AI Fluency for Small Business Owners: Five Skills to Self-Check

AI fluency for a small business owner means five things you can do on real work: hand the right tasks to AI, describe them clearly, check what comes back, keep private data out of the wrong tool, and stay answerable for the result. Yours is good enough when you score at least three out of four on each.

Most people test the wrong thing. They ask whether they know the tools, when the better question is what happens on a Tuesday when a customer's complaint, a supplier's price list or a draft contract lands on the desk. Courses can teach the vocabulary; only a scored task tells you how you behave under real conditions.

Follow me on Instagram@sagnikteaches

Anthropic publishes a free course called AI Fluency: Framework and Foundations, and its page names four competencies: Delegation, Description, Discernment and Diligence. I've kept those ideas but split the last one in two, because a small firm needs separate answers to "is my data safe?" and "who is responsible?". That gives five skills, each tested with one task from your own week. Read the course page yourself if you want the formal version; the five tasks below are the practical test.

Connect on LinkedInSagnik Bhattacharya

You need an hour, a chat assistant you already have access to, and a few real (or realistic) pieces of work. Score honestly. The scorecard only helps if a low number is allowed to stand.

Subscribe on YouTube@sagnikteaches

How the five-task self-check works

Each skill gets one task, one pass mark and a four-point rubric. You work the task, then score yourself using the descriptions, not your feelings. A score of 4 means a stranger could watch you and agree. A score of 1 means the habit does not exist yet.

Do the tasks in order. Skill one decides which work you carry into the rest, and the later tasks reuse the same material, so you finish with a coherent picture rather than five unrelated quizzes.

Skill 1: delegating, or knowing which jobs to hand over

Delegating is a judgement about the job, not about the tool. Some work suits AI because it is wordy, repetitive and easy to check. Some suits it badly because the answer must be exactly right, or because a person's trust is the point of the job.

The task: three piles from last week

Open your calendar and inbox for last week and write down twelve jobs you or your team did. Include the dull ones. Then place each in one of three piles: hand it over, share it, or keep it. "Share" means the AI drafts and a person finishes.

Here is an illustrative run for the owner of a six-person veterinary practice.

Score yourself like this. If your "hand over" pile contains anything where a wrong answer could hurt a customer or the books, deduct a point. If your "keep" pile contains only work you simply dislike, deduct another, because dislike is not a reason; risk and trust are.

• 1 point: you could not fill twelve jobs, or every job went in one pile.
• 2 points: piles exist but you gave no reasons, or at least one risky job sits in "hand over".
• 3 points: reasons given for every job, the risky jobs are in "share" or "keep", and you can say what would move one job to another pile.
• 4 points: as above, and you have also noticed one job that nobody does at the moment but which AI could take on.

If this task felt hard, the AI skills matrix covers the same ground for a whole team.

Skill 2: describing the job so the answer is usable

Describing is where most owners lose time. A thin request produces a generic draft, which you then rewrite by hand, and you decide AI "isn't that useful". A good brief is short but complete: who it is for, what facts to use, what tone, how long, and what to avoid.

The task: brief the same job twice

Pick one customer message you would normally write yourself. Brief the assistant twice. First give it the lazy version, then the full version. Compare.

An illustrative case from an optician. The message is a polite reminder that a customer's eye test is due.

Lazy brief:
Write a reminder about an eye test.

Full brief:
You are writing for a small independent optician. Write a text message
of no more than 320 characters to a customer whose last eye test was
about two years ago. Say that a test is due and that booking takes
two minutes at our desk or by phone. Do not mention prices, frames
or offers. Friendly, plain, no exclamation marks. Sign off with
"[your name], at the practice". Leave a gap where I can insert the
customer's first name.

The lazy brief typically returns a long, bland paragraph with three exclamation marks and a made-up offer. The full brief returns something close to this (illustrative):

Hi [first name], it has been about two years since your last eye
test, so one is due. You can book in two minutes by calling us or
popping in. Hope to see you soon. [your name], at the practice

Check what remains to fix. It is about 190 characters before the name goes in, well inside the limit. It has no invented offer. It says "popping in", which may be too casual for your regular customers, so you would change that one phrase and add it to your standing instructions so you never type it again.

That last step is the difference between a 3 and a 4. A person who is describing well turns repeated corrections into a saved instruction (a project, a custom instruction or a pasted style note), so the next brief starts further along.

• 1 point: the full brief is only slightly longer than the lazy one, or you cannot say what changed in the output.
• 2 points: you named the audience and the topic but not length, tone or exclusions.
• 3 points: audience, facts, tone, length and exclusions all present, and the draft needed only light edits.
• 4 points: as a 3, and you saved the corrections as a reusable instruction.

If you want ready-made phrasing for common jobs, the staff communications prompts show how a full brief reads when the audience is your own team.

Skill 3: checking what comes back

An assistant can write a smooth paragraph around a wrong figure. Fluency here is a habit, not a talent: you look for the kinds of error AI makes before you look at the style. The usual ones are an invented fact, a number changed in copying, a promise you never made, a confident tone over a thin source, and a missing exception.

The task: find the planted mistakes

You need a short draft with errors in it, and the easiest way is to make your own. Take a real, correct message you have sent. Ask the assistant to rewrite it, then edit the result so that it contains three mistakes of different kinds. Better still, ask a colleague to plant them without telling you where. Then review it as if it were about to go out.

Here is an illustrative example for a music school. The original correct message told parents that termly fees of $210 are due by the first lesson, with a $15 late charge after two weeks. The planted version reads:

Dear parents, termly fees of $240 are due before the first lesson of
term. A $15 late charge applies after one week. All lessons are
guaranteed to be made up if your child misses one.

There are three mistakes: the figure ($240, not $210), the grace period (one week, not two), and a guarantee the school never gave. A person scoring a 4 finds all three and explains the method. A good method is to read the numbers first, then every promise or deadline, then the tone last.

1. Strike through every figure and date and confirm each one against the source, never against the draft itself.
2. Underline every promise, guarantee or "always" and ask: did we actually say that?
3. Look for anything that is missing: the exception, the opt-out, the contact route.
4. Only then read for tone.

The checker should also learn the assistant's habits. After a month of use you will know whether yours tends to round figures, invent examples or soften refusals. Keep a short list of "this assistant tends to…" notes beside your desk; it makes the review quicker.

• 1 point: you found none or one of the three, or you read for style first.
• 2 points: you found two, mostly by luck rather than by method.
• 3 points: you found all three using figures, promises, gaps in that order.
• 4 points: as a 3, and you keep a written list of your assistant's usual errors.

A related exercise sits in the 20-minute practical test for staff, which you can hand to a new hire once you pass this one yourself.

Skill 4: using data safely

This skill is about what you paste, and into what. The first question is the plan you are on. Consumer plans of ChatGPT and Claude have a privacy setting that decides whether your chats help improve the models (ChatGPT calls it "Improve the model for everyone" and switches it on by default on Free, Plus and Pro; Claude calls it "Model Improvement"). Business plans such as ChatGPT Business, Claude Team and Microsoft 365 Copilot (whose app Microsoft now calls Microsoft Copilot) do not train on your content by default. Check the current setting on your own account instead of trusting memory; the labels change.

The second question is the content. The simplest rule is minimum necessary: give the assistant only what the task needs. A reply to a complaint needs the facts of the complaint, not the customer's full address, date of birth and account number.

The task: strip one document to the minimum

Take a real document you might want help with and reduce it by hand. A care agency owner might choose a family complaint letter. The original contains the relative's name, the client's name, a care worker's name, a street address and a phone number, plus a description of three visits.

The stripped version, which is all the assistant needs to suggest a reply, looks like this (illustrative):

A relative complains that on three visits last week the carer arrived
15 to 25 minutes late, left before the agreed 45 minutes finished,
and did not record the lunch as eaten. Draft a calm reply that
acknowledges each point, says we are looking into it, and promises
a call within two working days. Do not admit fault yet.

Names are gone. The address is gone. The phone number is gone. The facts that shape the answer remain. When the reply comes back, you add the names yourself in your own document, outside the assistant.

Three more habits belong in this skill.

• Know which plan each team member uses. A free personal account on a staff phone is a different risk from a company plan.
• Keep a short "never paste" list: card numbers, passwords, health details, anything under a confidentiality agreement.
• Do not share logins. Anthropic's consumer terms forbid sharing login details, and sharing one account across staff also removes any record of who asked what.

The simplest way to give staff the same rules is a traffic-light list; the traffic-light AI policy shows how to build one in an afternoon.

• 1 point: you could not say what plan you or your staff use, or the stripped version still held identifiers.
• 2 points: identifiers removed, but you never checked the plan's privacy setting.
• 3 points: identifiers removed, plan and setting checked, and a never-paste list exists in your head.
• 4 points: as a 3, with the never-paste list written down and given to anyone who uses AI for the business.

Skill 5: staying accountable for the result

Accountability means that when an AI-assisted email, quote or notice goes out under your business's name, it is yours. You cannot say "the assistant wrote it" to a customer, a regulator or a court. Fluency here is an agreement, written down, about who looks at what before it leaves.

The task: four lines for one output

Choose one AI-assisted output you produce regularly, such as quote follow-ups, a monthly newsletter or customer replies. Write four lines:

1. What the output is and who receives it.
2. Who may approve it, by name or role.
3. The check they must run (for instance: figures against the price list, promises against the terms).
4. Where the record lives: a folder, a ticket or a sent-items label.

A filled-in example for a music school's monthly newsletter to families:

1. Monthly newsletter to about 300 families, drafted with AI.
2. Approved by the owner, or the office manager if the owner is away.
3. Check: fees and term dates against the fee schedule; no discount
   not in the offers sheet; unsubscribe link present.
4. Final text saved in the "Newsletters sent" folder, same day.

Those four lines take minutes and settle most arguments later. If a parent complains about a fee in the email, you know who checked it and where the sent copy is.

A subtler part of accountability is disclosure. Some customers, contracts or professional bodies expect you to say when AI helped. Rules differ by industry and by the places you serve, so ask your adviser or professional body for your sector's position instead of guessing; no general guide can settle it for you.

• 1 point: no named approver, or the approver is "everyone".
• 2 points: a named approver but no defined check.
• 3 points: approver and check both defined for the output.
• 4 points: approver, check and record, and you have asked what disclosure your sector expects.

A filled-in scorecard, with the arithmetic

Here is how the same veterinary practice owner from skill one scored. These figures are illustrative.

The total matters less than the shape. An 11 with a 1 in checking is more dangerous than a 9 spread evenly, because checking is the safety net under everything else. Here is how I'd read the bands.

• 16 to 20: fluent for the work you tested. Widen the range of tasks and start giving staff the same test.
• 11 to 15: workable. Fix every skill scored below 3 before adding new uses.
• 6 to 10: start with low-stakes drafting only and nothing that reaches a customer unreviewed.
• 5 or below: go back to skill one and practise on invented material for a fortnight before using real work.

For the vet practice owner, the plan from 11 was short: three planted-mistake drills in a week, a saved instruction for the reminder messages, and a four-line sign-off for the monthly newsletter. Nothing needed buying.

When a high score is misleading

A self-check can flatter you, and the way it does so is worth knowing about. The commonest cause is scoring yourself on a task you picked because it goes well. If every task is a friendly email to a regular customer, you will pass skill two and three easily and learn nothing.

The second cause is the tidy output. A draft that reads well feels correct, and a fluent paragraph lowers your guard on the figure inside it. Picture an owner, purely as an illustration, who scores himself 4 on checking because the draft "looked right", and only finds the wrong delivery date when a customer rings. His method was to read for tone first, which is the first thing to change.

The third cause is testing with the wrong material. Practising with invented customers is fine for skills two and three, but skill four only counts if you handled a document that looked like the real thing.

Protect against all three by choosing at least one task that is awkward: a complaint, a figure-heavy quote, a message to a supplier you are in dispute with. Pass those and the easy ones will look after themselves. For a harder second opinion, hand your scorecard to someone who works with you and let them score the evidence; they will usually spot the gap you are protecting.

Turning low scores into a week of practice

You can lift any single skill in about a week of short practice, provided you aim at one skill at a time. Here is a plan keyed to the lowest score.

If delegating is your weak spot

Keep a "could AI have done this?" tally for one week. At the end of each day write the jobs you did and mark them hand over, share or keep. After five days patterns appear: the same three jobs keep turning up in "hand over". Start with one of those.

If describing is your weak spot

Write a brief template once. It needs six lines: audience, purpose, facts to use, tone, length, and exclusions. Paste the template into each new request and fill it in. After ten uses you will not need it written out, which is the aim.

If checking is your weak spot

Do the planted-mistake drill daily for five days, with a fresh draft each time. Time yourself. Most people find that thirty seconds spent on figures first catches more than three minutes of reading for sense.

If safe data use is your weak spot

Walk through each plan in use, open its privacy settings and note the state of the training switch. Then write the never-paste list on one page. Share it with whoever uses an assistant for the business and ask them to tell you if their practice differs.

If accountability is your weak spot

Write the four lines for your most frequent output. Pin them where the approver can see them. For anything with legal or financial weight, talk to an adviser about whether a written policy is warranted; the AI policy question for free ChatGPT use gives a way to decide.

Using the same five skills when you hire or delegate

Once you know where you stand, the five skills double as a hiring and onboarding lens. Instead of asking a candidate whether they "know AI", set a ten-minute task for each skill that fits the role: sort a list of duties, brief the assistant on a customer message, find mistakes in a draft, redact a sample letter, describe who signs off a quote.

A care agency manager hiring a coordinator might run only skills three and four, since a coordinator handles personal data daily and reviews rotas and notes before they go out. A music school might care most about skills two and five because the teachers write to parents. The point is that the same framework gives you a short, fair, role-specific test. The interview test for AI skills goes through running it in practice.

Do the same for existing staff, one skill at a time, and treat a low score as a training gap rather than a mark against the person. A member of the team who scores 1 on safe data because nobody told them the rules is exactly the case where the owner's own skill five, the written never-paste list, closes the gap.

If you would rather map these jobs and checks with someone, that is what my AI implementation consultation covers; the link is on the contact page.

Where the framework stops being useful

Five skills do not cover everything. They say nothing about picking between tools, controlling costs or judging whether a vendor's claims are true. They also do not tell you whether AI is the right answer to a given problem, because sometimes a checklist or a spreadsheet formula does the job better and with no risk of invented content.

Think of fluency as the floor, not the ceiling. It decides whether you can use a tool without hurting the business. What you build on top of it, such as automations that run without you, needs more: testing, monitoring and a plan for when something breaks. That is a separate job from the personal skills here.

One more limit: fluency is specific to the task. You may score 4 on drafting customer emails and 1 on checking a spreadsheet summary, since the failures look different. Run the five tasks again for each new kind of work, and treat a good score on one as no promise about another.

Questions owners ask before scoring themselves

Do I need a certificate to count as AI fluent?

No. A certificate shows you finished a course; it says nothing about whether you can brief an assistant, catch a wrong figure or keep customer data out of the wrong tool. The five tasks in this tutorial test the behaviour directly, using your own work, which is the thing that actually protects the business.

How often should I re-run the self-check?

Run it again whenever you change tool, add a new kind of task or hand AI access to a member of staff. A new tool resets some habits, because its defaults, privacy settings and mistakes differ. Keep the filled-in scorecard so you can compare like with like.

Should my staff take the same test?

Yes, with their own tasks. Skills one, two and three are about daily handling and vary a lot between people in the same office. Skills four and five are mostly set by the rules you give them, so a low score there usually points back at the owner, not the employee.

What if I score low on checking the output?

Treat it as the priority. A weak brief gives you a mediocre draft you can see for yourself; a weak check lets a wrong figure or invented claim reach a customer. Add a fixed review step with a named checker, and practise on drafts with planted mistakes until you catch them every time.

Is a free assistant enough to practise these skills?

For skills one to three, yes. For skills four and five the plan matters, because consumer and business plans handle your chats differently. Check the privacy settings of whichever plan you use before pasting anything about customers, and keep practice material invented or stripped of names.

Further reads

• How to Roll Out an AI Policy So Staff Actually Follow It — Get staff to follow the rules once you have written them.
• Best AI for Business Training: Onboarding, Compliance or Skills Gaps? — Choose AI help for training new starters and closing skills gaps.
• AI Tools for Learning and Development: A Small-Company Shortlist — A shortlist of tools for teaching staff to use AI.
• AI Hardware for Small Business: Do You Need a New Computer? — Check whether your equipment limits what you can try.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's free AI Fluency: Framework & Foundations course page (the four competencies it names), Anthropic consumer terms and privacy centre, OpenAI data controls help article and ChatGPT Business pricing page, Microsoft Learn page on Copilot data and privacy.