Replace an employee with AI only after you have written down what that person actually did. In most small businesses AI can take over parts of a job, not the whole seat, so the usual right move is to refill a redesigned role or hold the seat for 90 days before you decide. The Backfill Test below gives you four questions to answer first.
Someone resigns, and within a day somebody asks whether you could manage without them. It is a more common AI decision than any strategy document, and year-end makes it sharper, because many owners set next year's headcount in the coming weeks. This article covers what the evidence says is happening, a test for the decision, and a worked example.
Are small businesses actually replacing staff with AI?
Mostly not yet. The data show AI changing what people do in a seat faster than it empties seats. Three recent surveys, from the UK and the US, make the point from different angles.
In the UK, the British Chambers of Commerce surveyed 668 businesses, most of them SMEs, in early 2026 and published the results in March 2026. Among firms using only generic AI tools, 95% said AI had no impact on their headcount over the past 12 months. Among firms using bespoke AI rather than generic tools, 21% reported a headcount decrease, against 3% of the generic-only group.
In the US, Will Aarons and Asani Sarkar at the New York Fed analyzed the 2025 Small Business Credit Survey on October 8, 2026. The survey was fielded from September to November 2025, and its AI module covered 5,248 US employer firms. Of those, 46% used AI. Among firms with at least one employee, AI users' net expectation of higher employment over the next 12 months was 33 percentage points, against 15 for non-users, and the gap was still 13 points after controls for firm and owner characteristics. (Net expectation means the share expecting an increase minus the share expecting a decrease.) The authors describe these as associations and leave open whether firms achieve the growth they expect. They also found that 77% of owners reported no change in their labor costs from AI use.
The third survey looks at the decision itself. Ipsos polled 750 US small business owners and operators for the U.S. Chamber of Commerce Foundation between May 19 and June 4, 2026, with results published on July 14, 2026. The poll covered companies with up to 500 employees and excluded sole proprietorships. Among AI users, 73% said AI had affected employees' roles and responsibilities, and 65% said it had influenced hiring and personnel decisions. The St. Louis Fed's July 2026 survey of Eighth District businesses adds that 49% expect no noticeable staffing effect over the next 12 months, nearly 20% expect a slight reduction in staffing needs, and about 18% expect skill shifts rather than headcount changes. That bank does not publish how many firms answered.
Put together, these point one way. AI is reshaping jobs, and it is already shaping who owners hire. That reading is mine, not the surveys'. None of them tests whether a particular replacement decision turned out well, so the decision still rests with you.
What is the Backfill Test?
The Backfill Test is four questions an owner answers in writing before replacing someone who has left or advertising a new role: What did this person actually do each week? Which of those tasks can AI now do while a person checks the result? What can only a person in this seat own? And who will do the work that is left, and what will they stop doing? Write the answers down before you write the job advert.
I offer it as my suggested way of thinking, not as research. It exists because the decision usually runs in the wrong order. The gap appears, the pressure to act builds, and the description of the work comes last, if it comes at all. The test puts the description first.
How do you answer the four questions?
Each question has a method, and each one catches a different mistake.
1. What did this person actually do each week? Build the list from evidence, not memory. Use last month's calendar, their sent mail, the shared folders they touched, and a conversation with the person leaving and the colleague who relied on them most. Aim for 10 to 20 tasks. Do not start from the job description, because descriptions drift, as rewriting job descriptions for AI explains with the Core and Margin Rule. Write the before picture down now. My article on measuring whether AI is working makes the same case: you cannot show a change you never recorded.
2. Which of those tasks can AI now do while a person checks the result? Take each task and apply 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. A task that passes can move to AI with a named checker. A task that fails stays with a person, or needs a better check first.
3. What can only a person in this seat own? Look for the tasks where someone must answer to a customer, a colleague or a regulator, and for judgment calls with consequences. My governing principle is to outsource the doing, not the thinking. Anything that needs a person to stand behind it belongs to a person.
4. Who will do the work that is left, and what will they stop doing? Work does not vanish when a role does. It moves. If the AI-assisted tasks land on existing staff, name them, and name what comes off their plates in exchange. For the first month, check how it is going with the weekly routine in how to manage a team that uses AI.
What does the test look like on a real vacancy?
Take a hypothetical 16-person property management firm whose tenant-services coordinator has resigned. The owner is tempted to let AI cover the seat.
On question one, the owner and the departing coordinator list six tasks: sorting the shared tenant inbox, drafting repair updates, chasing contractors by phone, updating the maintenance log, calming tenants who have waited too long, and assembling the monthly landlord report.
On question two, the owner runs the Loop Test. Drafting repair updates passes: the person sending reads each one against the job record, and a wrong update is easy to correct. Sorting the inbox passes too. The maintenance log fails: a wrong entry would sit unnoticed until a contractor turned up at the wrong flat, so it stays manual with a named owner. The landlord report passes if a property manager checks every figure against the source.
On question three, the contractor calls and the upset tenants stay with a person. Someone has to be accountable to the landlord, and tenants who are angry want a human.
On question four, the two property managers take over the report check and the log, and stop writing the monthly report from scratch. The owner writes that down, with names.
The outcome is neither "replace with AI" nor "refill as it was." The owner advertises a smaller role built around contractor coordination and tenant relationships, with the inbox and drafting handled by the team using AI, and puts a review date in the diary 90 days after the new hire starts. The newcomer's first month is its own job, and AI training for new hires covers it.
What are your options once you have the answers?
Your four answers point to one of four moves. This table matches the pattern in your answers to the move and the thing to watch.
| What your answers show | Your move | What to watch |
|---|---|---|
| Most of the week still needs a person, and AI covers a minority of tasks | Refill, redesigned: advertise the new shape of the role, not the old description | The job description drifting back to the old one |
| The seat is where people learn the business | Refill, and keep the learning: AI drafts, the junior checks, a named senior mentors | Losing the people who would become your experienced staff |
| You cannot yet tell | Hold the seat for 90 days with a review date in the diary | A hold with no date is a freeze. Watch for errors and strain on the people covering |
| Most of the work was low-value, or passes the Loop Test with a real checker | Close the seat and redistribute the work, in writing | Work piling onto one person. Get employment advice before changing anyone's contract, hours or duties, because the rules differ by country and state |
A hold is not the same as a decision to leave the seat empty. Its job is to give you 90 days of evidence about what breaks, so the review is based on what happened rather than what you feared.
What if the seat is a junior one?
Then be slower to close it. Every junior role does two jobs: the work, and the development of the next experienced person. AI can take much of the first, and nothing replaces the second. If you are weighing a junior vacancy, read what AI does to entry-level jobs in a small business before you decide.
What I Tell Business Owners
When an owner tells me someone is leaving, the first thing I hear is almost always "can we manage without them?" My answer is that you might, but not yet, because you have not described the work. The mistake I see most often is deciding in the exit conversation, when the loss is fresh and nobody has the list. Across the business teams I train, the seats that go wrong are rarely the ones AI did badly. They are the ones nobody described before they were filled or left empty.
I have advised governments and system leaders, including the UK Department for Education, KHDA in Dubai and the Ministry of Education in Kazakhstan. It is a different world from a 16-person firm, but the order of operations is the same: describe the work, then decide who or what does it.
What should you do this week?
You do not need a vacancy to start. If someone has just resigned, do this before the advert goes up. If nobody has, pick the role you would find hardest to lose and ask what you would do if they resigned on Friday.
- Monday: Hold the advert. Tell the leaver and one close colleague you want an hour with them.
- Tuesday: Build the list of 10 to 20 tasks from the calendar, sent mail and shared folders.
- Wednesday: Run the Loop Test on each task and mark it pass or fail.
- Thursday: Write one sentence for question three and one for question four, with a named owner for each task that moves.
- Friday: Pick a move from the table and write the review date next to it.
If you keep the seat but your people are unsure how to use AI in it, start with what AI training your team needs.
Where to go next
If the outcome of your test is that your people need to use AI well in the seats you keep, that is what my AI training for teams is built for. If you would rather see where your business stands first, the Business AI Readiness Scorecard takes about five minutes. And if you are planning a leadership away day or an event on how AI changes the workforce, my business keynotes are written for that audience.
Sources and further reading
- British Chambers of Commerce, "Powering Productivity: AI and the Future of UK Work", March 2026 (668 UK businesses, most of them SMEs; early 2026 fieldwork): britishchambers.org.uk
- Will Aarons and Asani Sarkar, "AI Adoption and Employment Expectations: Evidence from a Survey of Small Business Owners", Federal Reserve Bank of New York, Liberty Street Economics, October 8, 2026 (2025 Small Business Credit Survey; 5,248 US employer firms in the AI module): libertystreeteconomics.newyorkfed.org
- U.S. Chamber of Commerce Foundation with Ipsos, "Small Business Owners Say AI Is Already Changing Work", July 14, 2026 (750 US small business owners and operators; fieldwork May 19 to June 4, 2026): uschamberfoundation.org
- Violeta Gutkowski, Charles S. Gascon and Rehann Silvanus, "New Survey Findings on AI Adoption and Its Effects on Employment and Productivity", Federal Reserve Bank of St. Louis, July 16, 2026 (Eighth District businesses): stlouisfed.org
- Dan Fitzpatrick, AI for business teams, The AI Educator
Dan Fitzpatrick, The AI Educator, is a Forbes contributor, international keynote speaker and bestselling author who has trained more than 150,000 people across 30+ countries. About Dan
