Someone will try to sell you an AI agent before the year is out.
Here is what that means. An AI agent is AI that takes actions rather than giving advice: it sends the email, books the slot, updates the record. For most small and medium-sized businesses, agents are worth trying now on narrow errands that are easy to undo, with a named person watching. They are not yet worth trusting, unchecked, with money, customers or personal data. Deciding which is which is a delegation skill, not a technical one.
What is an AI agent, in plain terms?
An AI agent is software that takes a goal, works out the steps and carries them out through your own systems, where an assistant stops at a draft for you to use. The line that matters is who presses the last button.
Picture a customer complaint. When you ask ChatGPT or Microsoft Copilot to draft a reply, you read the draft, fix the tone and press send. When an agent handles the same complaint, it reads the email, looks up the order, issues the refund and writes back. You find out afterward, if you look.
The UK's data protection regulator, the ICO, describes agentic AI in its Tech Futures report of January 8, 2026 as AI that "combines the capabilities of generative AI with additional tools and new ways of interacting with the world." The word doing the work there is tools. An agent is only as useful, and only as dangerous, as the inbox, calendar, accounts and customer records you connect it to.
How many small businesses are actually using AI agents?
Very few small businesses use agents regularly, and far more are curious than committed. A survey of 1,000 UK small business owners carried out in August 2026 and released on September 17, 2026 by the Small Business Institute with Alibaba.com found that 44% had heard of agentic AI, 5% used it regularly and 68% wanted to learn more. The same owners were confident with ordinary generative AI: 70% said so, yet only 48% used it regularly.
The US picture looks similar from a different angle. Census Bureau researchers working with the 2026 AI supplement to the Business Trends and Outlook Survey (Bonney and colleagues, April 2026) found that among firms using AI, 57% use it in three or fewer business functions, and most confine it to a handful of tasks, led by writing, document analysis and information search. AI-related cuts in employment were rare, reported by 2% of firms.
The loud version of the agent story comes from large companies. In a Zapier survey published December 15, 2025, 72% of executives at US companies with 1,000 or more employees said they were using or testing AI agents, and 20% said their AI systems now operate autonomously with minimal oversight. That is the world the pitch is written in. It is not the world of a 20-person firm with no IT department.
My reading of these numbers: the gap between hearing about agents and using them is not a failure. A small business is right to be slower than a corporation to hand over the keys, because it has no security team to notice when something goes wrong. The mistake is to treat the gap as either a reason to do nothing or a reason to buy something.
What changes when AI acts instead of advises?
Three things change when AI starts acting: mistakes travel further before anyone sees them, responsibility stays exactly where it was, and your staff's job moves from doing the task to directing and checking it.
Mistakes travel. A bad draft gets caught when a person reads it. A bad action has already happened. A wrong refund, a double booking or an email to the wrong customer is out of the building before anyone blinks.
Responsibility stays with you. The ICO's report is blunt: "organisations remain responsible for data protection compliance of the agentic AI they develop, deploy or integrate." Buying an agent from a vendor does not move accountability to the vendor. The same report flags "rapid automation of increasingly complex tasks resulting in a larger amount of automated decision-making", which matters in the UK because decisions about people carry extra legal duties. In the US there is no single agent law, but the rules that already govern how you treat customers and their data still apply to whatever your agent does in your name. And the rulebook is still being drafted: NIST launched its AI Agent Standards Initiative on February 17, 2026 because, in its words, "[a]bsent confidence in the reliability of AI agents and interoperability among agents and digital resources, innovators may face a fragmented ecosystem and stunted adoption."
The work changes shape. This is the future-of-work part, and it is the part owners underestimate. When an agent does the errand, someone has to brief it, set its limits and judge its results. That is management, done at small scale by people who have never managed. It is why I keep returning to the Scarcity Rule: when everyone has access to the same AI, a skill matters in proportion to how much scarcer it becomes as AI improves, not how useful it is today. Agents make the three scarce skills, framing, judgment and ownership, more valuable, not less.
The Errand Test: which jobs can you hand to an agent?
The Errand Test is my suggested way of deciding, one job at a time, whether an agent should act or a person should. It works because it asks the questions you would ask before trusting a new temp with a job, which is roughly what an agent is.
The Errand Test is four questions I suggest an owner asks before letting an AI agent act on the business's behalf: Could you write the errand down in one paragraph a temp could follow? Can you name everything it is allowed to touch? Would you know within a day if it went wrong? Could you put right anything it did, quickly and cheaply? An errand that passes all four can go to an agent, with a named person watching. An errand that fails any of them stays with a person, and AI assists.
Here is how it plays out in a hypothetical 25-person heating and plumbing business.
| Errand | One paragraph? | Named limits? | Know within a day? | Cheap to put right? | Verdict |
|---|---|---|---|---|---|
| Contacting customers due an annual boiler or furnace service and offering three slots | Yes | Yes: customer list and calendar only | Yes: customers reply or they do not | Yes: apologize and rebook | Agent, with the office manager checking the log daily |
| Paying supplier invoices under a set amount | Yes | Partly: it needs the bank account | No: errors surface at month end | No: money has left | AI prepares the payment run, a person approves it |
| Triaging emergency calls about a gas smell | No: judgment on every call | No | Too late | No | A person, every time |
Notice what decided each row. Not the cleverness of the AI, which could plausibly attempt all three. What decided it was reversibility and visibility: whether a mistake would show up in time and could be undone.
What I Tell Business Owners
What I tell business owners is that the agent question is a delegation question wearing a technology costume. I have trained more than 150,000 people in more than 30 countries to use AI at work, and the pattern I trust most is that the tool is rarely the hard part. The hard part is the brief, the limits and the check.
Across the business teams I work with, the first question about agents is almost always "which one should we buy?" The better question is "which errand would we give a sensible temp in their first week, and how would we know if they got it wrong?" If you cannot answer that for a person, an agent will not answer it for you.
The phrase that should make you pause is "don't worry, a human checks it." Before accepting it, I run the Loop Test.
The Loop Test is three questions I ask before accepting 'a human checks it' as an answer: Would that person notice a wrong result without being told to look for it? Would they notice in time to stop it mattering? And if they missed it, could the work be put back? Oversight that survives all three is a safeguard. Oversight that fails any of them is a signature.
Agents make the Loop Test harder to pass, because they work faster than anyone reads. The principle underneath all of it is the one I teach everywhere: outsource the doing, not the thinking. An agent can do the errand. Deciding what the errand is, and what good looks like, stays with you.
What should a small business do about AI agents this week?
This week, pick one errand, test it on paper and write down its limits before anyone buys or switches on anything. A 20-person business can do all of this in an afternoon.
- List five errands. Look for jobs that start with a trigger (an email arrives, a date passes, a form is filled in) and end with an action (a reply, a booking, an update). Chasing unpaid invoices, confirming appointments and sending quote follow-ups are typical.
- Run the Errand Test on each. Be strict about the last two questions. Most errands fail on "know within a day" or "cheap to put right", and that is useful information.
- Keep one errand that passes. Just one. If none pass, you have learned that your first move is an assistant helping a person, not an agent acting alone.
- Write its permission slip. A short list of what the agent may touch, what it must never touch, and what it must hand back to a person. If you already have an AI policy, add a line about agents to it; if you do not, start with the one-page version.
- Name the watcher. One person reads what the agent did, every day, for the first month. Put the time in their diary.
- Prepare four questions for any vendor. What can it touch? What does it record about what it did? How do I undo an action? Who is responsible when it gets something wrong? A vendor who cannot answer all four plainly is selling you a demo, not an employee.
When should a small business wait?
Wait on agents if your team is not yet using AI well as an assistant, because an agent automates habits, and it will automate bad ones as faithfully as good ones. If nobody in the business has a shared way of using AI on a real job, the first move is still one job, one owner, thirty days, not an agent.
Waiting also makes sense if the only errands you can think of involve payments, personal data about staff or customers, or decisions about people. Those are where the regulators are looking and where mistakes cost most. It is also worth being honest about the other trap: many businesses already pay for AI that nobody uses well, and an agent layered on top will freewheel the same way.
There is a counterargument, and it deserves its due. Some small firms, especially those with one overloaded owner doing every admin job, have the most to gain from a well-fenced agent and the least time to wait. For them the answer is not to wait. It is to start with the smallest errand that passes the test, and let the evidence, not the pitch, decide the next one.
Where to go from here
If you are planning next year and want your leadership team, your members or your conference audience to understand what agents will and will not change about work, that is what my keynotes for business audiences are built for. If you would rather start by finding out where your business stands before anything acts in its name, take the free Workplace AI Readiness Check: 21 questions, about eight minutes.
Sources and further reading
- ICO tech futures: Agentic AI, Information Commissioner's Office, January 8, 2026.
- New Research: UK small businesses want AI to win back time, not just write copy, Small Business Institute and Alibaba.com via PR Newswire UK, September 17, 2026.
- The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger and Aditya Pande, US Census Bureau Center for Economic Studies, April 2026.
- Announcing the "AI Agent Standards Initiative" for Interoperable and Secure Innovation, NIST, February 17, 2026.
- State of agentic AI adoption survey, Zapier (survey by Centiment), December 15, 2025.
Dan Fitzpatrick helps businesses use AI well through keynotes, practical AI training for teams, and AI strategy and governance for leadership teams.
