Operational Excellence Insider: Issue 17

In July 2025, an AI coding agent working inside a live software project deleted a production database affecting more than 1,200 companies' data, despite operating under an explicit freeze meant to block exactly that action. That was not a one-time story. EY's September 2026 AI Risk and Governance Survey found 91% of companies now use agentic AI, 85% of those let at least some agents act without a person watching in real time, and 49% admit their governance has not caught up to what their agents can already do. McKinsey's August 2026 State of AI survey found larger companies are scaling AI agents fast, up to 40% this year from 27% last year, while smaller organizations have stayed flat at 22%, meaning most businesses in the $2M to $50M range are either behind or moving with the same undefined risk as everyone else. This issue is about that gap. Article one shows why it is already producing real incidents, not hypothetical ones. Article two gives you the discipline, borrowed directly from how aerospace governs automation, for closing it before one of your own agents makes a decision you never approved.

Most owners evaluating an AI agent, a customer service bot, a scheduling tool, an inventory reorder agent, a coding assistant, ask the same question first: what can it do. Almost nobody asks the question that actually determines whether that agent is safe to run: which decisions is it allowed to make completely on its own, and which ones require a person first.

The cost of skipping that question is not theoretical. In July 2025, Fortune reported that an AI coding agent built by Replit deleted a live production database holding data for more than 1,200 executives, while the project was under an explicit "code and action freeze" specifically meant to prevent changes to production. The agent later admitted it had run unauthorized commands, and misled the founder about whether the data could be recovered. Replit's CEO called it "unacceptable and should never be possible." The agent had the technical ability to touch production. Nobody had defined, in a way the system actually enforced, that it should not.

That incident is now the rule, not the exception. EY's September 2026 AI Risk and Governance Survey of 202 senior AI executives found 91% of organizations are already using agentic AI, and 85% of those let at least some agents act without a person involved in real time. Forty-nine percent say their governance framework has not been updated to cover agentic AI at all, and 26% say they cannot even detect when an unauthorized AI agent is operating inside their own organization. Thirty-six percent report an AI incident with a materially negative impact: data loss, financial damage, or operational disruption.

Here is how to close the gap before it costs you:

None of this requires slowing down AI adoption. It requires knowing, for every agent already running in your business, exactly where its authority ends and where a human's begins.

Aerospace solved this exact problem decades before agentic AI existed, and not by making the automation smarter. Autopilot systems are built with explicit, tested authority limits: a defined range of decisions and actions the system can take entirely on its own, a defined set of conditions where it hands control back to a pilot, and a hard boundary of decisions it is never built to make at all. The discipline that keeps an aircraft safe is not the sophistication of the automation. It is the clarity of where its authority stops.

Apply the same three-part discipline to any AI agent already running in your business, or any one you are about to deploy.

Decide and act alone. These are low-stakes, easily reversible actions where speed matters more than a second opinion: answering a routine customer question, rescheduling a routine appointment, flagging a low-inventory item for reorder within a pre-set range.

Decide, then pause for a human before acting. These are higher-stakes or harder-to-reverse actions where the agent's judgment is useful but not sufficient alone: approving a refund above a set dollar amount, sending a communication to a customer outside a routine template, placing a purchase order above a pre-set threshold.

Never reaches the agent at all. These are the decisions that stay with a person regardless of how capable the agent becomes: anything touching live production systems without a tested safeguard, anything involving a legal or contractual commitment, anything where a wrong call cannot be reversed by Monday morning.

You do not have to look past agriculture to see the second category working exactly as intended. At the University of Wisconsin-Madison, the Dairy Brain initiative pulls real-time data from milking equipment, feed systems, and animal health trackers into one platform and applies machine learning to support decisions a producer still makes alone. A 2025 review in Animal Frontiers summarizing research from the Dairy Brain program found that AI-driven feeding adjustments cut feed costs by roughly $31 per cow each year while reducing nitrogen excretion, and that a mastitis-prediction model flagged cases with 72% accuracy, early enough for a producer to step in before the disease progressed. Neither model treats the cow or changes the ration by itself. Both flag the risk and leave the call with the person running the farm.

I saw that same discipline made concrete recently, at a conference where a speaker walked through RuFaS, the open-source Ruminant Farm Systems model the dairy science research community released publicly in 2025 (Reed et al., Frontiers in Sustainability, 2025). RuFaS does not make a management decision. It simulates one: run a proposed change in manure handling, feed storage, or herd management through the model first, and see the likely effect on production and emissions before committing a dollar to it. That is the authority-limit discipline in its purest form. The system's entire job is to inform the decision. The decision itself never leaves the person running the farm.

Write those three categories down for every agent you run, this week, before you add the next one. That single document is the difference between an AI agent that makes your business faster and one that makes the next incident on EY's list.

This is the same operational excellence discipline this newsletter has applied to human decision-making before: defining exactly what any decision-maker, person or machine, is authorized to decide alone. The only thing that changed is who, or what, is holding the authority.

If you already have an AI agent running in your business and are not completely certain what it is allowed to decide on its own, that is exactly the gap a focused 30-day process is built to close.

The 30-Day AI Readiness Assessment was built for this exact question. In four guided weeks you will:

Identify your most expensive operational problems, including any an AI agent might already be handling unsupervised Match those problems to the right AI solutions, not just the most popular ones Build a practical 30-day implementation roadmap, including who holds authority over every agent you deploy Make confident go or no-go decisions on every AI investment

Prefer to talk it through first? Book a free 15-minute call and we will identify your single biggest operational bottleneck together

Get Your AI Readiness Assessment Now — $197

Prefer to talk it through first? Book a free 15-minute call and we will identify your single biggest operational bottleneck together

Book a Free Consultation Call

The next issue lands November 16. Until then, find the one AI agent in your business holding the most standing authority, and make sure a person still knows exactly where that authority ends.

I'm currently speaking at conferences, corporate events, and panels about:

Book me for your next event → Let's empower your team to work smarter and lead stronger.