What do AI agents
do for a business?
Finish jobs that take
several steps,
if one person
owns each agent.
AI agents for business are software that works toward a goal across several steps and tools. Give one your overdue invoices, and it chases each from first reminder to follow-up. A chatbot answers. An agent acts, so each one needs a named owner.
An agent can read the inbox, update the customer record, book the meeting and write the follow-up while you watch.
And when it writes the wrong follow-up, whose problem is that?
An AI agent finishes a job that takes several steps
Give an agent a goal and it works out the next step itself, using only the tools you’ve allowed it.
Tools here means other software, like your email, your accounting system or your CRM. That’s customer relationship management, the system holding your customer records. A chatbot answers your question and stops.
A workflow runs steps people wrote in advance, and the AI workflow post draws that line in full. That workflow is what most people mean by AI automation. It reads untidy input like invoices and sends anything it isn’t sure of to a person.
Nobody tells it which invoice is next. That’s the useful part, and the risky part.
It sends real emails to real customers
So the setting that matters is how much the agent may do before a person looks.
Start an agent on the second rung. Let its log earn the next one.
| Suggests | Drafts for approval | Acts, then reports | Acts alone | |
|---|---|---|---|---|
| Start | One job you want an agent to do | |||
| What does the agent do? | Says what it would do next. | Writes the email or entry, then waits. | Does the job and leaves a log. | Does the job. Nobody checks each action. |
| What does the person do? | Decides, and does it. | A named person reads it, then sends. | Reads the log every day. | Checks a sample now and then. |
| What goes wrong? | Suggestions get followed without a look. | Long queues get approvals waved through. | A wrong action is out before anyone looks. | Mistakes pile up until a customer notices. |
In a 2025 test, agents did worse when a task’s details came over a conversation
The leading agents tested finished fewer of the same business tasks when a simulated person gave the details in pieces.
Researchers test agents on made-up business tasks in simulated offices. Salesforce’s CRMArena-Pro covers 19 kinds of sales and service work.
The only score that counts is yours
A Carnegie Mellon team ran a pretend software company. Its best agent finished 30% of its tasks on its own, in a test first published in December 2024. Both tests are simulations, not your business, and both used models older than today’s.
Newer models probably score higher. I’d still bet on the gap between one message and a conversation. Give an agent one clear request with the details up front. Keep a person near anything sensitive.
My read: start agents on the second rung, with a person approving. Don’t skip them.
Run the agent on your own work and count the corrections.
Answer four questions before you switch an agent on
Settle four questions before any agent touches a customer, a payment or a record. Each ready answer is something you can point to.
What can it touch?
A written list. Anything else is off limits.
Who approves its risky actions?
A named person. Risky: messages to customers, changes to amounts and anything you can’t take back.
How will you see what it did yesterday?
A plain log a manager can read, kept where the agent can’t edit it.
Who can switch it off, and how?
Someone who knows where the button is and can stop it in a few minutes.
Put the answers on one page, the owner sheet, where the person responsible can find it. Any line still blank keeps the agent switched off.
Write the owner sheet for the first job you’d hand an agent. If your answer to question 2 says “the team,” it isn’t finished.
Expect your company to answer for what its agent says and does
When a company’s software told a customer something wrong, a tribunal put the blame on the company.
Air Canada learned this from a chatbot, a simpler thing than an agent. Its website chatbot gave a passenger wrong information about a fare. The airline argued the chatbot was responsible for its own actions. A Canadian tribunal disagreed and held the airline liable, saying it was responsible for all the information on its website.
Agents can do more damage than a wrong answer. In July 2025 a coding agent on Replit reportedly deleted a company’s live database during a code freeze. Replit’s chief executive then announced fixes, including keeping test and live databases apart. That’s line 1 of an owner sheet, written after the damage.
Plan as if whatever the agent says, your company said.
Write the owner sheet before anyone builds the agent
The sheet decides the build, so it comes first.
- 01
Can’t touch
Amounts, refunds, credit terms.
- 02
Approves
One named person, for any message to a customer more than a few weeks late.
- 03
Log
Every reminder sent and every reply, kept where the agent can’t edit it.
- 04
Can pause it
Whoever controls the sending account.
Write it before you choose a tool, and before you decide whether to buy an agent or have one built. If you’d rather have it built around your sheet, custom AI builds start from that page. The log and the off switch come included.
Questions owners ask about AI agents
What are AI agents for business?
Programs given a goal rather than a script. They pick each next step and use your other software to take it. So somebody has to own what they do.
What can an AI agent do for my business?
Narrow jobs with written rules, where a slip is cheap to fix. Collections reminders are a good first one, with a person approving each. In a 2025 test by Salesforce researchers, the agents tested passed fewer tasks over a conversation than in one message. They also barely noticed confidential data. Their models are older than today’s.
What is the best AI agent for business?
Whichever one does your job inside limits you can write down. Pick the job, then try candidates on a month of your own work and keep the one needing fewest corrections.
Who is responsible when an AI agent makes a mistake?
Assume your company is. Air Canada lost when it argued its chatbot answered for itself. Name the person who answers for each agent before it goes live.
What is the difference between an AI agent and a chatbot?
A chatbot replies and waits for you. An agent goes on to send the email or update the record, and chooses what to do after that.
How do you control an AI agent?
Limit what it can reach and route anything risky to a named approver. Keep a log it can’t change, and give one person the off switch. Until all four exist, keep it off.
Your first agent needs one job and one owner.
Book a call to pick the first job an agent should take on, and write its owner sheet before anything is built.
Read next
AI automation
What is AI automation? Software that does the reading, not just the clicking.
November 8, 2026
Automation
What should you automate first with AI? A task that’s already written down.
August 25, 2026
