What Decisions Should You Let AI Agents Make?

Let an agent prepare any decision, let it own only the calls you could reverse within a month at a cost you have already accepted, and write down the point at which it must hand the call back to you. Everything else stays yours. That is the whole rule. The rest of this piece is why it holds, and why most people running agents today have not applied it.

The reason is not carelessness. Until recently AI advised and you acted, so the decision always passed through you. Agents change the order. They send the email, reschedule the client, approve the refund, reorder the stock, and report afterwards. Each of those was a decision. Nobody announced the transfer, because there was no moment of transfer. There was a setting, then a convenience, then a habit.

Are AI agents already making decisions for you?

Yes, if one of them can take an action without asking first. That is the working definition: a decision is any point where more than one outcome was possible and one of them happened. An agent that chooses which supplier to chase, which enquiry gets an answer today, or what wording goes to an unhappy customer has decided something, however modest.

The difficulty is that the decisions arrive unlabelled. A person you hire to handle your inbox comes with a job description, and you know roughly what they are allowed to commit you to. A tool you switch on comes with a list of features. Features describe what the agent can do. They say nothing about what it should be permitted to do, and the gap between those two is where authority quietly accumulates.

Run a simple test this week. List the last ten things your agents did without you approving them first. For each, ask whether you chose to give that agent that call, or whether it simply turned out to be possible. Most people find the second answer outnumbers the first.

Why does the outcome still land on you?

Because accountability does not delegate with the task. Your client does not care that the reply was drafted and sent by software. Your bank does not accept that a payment was released by an agent acting within its remit, if you never defined the remit. The consequence of the call arrives at your door whoever, or whatever, made it.

This is the same arrangement that governs handing work to a person, and the lesson from that is old. When you give someone a task without giving them the authority inside it, the task comes back. When you give them authority without limits, you find out where the limits should have been by hitting one. Agents behave the same way, with two additions: they act faster than a person, and they do not hesitate at the edge of their competence the way a nervous junior will. A person senses when a call is above their pay grade. An agent has no pay grade.

So the question is not whether to trust the agent. Trust is the wrong unit. The question is which specific calls have been assigned, to whom, with what ceiling. That is a design problem, and design is done in advance.

What is the rule for deciding what an agent owns?

Three tiers. Every call you meet belongs in exactly one.

Prepare. The agent gathers, compares, drafts, and recommends. You decide. This is the default for anything new, anything that touches a relationship you value, and anything you cannot easily explain afterwards. The agent saves you the labour of getting to the decision without taking the decision itself. For many people this tier alone is where most of the value sits, and it costs nothing in authority.

Own. The agent decides and acts, and you review the outcomes in batches rather than the calls one by one. A call earns a place here only when it passes three tests. It is reversible within about thirty days. The worst plausible result is a cost you have already accepted in writing, as a number. And you could state the criteria in a sentence a stranger would follow. Rescheduling a routine meeting qualifies. Refunding an order under a fixed sum qualifies. Declining a cold sales approach qualifies. Agreeing a discount does not, because the criteria are never a sentence long.

Escalate. The agent must stop and hand the call back, and it must be told exactly when. This tier is defined by thresholds, not by feelings: an amount, a named client, a first-time situation, a message containing certain words, any action that cannot be undone. The threshold is the important part. An agent told to "check with me if it matters" has been given a judgement it cannot make. An agent told to stop at any commitment above a stated sum has been given a rule it can follow without understanding your business.

Notice what the tiers do to your own workload. The prepare tier removes the legwork. The own tier removes the repetitions. The escalate tier concentrates what remains in front of you, so the calls that reach you are the ones that need you. This is not about delegating more. It is about making sure you are not, without having decided to, delegating everything.

How do you write the threshold down?

On a single page, kept where the agent can read it and where you can read it, one line per recurring call. Each line states the call, the tier, the ceiling, and the trigger that hands it back. The wording should survive being read cold by someone who has never met you.

Two disciplines make it work. First, set the ceiling before the agent is live, not after the first surprise. A limit chosen in the aftermath of a mistake is shaped by the mistake, usually too tight and abandoned within a month. A limit chosen beforehand is shaped by what you can afford. Second, review the page on a schedule, not when something feels off. Authority drifts in one direction, towards the agent, because every convenient exception gets absorbed and none gets withdrawn. A monthly look at which calls moved tiers keeps the drift visible.

You will find the first draft is wrong in places. That is fine. The cost of a wrong line on a page you review monthly is small. The cost of an undefined line you never see is the one that grows.

Does a small business need this, or only large ones?

A small business needs it more. A large organisation has layers: a second pair of eyes, a compliance function, a person whose job is to notice. A solo founder or a team of four has you. If an agent commits you to something unwise at half past three on a Thursday, nobody else is positioned to catch it, and you may not look at the output until Monday.

The material written for larger organisations tends to treat this as governance, with committees and audit trails. None of that scales down, and none of it is needed. What scales down is the page, the three tiers, and the monthly review. It takes an afternoon to draft and costs nothing to maintain.

It also protects the thing a small operation runs short of most, which is attention. Every call you have not consciously assigned is a call that may return to you at the worst moment, in the form of an apology to write or a mess to unpick. Assigning them in advance converts a diffuse background worry into a short list you can read.

That is the work Command Architecture is built for. It sets out who, or what, holds which authority, with the limits written in, so that the structure of your decisions exists outside your head and does not depend on your remembering what you permitted last month. It is a method for assigning the calls deliberately, and it applies as readily to an agent as to a person.

The agents will keep improving, and the range of what they can decide will keep widening. That makes the line you draw matter more, not less. A capable agent with no stated limits is not an assistant. It is an unfiled decision-maker with your name on the outcome.