Skip to content

AI That Acts, Needs a Boss#

AI that acts needs a boss — an agent without governance is just a liability with an API key.

If an AI system can touch a customer, a refund, or a record, it has a job.

A lot of teams gave it access before they gave it a job description.

Back in July, we asked a simple question on this blog: show me your AI policies.

Most owners couldn't. Not because they were careless, but because the tools moved faster than the rules. Everybody had a chatbot. Almost nobody had written down what it was allowed to do.

That was the chatbot question. This is the agent question.

The chatbots grew up. They stopped answering questions and started taking actions. They send the email. They update the record. They push the refund. They talk to your customer while you are asleep.

An AI agent without governance is just a liability with an API key.

The July question got harder#

A chatbot that gives a wrong answer is annoying. A human catches it and moves on.

An agent that takes a wrong action creates a record, issues a refund, sends a message, or pays a vendor before anyone sees it.

The gap between "AI that talks" and "AI that acts" is the whole ballgame.

The adoption curve kept moving. QuickBooks' July 2026 Small Business Insights survey says 4 out of 5 small businesses now use AI regularly, up from 48% in July 2024. That does not mean every small business is running agents yet. It means AI is already inside the workflows where agents show up next.

Deloitte's 2026 agentic AI research makes the governance gap harder to ignore: only 21% of surveyed enterprises had mature governance in place for agentic AI, even as 74% expected to use agents at least moderately by 2027.

The point is not that small businesses need enterprise bureaucracy. The point is that the oversight problem is arriving before most teams have named it.

Why "it worked in the demo" is not a policy#

The most common trap starts with a good demo.

You test an agent. It handles ten support tickets cleanly. It sounds sharp. It saves your rep two hours. You are impressed, so you turn it loose.

Then edge case number eleven shows up. A weird refund request. An angry customer using words that trigger the wrong response. A prompt that tricks it into doing something you never intended.

The demo was the honeymoon. Governance is the marriage.

Think about how you onboard a human. You would never hand a new hire your bank login on day one, point them at your customers, and walk out of the building. You give them a role. Limits. A manager. A way to escalate when they are unsure.

An agent needs the same thing, and it needs it in writing before it goes live, not after something breaks.

The four things every agent needs before it goes live#

You don't need a legal department for this. You need four answers.

  • Scope. What is this agent allowed to do, and where does it have to stop and ask a human? Write the line down. "Can draft refunds up to $100, must get approval above that." Specific beats vague every time.
  • A paper trail. Every action it takes should be logged somewhere a person can read later. If you can't answer "what did it do last Tuesday," you don't have an agent, you have a rumor.
  • An owner. One human name attached to it. Not "the team." A person who is responsible for what it does and who gets the alert when it does something odd.
  • An off switch. A kill button anyone on shift can find and use in under a minute, without calling a vendor. If turning it off is hard, you have already lost control of it.

That's it. Scope, logs, an owner, an off switch. You can write all four on an index card. Many companies have not written down even that much.

This is a management problem wearing a tech costume#

Nobody selling you software will say this.

None of this is about AI. It is the same thing your business has always run on: who is allowed to do what, who checks the work, and who is on the hook when it goes wrong.

You already solved this for people. You have approval limits. You have someone who signs off on the big stuff. You have a way to remove access when a person should not have it anymore.

Agents just move faster and never get tired, which means a small gap in oversight scales faster than a person ever could. The FTC has already said there is "no AI exemption" from the laws on the books. The practical translation for an owner is simple: if a system acts on your behalf, someone still has to own the result.

That is how we run our own agent work inside JOV. For this post, an agent could audit sources, draft the revision, and leave a review trail. It could not publish the post. Dyke still owns the final yes.

The good news is that if you can manage a team, you can manage an agent. The instinct is already there. You just have to point it at the software.

Start small, but start today#

You don't have to govern everything by Friday.

Pick the one agent or AI-enabled workflow that touches money or customers. Just one. Then answer the four questions for that single system:

  1. What can it do without a human saying yes?
  2. Where is the log?
  3. Whose name is on it?
  4. How do you turn it off?

If you can answer all four cleanly, you are ahead of the problem. If you can't answer even one, you just found the thing that keeps you up at night, and it was hiding in plain sight the whole time.

An agent with governance works for you. An agent without it is a bet you didn't know you placed.

Let's pressure-test yours#

If you are not sure which of your tools have quietly turned into agents, that is exactly the conversation worth having.

We will walk through what you have running, where the gaps are, and the four-question checklist for one workflow. You will leave with a clear picture of what is acting on your behalf, what needs a human approval step, and where the off switch lives.

Schedule a 30-minute agent guardrail review and we will map it out together.

Sources: