Let’s talk about the mistake that had no owner — and the person who paid for it anyway.
There is this moment every small business owner knows. The one where you look at the damage and you don’t even need to ask who did it, because the answer is going to make you want to throw something.
Picture a woman in Pretoria who runs a maintenance company. Fifteen staff, thirty clients, one tired owner. She sets up an AI agent to handle supplier inquiries and rebookings, because that is the kind of admin that eats a Thursday whole. The agent is good. Fast. Never complains about the spreadsheet.
One Friday, it misreads a clause in an email and cancels the wrong recurring order. The order that feeds her biggest client’s monthly service run.
Monday morning, the client’s site is not serviced. She calls the supplier. “Your system did it.” She calls the agent vendor. “It followed your configuration.” She calls her client. “I’m sorry.”
She pays the penalty. She eats the cost. She does the repair work herself, on her own time, to keep the account.
The machine acted. The machine answered for nothing. She answered for everything.
Now, none of this is new. Machines have broken things since the first machine. A crashed server, a misfiled tax return, a typo in a batch email — those all cost people money and sleep. But there is something different when the machine doesn’t just execute, it decides.
Because that is what agents are. They are not tools that wait for your finger. They are entities that act on your behalf — negotiate, choose, commit, spend. And we are building them in their millions, with almost no thought given to the one question every human organization was built to answer: who is responsible?
In the human world, responsibility has always followed the act. That is the whole trick of accountability. You can fire the person. You can sue the company. You can look the person in the eye and demand a repair, and something in the social machinery responds. Apology. Compensation. A changed process. The weight has somewhere to land.
The agent breaks the chain. The act happens. The consequence lands. And the blame disperses into a system that cannot feel it, cannot apologize, cannot be fired, and cannot be made to change. The supplier points at the software. The software vendor points at the configuration. The configuration was yours, somehow, even though no human ever made that specific call.
That is the part people are missing. This is not a customer service problem. It is a moral architecture problem.
We are building entities with the power to act and no capacity to answer. That combination has a name in every legal system on earth. It is called an unaccountable power. And unaccountable power does not stay neutral. It grows, because the people who deploy it keep the upside and push the downside onto whoever is standing closest.
Watch how quickly the blame laundering starts. When an agent misfires, the vendor says the tool is “as-is.” The employer says the human was “in the loop.” The human says the system did it. And somewhere in that three-way shrug, the bill arrives at the door of the one person who was actually there.
That is what human-in-the-loop has quietly become. Not oversight. A liability shield. The human is kept in the loop precisely so the machine’s errors have a pulse to attach to. The signature does not transfer authority to the human. It transfers the machine’s risk onto the human, and launders the error through a name.
Doctors feel this already. Traders feel it. Compliance officers live inside it. And now every small business owner who lets an agent touch money or promises is going to feel it too.
Here is the principle that has to survive: authority and answerability belong to the same person. If you hold the authority to act, you carry the answerability for the act. And if you delegate the act, you do not delegate the responsibility — you keep it, with your name on it, and you build the checks that make that real.
The moment a system can act without anyone answerable, you have not built leverage. You have built a machine that can do anything and answer for nothing — and the humans around it will spend their lives absorbing the consequences of decisions they never made.
So before you let the agent act, ask the only question that matters. Not “can it do the job?” It can. The question is: who answers when it doesn’t?
Because the machine will never have to sit across from the client and say the words. That is still a human thing. It might be the last human thing.
Hold on to it.