"What is an AI agent for a business?" Owners ask it after the third vendor in a month has used the word, or after a member of staff mentions that an agent now sends the reminders. The honest answer is simpler than the word suggests. An agent is software that is allowed to take steps on its own between your instructions: read something, decide something small, act in another system, and keep going until the job is done or someone stops it. Everything that matters to you as an owner follows from that one permission: what it may touch, where it must stop and ask, who reads what it did, and how it is switched off.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. The most expensive mistakes I saw were not made by software; they were made by people who let a process run without anyone owning it. An agent is that process with the people removed, which is why the owner's decisions come first. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs. This page is the owner's version of the question. Not legal advice; a working owner's plain view.
The short answer
A chatbot answers when you ask. An agent acts when you are not looking. Give it a goal, a set of systems it may reach and a set of rules, and it works through the steps on its own: it reads the inbox, pulls the record, drafts the chase, files the document, moves to the next one, and stops when the list is empty or when a rule tells it to stop. The technical definition, and how an agent differs from a chatbot, is on what is an AI agent; how it differs from plain automation and from a fixed workflow is on automation, workflow or agent. This page assumes you have read neither and need to run a company anyway.
In a company of fifty to five hundred people, an agent earns its place where the steps are boring, written down and reversible: gathering, drafting, filing, matching, chasing. It becomes dangerous where a step is a promise to a customer, a payment, a hire, a price or a deletion, because those steps are hard to undo and the agent will take them with the same confidence it takes the harmless ones.
Do you need one? Not until you have one boring process whose steps you can write on a page, a person who will own the agent that runs it, and a month to watch it work with that person checking the output. Most owners who ask do not yet have the page, and writing it is the first job.
Where it earns its place
The best first job for an agent is the one a careful junior employee would do without ever being asked to use judgement. Some examples from companies this size:
- Chasing what is late. Reading the list of overdue invoices each morning, drafting the reminder in your tone, and putting each one in a folder for a named person to approve and send. The agent never decides whether to escalate.
- Matching paper to records. Reading the supplier invoices that arrived, finding the purchase order each one belongs to, flagging the ones that do not match. A person decides what to do with the flags.
- Preparing the day. Pulling the jobs, the people, the addresses and the promised windows into one board before the dispatcher arrives. The dispatcher still owns the board.
Each of these shares the same shape: the steps are known, nothing leaves the building without a person, and if the agent gets one wrong the fix is a minute, not a month. Can AI be my assistant covers the same line for the owner's own desk.
What it must never do alone
Promise. A delivery date, a price, a discount, a credit, a job offer. Any sentence that commits the company is drafted by an agent at most, and sent by a person.
Pay or move money. No agent holds a bank login, approves a payment run or changes a supplier's bank details. This rule has no exceptions, and it is the one that the people selling agents most want you to relax. Not accounting or tax advice: a control question, and the answer is no.
Hire, discipline or let go. Nothing about a named employee's pay, job or conduct passes through an agent without a person deciding.
Delete or overwrite. An agent that can delete records can delete the wrong ones at the speed of a machine. Give it the power to add and to flag; keep the power to remove with a person.
Speak as the company in public. Reviews, social posts, statements to a customer who is angry. Drafted, perhaps. Posted, never.
Reach systems you did not list. An agent that can log into "whatever it needs" is not a tool; it is a new employee with every key in the building and no manager. The list of systems it may touch is written down before it is switched on, and anything not on the list is off limits.
The data rule
An agent reads more than a chatbot ever did, because reading is its job. So the data rule matters more here than anywhere else. Decide, in writing, which systems and which records it may reach, and which it must never see: customer lists in full, pricing, payroll, bank details, contracts under negotiation, anything about an employee's health or pay, anything a customer told you in confidence. Decide where its output goes and who reads it. And run it only on a company account with written terms that say who keeps the data, how long, and what it may be used for. Never through a free personal login, because when that person leaves, the agent and its history leave with them. Is ChatGPT safe for business use answers the account question; what is shadow AI is what happens when people set up agents on their own because nobody wrote the rule.
Five checks before you switch one on
First, write the process on one page. Every step, in order, with the rule at each step: what counts as late, what counts as a match, what to do when the field is empty. If you cannot write the page, no agent can run the process; it will only run your confusion faster.
Second, name the owner. One person inside the company who switches the agent on, reads its log every morning for the first month, and switches it off when it misbehaves. Not the vendor, not the IT provider. Who owns AI in a company is the longer answer; the short one is the person staff already ask.
Third, list the systems it may reach and nothing else. The inbox, yes. The order system, read only. The bank, never. The list is part of the page.
Fourth, decide where it must stop and ask. Anything that leaves the building, anything that changes a number a customer will see, anything about a named person. At those points the agent drafts and waits.
Fifth, insist on the log and the off switch. Before you buy, ask for the record of what the agent did yesterday, step by step, and ask how it is stopped. If either answer is vague, walk away.
Check yourself
Seven yes or no questions. Each no is a decision still to make.
- Is there one boring process in your company written on a page, step by step, with the rule at each step?
- Is there a named person who would own an agent running it, and read its log every day for the first month?
- Have you written the list of systems the agent may reach, and the list of records it must never see?
- Have you written where it must stop and ask before a step?
- Is every agent in use on a company account with written terms, not a free personal login?
- Could you switch each one off in a minute, and do you know how?
- Do you know how many agents are already running in your company today, set up by whoever had the idea?
Mostly no? That is the normal starting point. The AI readiness scorecard asks ten questions like these across the whole company.
What to ask a vendor who says "agent"
The word is now attached to everything, so the questions matter more than the label. Ask these and write the answers down.
- Show me the log. Every step the agent took yesterday, in plain words, with the time. If there is no log, there is no agent you should buy.
- Show me the stop button. How a person halts it, how fast, and what happens to the half-finished work.
- Which systems can it reach, exactly? A list, not a sentence. And how is that list changed, by whom?
- What does it do when it is not sure? The right answer is that it stops and asks a named person. The wrong answer is that it "uses its best judgement".
- Where does our data go, and who else can see it? Terms in writing, on a company account.
How to evaluate an AI vendor's claims covers the rest of the conversation. A vendor who cannot show the log, the stop button and the list of systems is selling a word.
Starting safely
Pick one process from the list above, the most boring one you have, and write its page. Choose the owner. Set the agent to draft only for the first month: it reads, decides, prepares and files, and every step that would leave the building or change a record waits for a person. Have the owner check the log every morning and keep a note of two things: what the agent got wrong, and how long the checking took. At the end of the month you will know whether the process was as clear as you thought, whether the agent is worth its cost after the checking time, and whether the owner wants to keep it. Only then decide which steps, if any, it may take without waiting.
Do I need an AI strategy is the five-decision version of the same page, if you want the wider frame before the first agent.
Picture it going wrong
A made-up example, to show the shape. An eighty-person wholesale distributor buys a collections agent on the strength of a demonstration. It is switched on against the accounting system and the email on a Friday afternoon, allowed to send, because the setup screen recommended it. Over the weekend it reads every open invoice, decides which are late by its own definition, and sends firm reminders to well over a hundred customers, including the eleven who had agreed payment plans with the owner in person, the three who had paid by cheque that week, and the largest account, whose invoice had been held back on purpose during a price negotiation. On Monday the phone starts at eight. Nothing the agent did was wrong by its rules. The rules were never written, the exceptions lived in the owner's head, nobody was reading the log, and the off switch turned out to be a support ticket.
Mistakes that repeat
- Buying the agent before writing the page.
- Letting it send, pay or delete in the first month.
- Giving it access to "everything it needs" instead of a list.
- No named owner, so the first person to notice a problem is a customer.
- Judging it by the demonstration instead of by its log.
- Counting the hours it saves without counting the hours someone spends checking it.
Frequently asked questions
What is an AI agent for a business?
Software allowed to take steps on its own between your instructions: read, decide small things, act in other systems and keep going until the job is done or someone stops it. In a company of fifty to five hundred people it earns its place on boring, written-down, reversible work, and it must never promise, pay, hire, delete or speak for the company alone.
How is an agent different from a chatbot or from automation?
A chatbot answers one question and waits. Plain automation follows a fixed set of steps with no judgement. An agent works through a goal, deciding small things along the way, and keeps going. The longer answer is on the foundations pages linked above.
Do I need an AI agent?
Not until you have one process written on a page, an owner for it and a month to watch it. If you have those three, one agent on one boring process is a reasonable test. If you have none, the page is the first job and it costs nothing.
What should an agent never be allowed to do?
Send a promise, move money, change bank details, hire or discipline, delete records, post in the company's name, or reach a system that is not on its written list. At each of those points it drafts and waits for a person.
Is it safe to let an agent read our email and files?
Only on a company account with written terms, only for the systems and records on its list, and only with the records it must never see written down first. Never through someone's personal login.
What should I ask a vendor who says their product is an agent?
Show me the log, show me the stop button, list the systems it can reach, say what it does when unsure, say where our data goes, and tell me what it got wrong last month. Vague answers to any of these end the conversation.
How do I know after a month whether it was worth it?
Three things: the mistakes in the log and whether they shrank, the checking time against the hours it saved, and whether the owner wants to keep it. If the checking time is most of the saving, the process was not as clear as you thought, and the fix is the page, not the tool.
Not sure where you stand? Take the 3-minute AI readiness scorecard.