription isn't faster patching or another detection product. It's changing the physics: memory-safe languages (Google wrote five million lines of Android in Rust, and memory-safety flaws fell from 76% of patched vulnerabilities in 2019 to under 20% in 2025, with none shipped in that Rust code), formal verification (mathematical checks that code satisfies defined safety properties, which caught a Microsoft encryption flaw that had passed tests, fuzzers, and human review), and machine-readable infrastructure configurations so agents can reason about them. Notice what kind of decisions those are. Not security-operations tool purchases, but multi-year engineering investments that trade this quarter's velocity for durable defensibility. That's a budget-and-strategy call, which makes it a leadership call. Weston's metaphor for AI in defense was "the Iron Man suit… there's still going to be someone at the center. You're just getting more superpowers that you can direct." For leaders, that suggests an accountability model too: automation can expand capability without removing the human owner.

What to do. The leadership translation of Weston's three steps. Ask your engineering leaders which of your systems still assume attacks are expensive; that's your exposure map. Ask which of your most critical code and infrastructure could move to safer languages, formal verification, or infrastructure-as-code with AI doing the heavy lifting; that's your investment list. Then fund one item on it this year. The new economics only favor defenders who spend their AI productivity on prevention instead of racing attackers at detection.

Sources: Black Hat USA 2026 keynote video, "The End of Rare" · SiliconANGLE keynote and theCUBE interview coverage

SIGNAL 03

What happened. The U.S. Chamber of Commerce's 2025 Empowering Small Business report found that 58% of surveyed small businesses use generative AI, up from 40% in 2024 and 23% in 2023. Adoption more than doubled in two years. That means more small companies are connecting AI tools to business email, documents, customer information, financial workflows, and internal knowledge. At the same time, Verizon's 2026 Data Breach Investigations Report includes a dedicated assessment of threats targeting small and midsize businesses. Its broader findings show that breaches continue to involve phishing, stolen credentials, software vulnerabilities, and ransomware, with generative AI now augmenting multiple attack techniques.

Why it matters. Small businesses were cyberattack targets before generative AI arrived. AI adoption gives attackers more to reach. Every new tool creates another account to protect, another place business data may travel, and another set of permissions to manage. Yet a twelve-person company rarely has a security team reviewing how those tools are configured or what information they can access. The risk is not that small businesses should stop using AI. It is that adoption is moving faster than the security practices around it.

What to do. Start with two controls. First, list every AI tool the business uses, who has access to it, and what company information it can reach. Remove tools and permissions nobody can account for. Second, require a person to verify any request to move money, change payment details, disclose sensitive information, or grant access through a different channel than the request arrived on. Then cover the basics Verizon recommends: multifactor authentication, prompt software updates, phishing training, and an incident response plan. AI adoption belongs in the security plan now, not after something goes wrong.

Sources: U.S. Chamber, Empowering Small Business report · Verizon Data Breach Investigations Report

THE LESSON

Put the three signals side by side. Cisco delegates most of the first draft of a closely scrutinized filing section to AI and keeps a human name on the certification. Weston shows one stage of exploit development falling to an average of $3.61, yet his picture of defense still has someone at the center of the Iron Man suit directing it. The leadership implication is ours: someone must also answer for what it does. Main Street is adding AI tools and data access faster than its security practices are changing, and the founder still answers for whatever goes wrong.

The pattern is the same at 90,000 employees and at twelve: you can delegate work to a system, but you can only delegate accountability to a person. Successful deployments tend to make that ownership explicit. Many preventable AI incidents begin where it is unclear.

Last issue argued that AI decisions are landing on the leadership table whether leaders notice or not. This week showed what the well-led version looks like: an executive owner, a budget discipline, training that ships with the tools, and a signature that never left a human hand. The organizations in good shape a year from now won't be the ones with the most AI or the least. They'll be the ones where every output that matters still has a name on it.

DO THIS WEEK

One action, whether you run a security program or a twelve-person firm.

List the outputs that leave your organization with AI's fingerprints on them: proposals, invoices, filings, customer answers, published content, code releases.

Next to each, write the name of the human who signs for it: the person who reviews it, owns it, and answers for it when it's wrong.

Every row with a name is delegation. Every row without one is abdication. Now you know exactly where to start.

THE ROUNDTABLE

If this issue read like an org chart with names missing, that's exactly the problem I want to explore.

On September 9, I'm hosting a small executive roundtable on Operational AI Leadership: how do you lead an organization when AI is increasingly making, recommending, or influencing decisions once reserved for people?

I'll start with a 25-minute walkthrough of the framework I've been developing, including where AI should decide and where people should remain involved, how accountability changes as AI moves into everyday operations, and what it takes to turn scattered AI experiments into a durable organizational capability.

Then we'll open it up.

The framework is still in the research phase, deliberately. I'm less interested in presenting a finished model than in pressure-testing it with executives, board members, and senior leaders dealing with these questions in real organizations.

Wednesday, September 9 at 10:00 a.m. Pacific.