AI Agents for Business: When to Use One

An agent earns a place in your business only when the job branches and a person still checks the output. If the steps never change, write a plain workflow and stop there. Anthropic’s December 2024 essay says to start with the simplest thing that works. It also splits a fixed workflow from an agent that picks its own tools.¹ We like that split, because vendors have blurred it for a year. Now even a reminder that fires the same way each time gets called an agent. That reminder is a workflow. An agent makes sense when the next step depends on what the last step found, and you cannot write every branch in advance. Even then a named person has to own the stop. An agent that nobody checks is just a risk with a nice name.

Cited in this note

40%

The people guide helps when the gap is who checks the work, not which model you rent. The definition covers the words, in case they mean different things around your table. The AI agent comparison goes one step further, for when someone wants a fancier label than the job needs. Most owners we sit with do not need the label. They need one job held in place until a named person says it can go.

Eight jobs, each with a stop

Inbox triage waits for a person before anything is sent. Meeting notes wait before they turn into tasks. Quote follow-up waits before a price goes out, and invoice questions wait before anyone promises a payment. Inventory exceptions wait before a purchase, and support replies wait before they reach the customer. Schedule changes wait before a customer is told, and a weekly report waits before anyone treats it as the record.

We would start from those eight and no others. Each one already happens in a shop, a clinic, or a trades office. Each one also has a moment when a wrong move hurts a customer or the books. If you use an agent, it gets the work ready and the person in the table below decides. If you cannot point at that person on a Tuesday, you do not have a real job yet. You have a demo.

Gartner said in June 2025 that over 40% of AI agent projects will be canceled by the end of 2027.² The reasons were cost, unclear value, and weak risk controls. Any job from the list with no named checker is a good fit for that pile. Cost grows when the agent keeps getting more systems because nobody drew an edge. Unclear value shows up when you cannot say which wait got shorter. Weak risk controls show up when the output goes out without waiting for anyone. A canceled project is not a moral failure. It is what happens to ideas that never became jobs. We would rather you skip a test than join that pile with a good story.

A worked example

Take a clinic schedule, where an agent can be honest. The front desk owns the book, and the practitioner owns which visits are in scope. A patient asks to move a Tuesday visit later in the week. Whether that works depends on what the book shows and on whether that practitioner does that kind of visit. It also depends on a note that billing will need later. You cannot write every branch on a card, so someone has to look. An agent can look, draft the reply, and then wait. The front desk still approves the message, because some patients need a call and some need two plain sentences. The draft does not know the family. If the move changes the care plan, the practitioner signs off too. That is the stop, and it has a real name on it, not a switch that no job description mentions.

Now take the shop’s weekly report. People like to call it an agent job because the word sounds busy. The report has the same shape every time, with yesterday’s sales, the items that ran low, and the invoices still open. Nothing in it depends on a find that changes the next step, so a plain workflow can build it. A person then accepts it as the record, or it stays a draft. If you buy an agent for that job, you pay for judgment you did not need. You also skip the judgment you did need, which is the bookkeeper deciding if the numbers are right. Until she does, the report is a rumor in a nice font.

A trades desk shows both shapes in one afternoon. Quote follow-up branches when a customer asks a question that changes the price. The agent can pull the old notes and draft a reply, but the owner approves the price before anything goes out. Invoice questions are stricter. A promise to pay is never the agent’s to make, so the person approves any promise. Inventory exceptions wait before a purchase, since the wrong part costs a truck trip. Support replies wait before the customer sees them. Schedule changes wait before anyone says the van is coming, and meeting notes wait before they become tasks the lead tech never agreed to. Inbox triage waits before anything is sent, even the cheerful draft that was about to go to the wrong thread. None of this needs a new department. All of it needs the wait to be real, because a wait that means “someone usually glances at it” is only a wish.

How we’d run it

First, write the job as if it were a plain workflow. Say who starts it, what they make, and who gets it. If you can finish that page without a branch, ship the workflow and stop. Anthropic’s advice to start simple is the whole method here. An agent is an exception you earn, not a default you announce.

Second, circle the step where the next action depends on what you found. On the clinic book, the next message depends on the open slots and on whether the visit is in scope. On quote follow-up, the next sentence depends on whether the customer’s question changed the job. If you cannot circle a step like that, you have circled a feeling. A feeling does not make a project.

Third, if you did circle one, name the person who approves, and treat the table as the least you do. Use the name of the actual person at your desk, not a role like operations. If that person is “whoever is free,” you have already described the cancellation. Free people rarely check the work, but named people do, even on their grouchiest mornings.

Fourth, decide what wrong looks like before you turn anything on. It might be a visit moved to the wrong day or a price the owner never said. It might be a payment promise the bookkeeper would not make, or a draft report treated as the books. The risk control is that list plus a place where the output sits until someone checks it. Gartner’s reasons are not abstract here. Weak risk controls just mean a send button with no waiting room.

Fifth, run one job from the list and leave the rest with a person, the old way. Inbox triage or schedule changes are enough to start. Do not light up quote follow-up, support replies, and the weekly report all at once. Give the job one branch, one checker, and one week of real messages. If the checker starts to skip the wait because the draft is usually fine, you have learned something useful. Usually fine is how the wrong message leaves the building.

Sixth, come back to the list only when the first job feels dull. That means the person still checks, the branch still matters, and you can say what value showed up without making up a chart. If you cannot, turn it off before cost becomes the reason someone else does. Ending a small test is a win. Feeding it until the hype fades is how that canceled pile gets its next member.

Common mistakes

The first mistake is calling every automation an agent, and the vocabulary notes exist to stop it. A fixed sequence that files a photo and waits for a price is a workflow. It can be excellent, but it does not get better when you rename it. We have watched owners pay an agent premium for a checklist and then feel cheated when it acted like a checklist. The checklist was honest, and the brochure was not.

Letting the output send itself is the mistake that reaches a customer. The table is blunt on purpose. Someone decides what is sent, someone approves what the customer sees, and someone approves the schedule message. Take those lines out and a confident drafter holds the keys. Confidence is not permission. The front desk knows who needs a phone call, and that knowledge will not be in the prompt you wrote on a calm afternoon.

Starting from the model instead of the branch is how tests turn pricey and vague. Someone wants “an agent for the business” the way they might want a new hire with no desk. But there is no seat called the business. There is inbox triage, or there is not. There is a weekly report that someone must accept, or there is a PDF that will be quoted as fact. Pick a row. If none of the rows hurt this month, you have curiosity and not a project. So read the definition post and hold off on a platform deal.

Ignoring the forecast because it sounds like a big company’s problem is how small firms get a smaller version of the same ending. Over 40% is not weather that only hits large firms. Cost, unclear value, and weak risk controls fit a clinic or a trades desk just as well. Your version will not make the news, but it will make a monthly bill and a queue of drafts that nobody trusts. You do not need a bigger disaster to take the warning.

Giving the check to a committee is a polite way to have no checker. “Someone will review it” means the draft sits there until it does not. Write one name, and give that person the power to say no without a meeting. If the job matters enough for an agent, it matters enough for a name. If it does not matter enough for a name, it does not matter enough for an agent.

What we’d do

We would start with a workflow and make the agent earn the exception, even when the demo is charming. A charming demo is not a branch. A real branch is a schedule change that alters what you may say, or a customer question that changes a price. If we cannot see that in a real job from last week, we ship the fixed path and go home. Going home is a fine result, because not every visit should end with a new system.

We would put a person’s name on the check before anything is turned on, and we would treat the table as a contract. The person decides what is sent and accepts the tasks. The person approves the price, any promise to pay, and the order. The person also approves what the customer sees and the schedule message, and accepts the report as the record. If a proposal will not attach a name to the row you picked, that is the missing risk control showing up early. Believe it.

We would not run the whole list just to feel modern. One job held at the stop beats eight half-starts that get canceled for the reasons Gartner already gave. If you want a test this week, pick schedule changes or inbox triage and write down the name of the approver. Then refuse every extra system until that person has lived with the wait. You can add the agent once the wait is real. It does not get to create the wait by existing.