Skip to content

3questions

What to automate with AI first in a small business, and what to leave alone

What a small business should automate with AI first, shown on one invented plumbing company's list: eleven ideas, three questions that cross four of them off, and a first build chosen by its worst day. The questions we ask in an assessment, so you can answer them yourself; how the assessment runs and what you get; a week-by-week roadmap you can follow alone; what 2026 government surveys show small firms really do with AI; a one-page sheet; and a ten-question quiz to check what stuck.

Excelsior Mind26 min readUpdated

The first thing a small business should automate with AI is a job that happens every week, follows a rule you could write down for a new hire, and fails where a person sees it before a customer does. Below, one owner's list of eleven ideas goes through that test, and you will see why the smallest worst day goes first and why the winner may need no AI at all. It is the method we use in an AI assessment, which some firms call an AI audit, and you can run it on your own business. By the end you can test any idea with three questions, put what survives in a safe order, and follow a week-by-week roadmap to a first measured trial.

Maria's plumbing company (invented). Maria owns a plumbing company with four people: herself, Tom in the office and two plumbers on the road. She keeps hearing that businesses like hers are falling behind on AI, so one evening she writes down everything she would hand to it: eleven ideas. She does not know which to start with, or which would embarrass her in front of a customer. Maria, Tom and every figure about their company are made up; their list runs through this page as one worked example of a real method.

Everything on this page is free to use: the method, the eighteen questions we ask and a one-page sheet. What we sell is running it with you and building what survives.

In short

  • Three questions cross ideas off: often enough to be noticed, a rule a new hire can follow, a wrong answer a person sees first.
  • Rank what survives by its worst day, and keep a person on anything a customer sees.
  • About one firm in five uses AI inside a business function; close to one in two has someone using a tool somewhere.
  • The use is narrow: sales and marketing first, and most users run it in three or fewer functions.
  • It adds to the work people already do: 2.3 percent of AI-using firms added staff and 2.0 percent cut.

What an AI audit of a small business is for

An AI audit of a small business, which we call an AI assessment, is for one result: your list handed back shorter, with a reason beside each removal that you can apply again on your own. In a small business every idea that gets built becomes a job someone has to own, and a build that skips the assessment risks automating the wrong job, or the right job on a rule nobody had settled. Every idea crossed off is one less build to pay for, own and watch.

What you hold at the end

Shorter than the list you brought

  • A written account of how the work runs now

    Including the workaround somebody invented years ago that the office now depends on.

  • An order, with the reason beside each position

    So the order survives somebody asking why the third item is not the first.

  • The list of what you will not build

    The half nobody writes down. Its reasons are the part you use again.

  • A plan for the first build

    Picked by its worst day, with the person who approves it named.

Source: our method

Where to start with AI: one job written down

To start with AI in a small business, write one real job down first, in the order it happened, including the steps that exist only because a system does not do what it was bought to do. The workflow automation article on this site shows how, with a piece of paper.

Maria's plumbing company (invented). Tom writes down how a new customer gets booked. He takes the call and writes the details in the phone log, types them again into the quote, types them a third time into the invoice, and texts the plumber by hand. Four steps, three of them the same typing, about 25 times a week.

Your turn. Write down the job you did most often last week, tonight, the same way.

The three questions: is it worth automating at all?

Three questions decide whether an idea for AI is worth automating at all, and each one takes a different kind of idea off a list. Run each idea past them in order, and stop at the first no.

One: it happens often enough that a failure gets noticed within a week

A job that runs several times a week has someone waiting for its result, so a failed run is noticed within days. A job that runs four times a year can break in March and be found in June by an annoyed customer. Weekly or more is a yes when you can name the person who would miss a run; quarterly is a no.

Two: the rule can be written down for a new hire

When three people in an office describe the same decision three ways, nobody is wrong. The rule was never decided, and software built over the top freezes whichever version its builder heard. Ask the three separately, never in a meeting; if you work alone, write the rule down today and again next week without looking, and compare. Different answers mean deciding the rule comes before any build.

Three: a wrong answer lands where a person sees it first

A wrong figure in a sheet someone checks on Friday costs an hour. A wrong answer sent to a customer in writing costs the customer and then the business. An idea passes when a person reads the result before a customer does, or when the message is a fixed one whose exact words a person approved once. Better software does not change this question's answer, because where a mistake lands is a fact about the business.

The three questions, in order

Any one no is enough to cross an idea off

  1. Does it happen often enough that a failure gets noticed within a week?

    Yes, carry on

    If the answer is No, so not building

    Too rare for a break to be found

    A weekly job reports its own failure. A quarterly one hides it.

  2. Can the rule be written down for a new hire?

    Yes, carry on

    If the answer is No, so not building

    No settled rule yet

    Deciding the rule comes first. Ask three people, separately.

  3. Does a wrong answer land where a person sees it first?

    Yes, carry on

    If the answer is No, so not building

    It would land on a customer

    Better software does not change this answer.

  4. Rank what survives by its worst day

    The most contained failure goes first.

An invented example, not dataEleven ideas in, four crossed off at the three questions.

Source: our method; NIST AI 100-1, January 2023, checked current October 31, 2026Avoiding a risk is one of the four responses the framework lists, which makes not building a legitimate result.

Maria's plumbing company (invented). Maria runs her eleven ideas past the three questions. Four come off and seven stay. This is her whole list, an invented example, with what each question said:

  1. Update the price list each year. Off at question one: once a year.
  2. Mail the holiday cards. Off at question one: once a year.
  3. Decide which jobs get a discount. Off at question two: three people gave three rules.
  4. Answer customers' price questions by text. Off at question three: a wrong price would reach a customer in writing.
  5. A Friday list of open quotes. Stays: weekly, a plain rule, read only by Maria.
  6. Copy each new customer's details into the quote and the invoice. Stays: Tom reads both before they go out.
  7. Draft the follow-up to each open quote. Stays: Tom reads and sends every one.
  8. Send payment reminders. Stays: a fixed message approved once, sent by the program that holds the invoice.
  9. Send an appointment reminder the day before. Stays: a fixed message approved once.
  10. Draft replies to after-hours enquiries. Stays: a person approves each reply.
  11. Count missed calls each week. Stays: nobody outside the office sees it.

Your turn. Try your own numbers. From 12 ideas (an invented example), if 65 percent happen often enough, 70 percent of those have a writable rule and 60 percent of those land where a person sees them first, about 3 are left.

In your numbers

How many of your ideas are still standing at the end?

Write them on paper first, without judging any of them.

12

Weekly or more counts. Quarterly almost never does.

65%

If the honest answer starts with it depends, count it out.

70%

Not whether it would be wrong. Where it would land if it were.

60%

Ideas still standing at the end

3.3left

Ideas you can nameyour 12
12ideas
Happening often enoughyour 65%
7.8ideas
With a rule you could write downyour 70%
5.5ideas
Where a wrong answer lands somewhere safeyour 60%
3.3left

Shares produce fractions, so read anything with a decimal as a rough count. Move the last slider first: it usually moves the answer most. This counter puts no number on a failure rate, a payback period or the cost of a build, on purpose.

The federal AI Risk Management Framework from the National Institute of Standards and Technology agrees that not building is a legitimate result: it lists the responses to a risk as "mitigating, transferring, avoiding, or accepting", and it "does not prescribe risk tolerance". Version 1.0, from January 2023, was still current on October 31, 2026 (NIST AI 100-1; NIST, AI RMF page).

What should a small business not automate with AI?

A small business should not automate three kinds of work with AI: any fact a customer relies on, such as a price or a policy, unless the system reads it from the record that governs it; the judgment calls made a few times a year, such as taking on a client; and anything whose failure is silent, such as a reminder that stops sending without an error. Those three stay with a person in every assessment we run. A fourth test costs nothing: anything you would be embarrassed to say a machine did.

What stays with a person

Four kinds of work we leave alone

  • Facts a customer relies on

    Prices, policies, appointment rules: read from the record, or say there is none.

  • Rare judgment calls

    Taking on a client, changing a price, letting a customer go.

  • Anything whose failure is silent

    If nothing visibly goes wrong when it stops, nobody finds out for months.

  • Anything you would be embarrassed to explain

    The free test, applied before the other three.

Source: our method

A model asked about a price will answer, and the risk is a wrong answer in the right format that nobody checks. The right build is smaller: it fetches the value from the record that governs it and says so when there is no record. It may report what a source says or say it does not know.

Maria's plumbing company (invented). Answering price questions was the idea Maria wanted most, and it is exactly what the third question is for. If she builds it one day, the system will read each price from her price list and, where the list is silent, say that Tom will confirm. It will never write a price of its own, Tom approves each reply before it is sent, and he still confirms a quote before it goes out.

Your turn. Mark every idea on your list that states a fact to a customer, and ask where that fact is read from.

In a property business, specifically

We started in real estate, where square footage, a tax figure or a permit reads like data and behaves like liability, because a buyer may rely on it. The build worth doing reads each value from the listing record or the public tax roll and says when the record is silent. Whether to take a listing or advise a price cut stays with the agent.

Who should stay on the job

A person should stay on any job a customer sees. That is our rule, and it matches what firms report so far: among AI-using firms in the Census working paper, 44 percent said it added to tasks their workers already do. Headcount barely moved, with 2.3 percent of AI-using firms reporting an increase and 2.0 percent a decrease (Bonney and others, CES 26-25).

What AI did to the work

The person stays

Each bar is its own share of the firms using AI. One firm can be in more than one count.

Tasks

Added to tasks workers already do

44 percent

Took on a task an employee did before

about 10 percent

Brought in a new task

about 11 percent

Headcount

Any change in headcount

about 5 percent

Headcount up

2.3 percent

Headcount down

2.0 percent

Software and equipment

Replaced existing software or equipment with AI tools

16 percent

66 percent

of firms with any change to tasks saw only AI helping a person with their work.

Source: Bonney and others, working paper CES 26-25, Census Bureau, April 2026; read October 31, 2026

We build AI to add to what your people do: the person stays, and the system takes the repeated typing, copying and checking off their desk. The system drafts the reply, fills in the details and finds the past job that looks like this one; a person reads it and approves it. The step that sends cannot run without that approval, and that is built into the system rather than written in its instructions: it either has no way to send, or it checks an approval it cannot give itself.

A neighbourhood bakery counter: croissants, brioche and pastries on red plates with handwritten labels, coffee urns on the shelf behind and a worker at the back of the shop.
Every label on this counter was written by hand, by someone who knows the regulars. The reorders, invoices and reminders behind it are the repeated work worth a look. Photo: Infrogmation of New Orleans, CC BY 2.0.

Four checks decide who runs a job. Does the result bind the business, such as a price or a promised date? Does it reach a customer? Does it feed another system people act on, such as the books? Could it damage trust if wrong? One yes means a person approves every result, at least through a four-week trial; with two or more, the binding part stays with that person alone.

The four checks

One yes, and a person approves every result

We run every job past these four before deciding who runs it.

  • 1

    It binds the business

    A price, a contract, a payment, a promised date.

  • 2

    It reaches a customer

    Or anyone outside the office.

  • 3

    It feeds another system

    The schedule or the books that people act on.

  • 4

    It could damage trust if it is wrong

    The question to ask out loud, not to assume.

Source: our methodTwo or more yes answers: the binding part, such as setting a price or signing, stays with the person alone.

The one exception is a fixed message whose exact words you approved once, with only details read from the record filled in, such as the name, date and time of an appointment reminder. Check the record it reads from before you switch it on, and name who corrects that record when it is wrong. Anything a model writes to a customer is approved every time.

Maria's plumbing company (invented). Quote follow-ups pass all three questions, and they reach customers. So the system drafts each one, Tom reads it and sends it, and the price is never written by the system.

Your turn. For every idea that survives, name the person who approves it.

A mechanic in a repair shop leans over the open bonnet of a car, reading a diagnostic program on a laptop set on the engine bay.
The laptop reads the fault codes. The mechanic decides what the car needs, and tells the customer. Photo: Shixart1985 (Nenad Stojković), CC BY 2.0, via Wikimedia Commons.

Which AI idea should a small business build first?

A small business should build first the AI idea whose worst day is smallest, not the one with the biggest saving. Ranking by the average outcome gets it wrong because an average blends two different things: most jobs come in close to plan, a few go badly wrong, and the blended number describes neither. With a short list, the natural move is to estimate what each idea is worth and start at the top. We sort by how contained the worst case is.

Maria's plumbing company (invented). The obvious first build is the typing Tom found. Tom times it: about ten minutes a retyping, twice a booking, 25 bookings a week. Copying the details automatically would take about eight hours a week of typing off him, less the minutes he spends checking each record. When it goes wrong, one bad entry that slips past Tom lands in the quote, the invoice and the books at once. The Friday list of open quotes saves about two hours, and when it goes wrong a quote is missing from the list until Maria checks it against her records: some rework, and no customer has seen anything. Ranked by the saving, the copying wins. Ranked by the worst day, the Friday list goes first and the copying second, once the office has seen one system earn its trust.

Your turn. Write each idea's worst day beside it, and sort by that.

The failure rate nobody can give you

We found no 2026 public survey of how often a small business's automation fails. The most quoted project failure figure comes from a report sold by a private research firm whose data has never been opened to inspection. Two researchers at Vrije Universiteit Amsterdam applied its definitions to real project records and found they "result in meaningless figures" (Eveleens and Verhoef, IEEE Software, 2010): a project counted as late or over budget, never early or under. When anyone quotes you a failure rate or a saving, ask how it was measured.

What the order gets sorted by

First, containment: what happens when it is wrong, and who finds out. Second, whether the rule is settled, because software will not make a management decision. Third, and only third, the size of the saving, because the biggest saving is often the job with the most judgment in it.

Our plan also scores each job from 1 to 5 for impact and for risk, each with its reason. Impact is the hours a week the job takes: under one hour is a 1, from one to under two a 2, from two to under five a 3, from five to under ten a 4, ten or more a 5. Risk is what a wrong result costs: fixed in minutes inside the office is a 1 and some rework inside the office a 2; one yes on the four checks makes it at least a 3, two or more at least a 4, and a 5 binds the business and cannot be taken back. A job that takes under two hours a week is left alone, unless switching it on costs nothing, as with the reminder Maria finds below. On that scale Maria's Friday list (still the invented example) is impact 3 and risk 2: the digging takes her about two and a half hours a week, and a wrong list costs some rework inside the office. Her copying job is impact 4, about eight hours a week of typing, and risk 4, because it feeds the books and reaches customers, so Tom checks every record for as long as it runs.

Does every job on the list need AI?

Not every job on the list needs AI. A fixed step, such as copying a customer's details from one program to another, is ordinary automation: a plain rule that runs the same way every time and costs less to keep. AI earns its place where meaning matters, such as drafting a reply or reading a note someone typed in their own words. We build both, and our plan labels every step as one or the other.

From one job to the first build

Six steps, and the third does the most work

  1. The account

    One real job, written down as it ran.

  2. The ideas

    Every one you can name, unjudged.

  3. The cuts

    The three questions, one at a time.

  4. The order

    Worst day first, a settled rule second, the size of the saving last.

  5. The first build

    Small, visible when it fails, with a person approving.

  6. The re-read

    The same list, months later, because the business moved.

Source: our method

How do you run an AI audit of your own business?

To run an AI audit of your own business, start with one evening and five steps: list, ask, trace, cross off, then wait two weeks and read it again.

  1. Write every idea down without judging it, and beside each one how often the job happens.
  2. Ask the rule question to three people, separately.
  3. Before writing any benefit, write where a wrong answer would land and who would find out.
  4. Cross off every line that happens less than weekly, every line with no settled rule, and every line where words nobody approved would reach a customer, or a wrong answer would reach nobody.
  5. Put the list away for two weeks, then read it again.

One evening, then two weeks

Run the test yourself

  1. Write every idea down, without judging it

    And how often each job happens.

  2. Ask three people the rule question, separately

  3. Note where a wrong answer would land, and who would find out

    Before writing any benefit.

  4. Cross off the rare, the unsettled, and what would reach a customer unapproved, or nobody

  5. Read the list again in two weeks

What is left is usually short.

Source: our method

The free file

The three-question sheet: one page with the three columns ready and the five steps under them.

Download the sheet

A free printable PDF, one page. Nothing to sign up for.

Your AI roadmap, week by week

Nothing here needs us, and the method costs only your time until an idea has passed every test. Where no job repeats yet, writing the process down comes first. This roadmap ends at a first measured trial; building beyond it is a separate job.

  1. Week one. Write down one job, answer the questions and run the one-evening test.
  2. Weeks two and three. Let the list sit, decide the unsettled rules with the people who do the work, and check whether a program you already pay for does any job on it.
  3. Week four. Pick the idea with the smallest worst day and name its approver.
  4. Weeks five to eight. Before the first run, write five things on one page: where the result is read from, who approves it, how a missed run would be noticed, how to switch it off, and the hours the job takes today. Then run a four-week trial with a person approving every result, and count the hours again at the end. If it reaches a customer or binds the business, the person stays on it for good; the one exception is a fixed message whose words were approved once. If it does neither, and four weeks needed no corrections, it can run alone on your written say-so, with a log you can read and a named person who would notice a missed run within the week and reads the log every month. If it needs correcting most weeks, stop and settle the rule before touching the tool.
  5. Every few months. Read the list again.

On your own, week by week

From one written job to one measured result

  1. Week one

    One job written down, the questions answered, the evening test run.

  2. Weeks two and three

    The list rests. Rules get decided. Tools you already pay for get checked first.

  3. Week four

    One idea, the smallest worst day, and the name of the person who approves it.

  4. Weeks five to eight

    A four-week trial with a person approving every result, hours counted before and after.

  5. Every few months

    The list, read again.

Nothing on this path costs money until an idea has passed every test.

Source: our method

Maria's plumbing company (invented). In weeks two and three Maria, Tom and the senior plumber settle the discount rule at last, so that idea can be tested again at the next read. Maria also finds that the invoicing program she already pays for can send payment reminders. A reminder is a fixed message, the one exception named above: she approves its words once, sends one to herself to check the amount and the timing, and switches it on. That idea needed no build.

Maria's plumbing company (invented). In week four Maria picks the Friday list and names herself as its approver. It is small on purpose: every quote sent more than a week ago with no answer, read from the program the quotes live in, on one page in her inbox each Friday morning. No AI is needed for it. Before the trial she times her Friday digging at about two and a half hours. For four Fridays she checks the list against her own records in both directions, every line on it and every open quote it should hold, which takes half an hour. If no list arrives on a Friday morning, she knows the same day. By week eight she knows the list saves about two hours a week, measured rather than promised. Eleven ideas became seven: one was a switch she already owned, one runs and is measured, and the copying is next, with Tom checking every record.

Your turn. Decide now what the found hours are for. A found hour has to be spent on purpose, or it gets absorbed.

If you would rather go through your own list with a second pair of eyes, that is the assessment: the next three sections say how it runs, what we ask and what you get.

What is an AI readiness assessment and how does it run?

An AI readiness assessment is a one-hour call with you, or with you and your manager, and a written plan after it. We are based in New York, and the call works for an owner anywhere.

Why this is different from an automation agency

A tool-first pitch starts from what the software can do. We start from the person who does the work and the hours they get back, keep a person on anything a customer sees, and tell you in writing what not to build.

The steps, in order

Before the call. Your written answers to the questions below, and the four things listed after them. Bring your manager if someone else runs the admin: you know how the business should run, and they know how it does.

The call. One hour on the last real week, because "walk me through last Tuesday" finds the real jobs. We pitch nothing, name no tools and promise no savings.

The notes. Within two business days, sent back for you to correct.

The scoring. Every step labelled, impact and risk with reasons, and one verdict per job: let a system run it, keep a person on it, or leave it alone.

The plan. Built from the scoring, about three business days after the call, then a short call to walk you through it.

The route you pick. We build and run it, or you build from the plan. Either way it runs in your accounts.

If you ask us to run it

What happens, in order

  • One hour

    on the call with you, or you and your manager

  • Two business days

    for our notes to come back to you

  • One plan

    written for you, yours to build from

  1. Before the call

    Your written answers, the four things to send, and your manager if they run the admin.

  2. The call

    The last real week, walked through. Nothing pitched.

  3. The notes

    Sent back to you to correct, with one question: is anything missing?

  4. The scoring

    Every step labelled, impact and risk with reasons, one verdict per job.

  5. The plan

    What to build and in what order, how long each piece takes, the precautions, and which system does which job.

  6. The route you pick

    We build it and run it, with the assessment included, or you build it from the plan. You own it.

Source: our method

The questions we ask in an assessment

Answer these in writing, about the last real time something happened, before you talk to anyone: the answers are the first half of an assessment.

The business in five minutes

  1. What do you sell, to whom, and where? How many people work here, and who does what?
  2. What is a customer worth to you on average, and do they come back?
  3. How many new enquiries come in during a normal week, and from where?

Last week, walked through

  1. Take last Tuesday, from the first message to the last. What happened, and which of it happens every week?
  2. Walk one job from the first enquiry to getting paid. Who touches it, how long does each step take, and how often a week does it happen?
  3. When an enquiry comes in at 2pm on a Tuesday, what happens? And at 9pm on a Saturday?

The job list

  1. Which two of your programs do not talk to each other, and how does information get between them today?
  2. How many times does the same piece of information get typed into something, by anyone?
  3. For each repeated job: if you stopped doing it tomorrow, what would break, and who would notice?
  4. If two of your people hit the same case on the same day, would they handle it the same way?

The numbers

  1. Roughly what does an hour cost you for the person doing the most repetitive job, counting wage, taxes and benefits?
  2. How many calls do you miss in a week, and what happens to a missed call now?

What you want to build

  1. Which single task do you most resent doing?
  2. If you had 40 hours a week back across the business, what would you do with them?
  3. If this worked, which number would move?

What you would never hand to a machine

  1. Where do you not want a system involved at all?
  2. What has to be true before anything goes to a customer, and who approves that today?
  3. What is sensitive here, such as customer details, prices or staff information?

The questions we ask

Six groups, answered alone before any call

  • The business in five minutes

    Questions 1 to 3: what you sell, what a customer is worth, how many enquiries.

  • Last week, walked through

    Questions 4 to 6: one real day, one real job, one evening enquiry.

  • The job list

    Questions 7 to 10: double typing, and what would break if a job stopped.

  • The numbers

    Questions 11 and 12: what an hour costs, and the calls that go unanswered.

  • What you want to build

    Questions 13 to 15: the task you resent, the 40 hours, the number that moves.

  • Never handed to a machine

    Questions 16 to 18: your limits, your approvals, what is sensitive.

Source: our method

Before the call, send a week of your calendar, the tracker you run the work from, two real messages you sent customers, and your quote template. Black out customer names and anything sensitive first: we need the shape of the work, not the details. A good answer to question 8 sounds like Tom's invented one: "Every new customer is typed in three times: the phone log, the quote and the invoice. About 25 a week."

What do you get at the end of an AI assessment?

At the end of an AI assessment you get a written plan, made from what you told us: what we would build and in what order, how, how long each piece takes, and which system does which job. The plan names the precautions: what stays with a person, what is checked before anything reaches a customer, and how a failure shows. For every repeated job it gives the scores with reasons, the verdict, current hours and a four-week trial target, plus what we crossed off and why, and our open questions instead of guesses. Hours back are a trial target, never a promise, and every number says how it is known. Your time is the written answers before the call, the hour on the call and about another hour later for the notes and the walk-through.

Every job gets one verdict

Three verdicts, each with its reason

  • Let a system run it

    Fixed steps a system does the same way every time, with nothing reaching a customer.

  • Keep a person on it

    The system prepares, a person approves, and the send step waits for that approval.

  • Leave it alone

    Too small, too rare, or nobody would notice it stopping. The reason is written down.

Source: our method

Two routes from the plan

If you ask us to build it and run it, the assessment is included in that work, and only then do we talk with each person who does the jobs being built. If you build it yourself, the plan is your map.

Systems break: a program changes, a rule moves, and a job that ran for months stops without an error. A system needs proper evaluation and regular check-ins after it goes live. We have already made the mistakes a first-time builder makes, and the fixes are in the way we build. That is why we recommend a monthly retainer: we build it, watch it and fix it when a program around it changes, which is safer and more efficient than finding or training someone to learn those mistakes on your business.

What an assessment does not do

An AI assessment does not decide for you or see what nobody will say, and it goes stale within months, because your business moves.

How small businesses use AI in 2026

About one US business in five uses AI: the Census Bureau's Business Trends and Outlook Survey put AI use in any business function at 19.8 percent of US businesses as of May 3, 2026 (Census Bureau, May 26, 2026). The Federal Reserve's Small Business Credit Survey asked firms with 1 to 499 employees a wider question, whether the business or its employees use AI in some way, and 46 percent said yes, mostly for writing and marketing (Federal Reserve Banks, March 3, 2026; 6,525 responses, 5,248 of them to the optional AI questions, fielded September 3 to November 14, 2025).

Different questions, different counts

One firm in five, or close to one in two

Three surveys asked three different questions. Read each panel on its own.

Business Trends and Outlook Survey, as of May 3, 2026

Used AI in any business function, all US businesses

19.8 percent

Census working paper, November 2025 to January 2026

Used AI in a business function

18 percent

Small Business Credit Survey, fielded September 3 to November 14, 2025

Business or employees use AI in some wayFirms with 1 to 499 employees

46 percent

Source: Census Bureau, May 26, 2026; Bonney and others, working paper CES 26-25, Census Bureau, April 2026; Federal Reserve Banks, March 3, 2026, 6,525 responses; read October 31, 2026The three counts answer different questions, so they are not one series and they do not add up.

A Census Bureau working paper on the survey's 2026 AI supplement, reference period November 2025 to January 2026, splits use by size (Bonney and others, Census Bureau working paper CES 26-25, April 2026). Firms with 1 to 4 employees were at 18 percent; firms with 250 or more were at 31 percent and expected 40 within six months.

Use of AI by firm size, now and in six months

The bigger the firm, the bigger the jump it plans

One survey and one reference period, November 2025 to January 2026.

1 to 4 employees

Using AI

18 percent

Expected within six months

about 21 percent

100 to 249 employees

Using AI

25 percent

Expected within six months

34 percent

250 or more employees

Using AI

31 percent

Expected within six months

40 percent

Source: Bonney and others, working paper CES 26-25, Census Bureau, April 2026; read October 31, 2026Solid bars are measured use. Outlines are what the firms expected, not a measurement.

The trade matters as much: as of May 3, 2026, the Information sector was at 39.7 percent and Retail Trade at around 14 percent (Census Bureau, May 26, 2026).

Use of AI by trade, as of May 3, 2026

Information and finance lead, retail trails

Information

39.7 percent

Finance and Insurance

33.9 percent

All businesses

19.8 percent

Retail Trade

around 14 percent

Source: Census Bureau, May 26, 2026; read October 31, 2026Shares of the businesses in each sector, from one survey reading.

The use is narrow. Among AI-using firms the working paper found sales and marketing the most common use (52 percent of users), and 85 percent of firms whose workers use generative AI used it for writing and editing (Bonney and others, CES 26-25).

Among firms using AI

Sales and marketing first

Sales and marketing

52 percent

Strategy and business development

45 percent

IT

41 percent

Research and development

40 percent

57 percent

of the users had AI in three or fewer functions; 24 percent in only one.

Source: Bonney and others, working paper CES 26-25, Census Bureau, April 2026; read October 31, 2026Shares of the firms using AI, not of all firms. One firm can use it in several functions, so the bars do not add up.

Most small businesses have not started

With four firms in five not using AI inside their work (Census Bureau, May 26, 2026), few owners have a neighbour to copy, so ask any confident advice, this page included, what it rests on. Of the firms with no plans, 65 percent said AI is not applicable to the business (Bonney and others, CES 26-25). Count the retyping in your own business before you agree with them.

Among firms with no plans for AI

Most say it does not apply to them

AI is not applicable to the business

65 percent

Not knowing what AI can do

22 percent

Concerns about privacy or security

20 percent

Source: Bonney and others, working paper CES 26-25, Census Bureau, April 2026; read October 31, 2026A firm could give more than one reason, so the bars do not add up.

Common questions about using AI in a small business

How do I use AI in my business?

To use AI in a small business, write one job down, list your ideas, cross them off with three questions, and build the survivor with the smallest worst day. The method costs nothing to run alone; the crossing off is harder alone, because every idea on your list is there for a reason you agree with.

What is an AI audit for a small business?

An AI audit for a small business is a structured look at how the business runs, to decide which repeated jobs AI takes first and which stay with a person.

What can I automate in my business first?

The first thing to automate in a small business should be something plain that passes all three questions and fails where you can see it: a confirmation, a reminder, a copy from one system to another.

How long does an AI assessment take?

An AI assessment takes one hour on the call with you, and the written plan follows about three business days later.

Do I need to know anything about AI first?

No, an owner needs no AI knowledge before an AI assessment. The three questions are about the business. Choosing the tools is our job.

Does using AI mean fewer staff?

Using AI has not meant fewer staff at most firms so far: among AI-using firms in the Census Bureau's 2026 working paper, headcount went down at 2.0 percent and up at 2.3 percent (Bonney and others, CES 26-25).

Does this method work for my kind of business?

The method works for any small business with office work that repeats: a plumbing or HVAC company, a dental office, a salon, a restaurant, an accounting or insurance office, a real estate team. It fits less where no job repeats yet; there, writing the process down comes first.

Ten questions

What did you learn? Check it in two minutes

Ten questions on what this page taught. It takes about two minutes.

Take the quiz
  1. 1 of 10

    In the Census Bureau's survey, about what share of US businesses used AI in a business function as of May 3, 2026?

    • A.About 1 in 5
    • B.About 1 in 2
    • C.About 3 in 4
    • D.Almost all of them
    Show the answer

    A. About 1 in 5 It was 19.8 percent, from the section on how small businesses use AI in 2026. The 46 percent figure answers a wider question.

  2. 2 of 10

    Which of these is the first of the three questions?

    • A.Is it the biggest time saving on the list?
    • B.Does it happen often enough that a failure gets noticed within a week?
    • C.Is there an AI tool that already does it?
    • D.Would it impress customers?
    Show the answer

    B. Does it happen often enough that a failure gets noticed within a week? How often a job runs decides whether anyone notices when it breaks. That is the first question.

  3. 3 of 10

    Three people in your office describe the same decision three different ways. What does that tell you?

    • A.One of them is doing it wrong
    • B.The rule was never decided, so deciding it comes before any build
    • C.The job is ready to automate
    • D.You need more staff
    Show the answer

    B. The rule was never decided, so deciding it comes before any build From the second question: software freezes whichever version its builder heard, so the rule gets decided first.

  4. 4 of 10

    An idea would send a wrong answer to a customer in writing before any person reads it. What does the third question say?

    • A.Build it with a better model
    • B.Cross it off, or put a person's approval in front of it
    • C.Send it and check a sample later
    • D.Ask the customer to confirm the answer
    Show the answer

    B. Cross it off, or put a person's approval in front of it From the third question: where a mistake lands is a fact about your business, so better software does not change the answer.

  5. 5 of 10

    Among AI-using firms in the 2026 Census paper, what share reported a decrease in headcount?

    • A.About 2 percent
    • B.About 20 percent
    • C.About 44 percent
    • D.About 66 percent
    Show the answer

    A. About 2 percent It was 2.0 percent down and 2.3 percent up. The common effect was AI adding to tasks people already do.

  6. 6 of 10

    What do we sort a shortlist by first?

    • A.The biggest saving
    • B.How contained the worst case is
    • C.Which idea is newest
    • D.Which tool is cheapest
    Show the answer

    B. How contained the worst case is From the section on what the order gets sorted by: containment first, a settled rule second, the size of the saving last.

  7. 7 of 10

    What is the right build for a fact a customer relies on, such as a price or a policy?

    • A.Let the model write its best guess
    • B.Read it from the record that governs it, and say so when there is no record
    • C.Leave the answer blank
    • D.Ask the customer to check it
    Show the answer

    B. Read it from the record that governs it, and say so when there is no record From the section on what a small business should not automate with AI: a system may report what a source says, or say it does not know.

  8. 8 of 10

    In our builds, what has to happen before a message a model wrote reaches a customer?

    • A.Nothing, if the draft looks good
    • B.A person approves it, and the send step cannot run without that approval
    • C.The customer agrees once, up front
    • D.It goes out, and a person checks it later
    Show the answer

    B. A person approves it, and the send step cannot run without that approval From the section on who stays on the job: the approval is built into the system, never just written in its instructions.

  9. 9 of 10

    What happens on the assessment call?

    • A.We demonstrate tools and quote a price
    • B.We walk through your last real week and pitch nothing
    • C.We build the first automation live
    • D.We review your website
    Show the answer

    B. We walk through your last real week and pitch nothing From the section on what an AI readiness assessment is and how it runs: one hour on the last real week, with no pitch, no tools named and no savings promised.

  10. 10 of 10

    Why does this page give no failure rate for AI projects?

    • A.Because the rate is zero
    • B.No 2026 public survey measures it for small businesses, and the most quoted figure cannot be checked
    • C.It is a trade secret
    • D.Every project works with the right tool
    Show the answer

    B. No 2026 public survey measures it for small businesses, and the most quoted figure cannot be checked From the section on the failure rate: when anyone quotes a failure rate or a saving, ask how it was measured.

Next step: book an AI assessment

The offer is the one-hour assessment, because the crossing off is the hard part to do alone: we go through your list with you, then send back a written plan of what we would build first, how, and what we would cross off. On the call we pitch nothing, name no tools and promise no savings, and you correct our notes before the plan is written. Book a Consultation: (914) 506-5884, hello@excelsiormind.com.

Sources

  • U.S. Census Bureau, Adam Grundy, Cory Breaux and Dhanapati Khatiwoda, "Large Firms With at Least 20 Employees Biggest AI Users", America Counts story, May 26, 2026. Business Trends and Outlook Survey, data from December 14, 2025 to May 3, 2026. Read October 31, 2026.
  • Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger and Aditya Pande, "The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks", U.S. Census Bureau, Center for Economic Studies working paper CES 26-25, April 2026. The 2026 AI supplement to the Business Trends and Outlook Survey, reference period November 2025 to January 2026; all shares firm-weighted. Read October 31, 2026.
  • Federal Reserve Banks, Small Business Credit Survey, 2026 Report on Employer Firms, press release of the Federal Reserve Bank of Cleveland, March 3, 2026. Fielded September 3 to November 14, 2025; 6,525 responses from firms with 1 to 499 employees, 5,248 of them to the optional questions on AI. Read October 31, 2026.
  • National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023, sections 1.2.2 and MANAGE 1.1 and 1.3; NIST's framework page, read October 31, 2026 (version 1.0 current, a revision in progress). An older source, kept for the argument: it is the framework in force.
  • J. Laurenz Eveleens and Chris Verhoef, "The Rise and Fall of the Chaos Report Figures", IEEE Software, January/February 2010. An older source, kept for the argument: cited for how a failure rate was defined, not for any rate. Read October 31, 2026.
  • A search for a 2026 public survey measuring how often small-business automation fails returned none on October 29, 2026.

Sources checked October 31, 2026.

Written with an AI assistant and checked by a person.

Know somebody who would argue with this? Send it to them.