“If we hand emails and SOPs over to AI, we’ll be so much more efficient.” More and more small and mid-sized businesses are acting on that instinct โ and the instinct is right. AI deserves a place at the center of your operations. But before you flip the switch, there is one question worth asking yourself.
Can you stop it?
The Question That Comes Before “Is It Convenient?”
When evaluating automation, most people ask, “How much easier will this make things?” The more important question to ask first is: “If something goes wrong, can I undo it?”
Imagine you have AI set up to automatically send emails to your clients every morning. One day the content is slightly off. Nobody catches it โ and it has already gone out to a hundred companies. Those clients are reading it as an official communication from you. There is no undo button.
The more seamless a system is, the harder it is to stop once it is moving. That is precisely why you need to design for stoppability and reversibility before you design for convenience. This is the core principle of Reversible Management: sound judgment requires the ability to course-correct.
How One Person Ran a Shareholder Meeting Solo
When AI is used well, a small team can accomplish a remarkable amount. At one startup, a single person handled a shareholder meeting โ a process that would typically require three to five people and several months of preparation โ almost as a side project. Personalizing emails for each recipient, tracking document submissions, coordinating schedules, organizing materials, liaising with outside experts, working through the legal pieces. And by all accounts, the quality was the best it had ever been.
So far, this sounds like a simple “AI is amazing” story. What matters is what comes next.
The reason it worked was that the person running it had genuine expertise in the subject matter. They did not accept every sentence and document the AI produced at face value. There were moments when they flagged a problematic phrase, or recognized that a particular clause might raise legal issues, and stepped in to fix it. That is what made it stoppable.
If someone without that knowledge had handed off the same work entirely, they would not have caught the errors. The AI’s mistakes would have gone out the door. In the worst case, legally problematic documents could have been formally sent. That would not be an “AI error” โ it would be a human error made with AI.
Three Conditions for Stoppable Automation
So how do you design for this in practice? It is not complicated. Before you launch any automation, confirm that these three conditions are in place.
- There is a paper trail. Who created what, when, who approved it, and when it ran. AI-driven work needs to be traceable after the fact. This also gives you a record to stand on for management decisions.
- A human approves before execution. Whatever AI produces, a person signs off before it goes live. No automatic execution without human sign-off. For email, this can be as simple as adding one step to review an outbox summary before sending.
- You can switch back to manual at any time. Have the authority to pause the system, and a clear path to restore the previous state.
With these three in place, automation is not frightening. Without them, no matter how convenient a system looks, it is dangerous. Targeting 95% automation with 5% human oversight is more sustainable and safer in the long run than aiming for 100% automation. Full automation may sound like the ideal, but when it breaks, nobody can stop it.
Design for AI-Driven Operations from the Start
There is another failure mode that shows up in practice. When organizations add AI to a process without thinking ahead, the account permissions and tool-access structures end up bolted on as an afterthought โ which creates more work, not less.
The order of operations needs to flip. Instead of “let’s try AI and see,” the mindset becomes: “This process will be AI-driven from day one โ so we’ll choose our systems, set up permissions, and design the workflow with that in mind from the start.” The ability to stop when needed and revert when necessary gets built into the architecture at the beginning. That is the new baseline for business process design.
Draw the Line by People, Not Tasks
Trying to categorize which types of tasks are “safe to delegate to AI” is harder than it sounds. Automated email looks like a safe category โ until those emails contain language tied to legal notices or contractual obligations. At that point, if no one who understands the content is watching, the risk is real.
The more accurate approach is to draw the line by person, not by task type. Does the person handing a task off to AI genuinely understand that domain? AI is a tool for organizing information and preparing the inputs for a decision. The decision itself belongs to a human. The right use of AI is not to have AI do the thinking โ it is to use AI so that humans can think more precisely.
The Era of Shared Expertise
You may be thinking: “We can’t afford to keep specialists in-house.” That is exactly why the model is changing. The traditional assumption was that each company needed to hire its own dedicated expert. But as AI capabilities advance, a small number of highly skilled specialists can support multiple organizations simultaneously, using AI to scale their reach.
For SMEs, the answer is no longer “build everything in-house.” It is “share expertise and pair it with AI.” Those shared specialists become the quality-control layer for AI-driven work. AI executes; specialists oversee. Running AI without any specialist oversight is the most dangerous state of all.
The Era of Blaming AI Is Over
One final point โ the most important one. When AI makes a mistake, the responsibility falls not on the AI, but on the person who used it. The more capable AI becomes, the more what is called into question is the judgment and expertise of the human operating it.
You only need to do one thing starting today. Look at the work your organization is currently delegating to AI and ask yourself, just once: “If this goes wrong, can I stop it?” Build the ability to stop before you start. Move forward while keeping the option to reverse course. That is the surest way to turn AI into a tool that works safely for the long term.
Comments