You've decided an AI agent is going to take over part of the work your team does today. Maybe it answers the website chat after you close, sorts the shared inbox, or drafts quote summaries for someone to check. Setting it up is the easy part. The hard part is the conversation with the people who will work next to it.
Some of them may walk into that conversation already uneasy. Pew Research Center surveyed 5,273 employed U.S. adults from Oct. 7-13, 2024, and its report, published February 25, 2025, found that "About half of workers (52%) say they’re worried about the future impact of AI use in the workplace" and that "32% think it will lead to fewer job opportunities for them in the long run". On the other side: "Only 6% of workers say workplace AI use will lead to more job opportunities for them in the long run." And "Some 17% of workers have not heard about the use of AI in the workplace."
Those numbers describe U.S. workers in general, surveyed in October 2024. They are not a measurement of your staff, or of small-business staff in particular. But they're a fair warning: some of your people may assume the worst before you say a word, and a few may never have thought about AI at work at all. What you decide, what you say and the order you do it in carry most of the weight.
This guide walks through six steps in the order you'd actually do them, with the way each one tends to go wrong. It isn't legal or employment advice. If anyone's hours or role will change, check the rules that apply to your business.
What an Agent Changes, and What You Decide
Here's an honest backdrop you can repeat without overpromising. A U.S. Census Bureau Center for Economic Studies working paper (CES-24-16, March 2024, revised October 2024), "based on the Business Trends and Outlook Survey for September 2023 to February 2024", reports: "AI users often utilize AI to substitute for worker tasks and equipment/software, but few report reductions in employment due to AI use." It adds: "Many firms undergo organizational changes to accommodate AI, particularly by training staff, developing new workflows, and purchasing cloud services/storage."
Read that carefully. It's what businesses reported about themselves, in a research paper rather than an official Census Bureau statistic, covering late 2023 and early 2024. It's a snapshot, not a forecast, and it guarantees nothing about your business.
The useful part is the split it describes. An agent takes over tasks. What happens to anyone's hours or job is not something the software decides. You decide it. So decide it before you announce anything.
Step 1: Decide Before You Announce
Before you say a word to your team, write down your answers to four questions:
- Which tasks will the agent take? Be specific. "Answers routine questions in the website chat after we close" is a task. "Helps with customer service" is a wish.
- Which tasks stay with people? Name them, so nobody has to guess whether their part of the job is next.
- Where does the agent hand work to a person, and to whom? Every agent needs a point where it stops and passes something on.
- What happens to the time it frees? There are only a few honest answers: more of the work your team never gets to, different duties, or fewer hours. That's your decision, and only you can make it. Whichever one is true is the one you say.
If you haven't settled that last question yet, that's fine. But know that you haven't, so you don't accidentally answer it in the meeting.
Where it goes wrong: You announce first and decide later. The gap fills with rumor, or with a quick reassurance that later turns out to be false.
Step 2: Tell the People Whose Work It Touches First
Start with the people whose daily work changes: the person who answers the chat, the one who sorts the inbox, whoever enters the bills. Tell them in person, or on a call if you work apart, and leave time for questions. Then tell everyone else.
Do it before the agent is visible to customers. Nobody should find out from a customer who mentions "your new chat thing," from a strange message in the shared inbox, or from a login they didn't expect.
A group email isn't enough for the people most affected. Email can't answer the follow-up question someone is too nervous to type.
Where it goes wrong: Staff discover the agent on their own and fill in the blanks with the worst version.
Step 3: Describe the Agent by Its Job, Not by Hype
Skip the big talk about what AI can do. Your team needs three plain sentences:
- What it does. "It answers common questions in the website chat after hours and offers open appointment times."
- What it will never do. "It never quotes a price that isn't on our price list and never gives medical, legal or billing answers."
- When it hands off, and to whom. "Anything it can't answer goes to the office manager first thing in the morning, with a short summary."
Those are examples, not a script. Use your own answers from Step 1. If you can't fill in those three sentences yet, the process probably isn't written down yet either. Our guide to writing a process down before handing it to an AI agent covers that part, so do it first.
Where it goes wrong: You describe an agent that can do anything. Some staff then trust it too much, and others ignore it, because neither group knows where its job ends.
Step 4: Promise Only What You've Decided
This is the step that protects your credibility. Sort everything you might say into three piles:
- What's fixed. Say it plainly, including what happens to the time the agent frees, if you've decided.
- What isn't decided yet. Say that it isn't decided. People handle "I don't know yet" far better than a guess that changes later.
- When they'll hear more. Give a real date for the next update, and keep it.
The tempting line is "Don't worry, nobody's job is changing." Only say it if it's settled and you mean to stand behind it. The opposite mistake counts too: don't hint at cuts you haven't decided on, just to sound realistic.
Where it goes wrong: A reassurance gets contradicted three months later. That costs you more trust than the agent will ever save in time.
Step 5: Name the People Staff Go To
Staff need to know who to talk to. Name three things, even if one person covers all of them:
- Who owns the agent day to day. This is your team's point of contact for it.
- Who staff go to with a problem. Usually the same person, plus a fallback for days they're out.
- Who decides whether a change gets made. When someone says "it keeps getting this wrong," who says yes or no to fixing it?
The National Institute of Standards and Technology (NIST), the federal standards agency, makes the same point in its voluntary AI Risk Management Framework (AI RMF 1.0, January 2023): "Policies and procedures are in place to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems." For a ten-person business, that comes down to one owner's name and one fallback name, written down where everyone can find them.
Checking the agent's work on a schedule is a separate job. The process write-up guide linked in Step 3 covers who reviews the work and how often, so there's no need to repeat it here.
Where it goes wrong: Everyone assumes someone else is in charge of the agent, so problems get mentioned in passing and never fixed.
Step 6: Make Its Mistakes Easy and Blame-Free to Report
Your staff will spot the agent's mistakes before you do. A wrong answer in the chat, a message sorted into the wrong pile, a bill entered with the wrong date. Make it easy for them to say so:
- One simple place to flag a problem. A shared channel, a short form or a notebook at the front desk. Pick one and tell everyone.
- No blame. The agent's errors are not the staff member's fault, and nobody is expected to quietly cover for it.
- Close the loop. When something gets changed because someone flagged it, tell the team what changed. That's what keeps people reporting.
NIST's framework calls for the same two habits. One: "Feedback processes for end users and impacted communities to report problems and appeal system outcomes are established". Two: "Incidents and errors are communicated to relevant AI actors", which in plain words means the people who run and work with the system hear about its mistakes. The framework is voluntary and written for organizations of every size, so treat these as sound principles, not rules a small business must follow.
There's one more thing to say out loud. NIST's profile for generative AI (NIST AI 600-1, July 2024) warns that "humans may over-rely on GAI systems" (GAI means generative AI, the kind that writes text) and calls this "automation bias, or excessive deference to automated systems." So tell your team in plain words: the agent can sound completely sure of itself and still be wrong. When their judgment and its answer disagree, their judgment wins.
Where it goes wrong: Staff quietly work around the agent's mistakes, or they go along with it when it's wrong, because nobody told them they were allowed to overrule it.
A Sample of What You Might Say
This is an illustrative example written for this article. It is not a quote from any business. Swap in your own decisions, and cut anything you haven't actually decided.
"I want to tell you about a change before you see it anywhere else. Starting [date], we're adding an AI agent to [one job, for example answering the website chat after we close]. Here's what it will do: [the task]. Here's what it won't do: [the limits]. When it can't handle something, it passes it to [name].
Here's what's decided: [what happens to the time it frees, said plainly]. Here's what isn't decided yet: [anything still open]. I'll update you by [date].
[Name] owns the agent day to day, and [name] covers when they're out. If it gets something wrong, tell [name] or post it in [place]. That's never your fault, and you don't need to fix it quietly. It can sound confident and still be wrong. When that happens, trust your judgment over its answer. I'll let you know what we change because of what you flag.
What questions do you have?"
Frequently Asked Questions
Should I tell staff before customers see the AI agent?
Yes. Tell the people whose work it touches first, in person, before the agent answers a single customer. Nobody should learn about it from a customer, an odd message or a login they didn't expect.
What if I haven't decided what happens to anyone's hours or role yet?
Then say that it isn't decided, and say when you will tell them more. Don't fill the gap with a reassurance you might not keep. If hours or roles will change, check the rules that apply to your business before you announce it.
Who should own the AI agent in a very small business?
One person, by name, with one fallback. That person is who staff go to with a problem and who decides whether a change gets made. Write both names down where everyone can find them.
Is this the same as a policy on staff using AI chat tools?
No. This is about an agent your business puts to work beside your staff. What staff may do with their own AI chat tools and accounts is a separate decision, which we cover in our article on staff using AI chat tools on personal or work accounts.
What Stays Your Job
An agent can take real, repetitive work off your team's plate: the first reply to routine questions, the sorting, the typing, the drafts a person then checks. It can't make the decisions in Step 1, keep the promises you made in the meeting, or act on what your staff report. Those stay with you, and they're what your team will judge the whole change by.
If you're working out which single job an agent should take on first, our pricing page explains that agent work is scoped and priced individually rather than sold as a plan, and that each automation gets its own scope, price and monthly usage, agreed with you before we start. Or start a conversation with us and tell us which job you have in mind.