Which Marketing Tasks Should a Small Business Never Hand to AI?

Never hand AI the final say on anything that makes a factual promise, speaks for a real person, or can't be undone: health and results claims, testimonials, prices and offer terms, public replies to complaints, crisis messages, and decisions about who gets contacted or what customer data is uploaded. AI can draft all of these. A named person must approve them.

"Never hand to AI" rarely means "never use AI anywhere near it". It means the AI doesn't get to decide or publish. The useful split is three-way: tasks AI can run on its own, tasks where AI drafts and a person signs off, and a small set where a person should do the thinking from scratch because the judgement is the job. Most of the list below sits in the middle group, which is good news: you keep most of the time saving and lose most of the risk.

Follow me on Instagram@sagnikteaches

Run it, draft it, or keep it: the three-way split

The sections below take each group in turn, with an illustrative example of how it goes wrong in a small business.

Connect on LinkedInSagnik Bhattacharya

Claims and promises: AI drafts, a person signs off

1. Health, safety and results claims

Language models write persuasive copy by reaching for strong claims, and in health businesses a strong claim is often a false or unlawful one. An illustrative pharmacy asked AI for a social post about its own-brand hay fever tablets. The draft read: "Clinically proven to stop hay fever symptoms within 24 hours." The pharmacy had no such evidence for that product, and medicine advertising is tightly regulated almost everywhere. The pharmacist rewrote it around what the product label actually said.

Subscribe on YouTube@codingliquids

The rule: any sentence about what a product or treatment does, how fast, for whom or how well, is approved by the qualified professional before it's published, with the evidence noted. That covers dental whitening results, osteopathy for back pain, hearing-aid performance, pet treatments and supplements. Backing up every claim AI writes in your marketing sets out a quick evidence log.

2. Prices, discounts and offer terms

A wrong price in a marketing email is a promise customers will try to hold you to. An illustrative osteopath asked AI to turn "50% off first assessments in March" into an email. The subject line came back as "50% off all treatments this March". Fourteen existing patients booked follow-ups expecting half price before anyone noticed. The clinic honoured it for them, at a cost of a few hundred dollars, and changed the process.

The rule: the person who set the offer checks every figure, date, eligibility condition and exclusion in the final version, including subject lines and image text, which are where AI rewording slips through. Put the full terms in one place and link to them rather than letting AI paraphrase them each time.

3. Comparisons with competitors

"The cheapest hearing aids in town", "faster than any other practice", "the only clinic that…" are claims you must be able to prove on the day they're published. AI writes them because they sound good, not because they're true. An illustrative hearing-aid shop's AI-drafted flyer said it offered "the lowest prices in the area". A competitor's customer brought in a cheaper quote, and the shop had nothing to show that the claim had ever been checked. Comparative advertising is lawful in many places when it's accurate and fair, and risky when it isn't.

The rule: no comparative or superlative claim goes out unless the owner has the evidence on file, dated. Most small businesses are better off cutting these lines altogether and saying what they actually offer.

Anything that speaks for a real person

4. Testimonials, reviews and endorsements

AI-written testimonials count as fake reviews under consumer rules in many markets, whatever label you add. AI can tidy the spelling of a real customer's words with their approval; it can't supply the words. The full workflow, including a consent email, is in where the legal line is for AI and testimonials.

5. Images and video of real people and places

A generated picture of "our clinic" that isn't your clinic, a generated "team photo", or a generated patient with a perfect result all tell customers something untrue. An illustrative podiatrist used an AI image of a bright, spacious treatment room on her booking page. New patients arriving at her smaller, older room mentioned it more than once. She replaced it with a well-lit phone photo of the real room and the comments stopped.

The rule: AI may edit real photos (lighting, cropping, removing clutter) but must not invent people, premises or results. Anyone identifiable in a photo, staff included, agrees before it's used.

6. Personal messages at emotional moments

Some messages are valuable precisely because a person wrote them. A veterinary practice's condolence card after a pet is put to sleep, a thank-you note to a long-standing patient who is moving away, a message to a customer after a bad experience. An illustrative vet practice trialled AI-drafted condolence cards to save the nurses time. A client who had lost her dog of fourteen years received a card that called him by the wrong breed. It had been merged from the wrong record. She rang the practice in tears.

The rule: messages at emotional moments are written by the person who signs them. AI can help someone who struggles with words find a structure, privately, but the message itself should come from them, and it should never be sent automatically.

Public replies that can escalate

7. Replies to complaints and negative reviews

AI is good at calm, polite drafts, which makes it useful here, but a public reply is permanent and every word is read by future customers. Two failure patterns recur. The first is the reply that sounds like a template ("We're sorry to hear about your experience. We strive for excellence…"), which reads as indifference. The second is worse: the reply that reveals something private. An illustrative dental practice's AI draft, written from the practice's notes, replied to a one-star review with "As you'll recall, you missed two appointments before your root canal." Accurate, and a confidentiality breach in public.

The rule: AI drafts from the review text only, never from patient or customer records, and a manager edits and posts. Health businesses should never confirm in public that the reviewer is a patient. There's a fuller method in replying to negative reviews with AI.

8. Crisis and sensitive-moment communication

When something goes wrong, such as a data breach, a clinical incident, a flood that closes the premises, or a local tragedy, what you say and when you say it are judgement calls with legal and human stakes. AI can help you check a statement for clarity once you've written it. It shouldn't write it, and it certainly shouldn't be posting cheerfully in the meantime.

That second point catches businesses out. An illustrative pharmacy had a month of AI-drafted, scheduled posts running. On the morning its system went down and prescriptions were delayed, a post went out at 9am: "Nothing beats our speedy service!" The replies were predictable. The fix was simple and worth copying: a one-line crisis checklist whose first step is "Pause all scheduled posts and emails."

Decisions about people and their data

9. Who gets contacted, and whether they agreed

AI can segment a list in seconds. Whether someone may be contacted at all is a decision about consent, which is a legal matter in most places and a trust matter everywhere. An illustrative hearing-aid shop asked AI to "find everyone who might need new hearing aids" in its customer database and send them an offer. The list included people who had unsubscribed and eleven customers who had died, whose families received the email. The data was messy; the AI had no way to know, and no one had checked.

The rule: the owner decides who may be contacted, using the consent records, and a person checks the final list's size and a sample of names before any send.

10. What customer data goes into AI tools

Pasting a customer list, patient notes or a spreadsheet of names and phone numbers into a consumer chat tool is a decision about personal data, and it should be made deliberately, not by whoever is trying to save ten minutes. Use business plans that don't train on your data by default, remove names and contact details wherever the task doesn't need them, and write down who may upload what. Whether it's safe to put customer data into ChatGPT covers the settings and the safer habits.

11. Targeting based on sensitive traits

Ask an AI to plan ads for a podiatrist and it may suggest "target people with diabetes". The major ad platforms have removed or restricted targeting by health conditions and other sensitive characteristics, and even where some version is possible, customers who realise they were targeted because of a health condition rarely feel grateful. Target by interest in the service or by location, and let people who need you find you. This one is a decision to keep, and usually a decision to say no.

Strategy and relationships

12. Your positioning: why customers should choose you

AI averages. Asked why patients should choose a dental practice, it produces "friendly team, modern technology, patient-centred care", which is what every practice says. Your real reason (open until 8pm, the dentist who is good with nervous patients, same-week emergency appointments) comes from knowing your customers and competitors. Use AI to challenge your positioning ("What would a sceptical customer say about this?") and to find the words for it, but decide it yourself.

13. Relationships with referrers, partners and the local press

The doctor who refers patients to your podiatry clinic, the gym that recommends your osteopathy practice, the local journalist who covers your open day: these relationships run on people knowing each other. A mail-merged AI letter to twenty doctors' practices reads as exactly that. The owner writing four personal notes, perhaps with AI's help on the first draft, does far more. Relationship messages should come from the person who owns the relationship, in their own voice.

14. Ad budgets without limits

Letting the ad platforms' AI choose audiences and bids is often sensible; their automated campaign types are built for it. Letting it spend without limits isn't. An illustrative veterinary practice switched on an automated campaign with a "maximise conversions" setting and a monthly budget but no conversion tracking set up properly. It spent the month's budget in nine days, optimising for clicks on a "contact" button that didn't record whether anyone booked.

The rule: you set the daily and monthly caps, confirm conversion tracking works before launch, and check spend against bookings weekly. Setting up conversion tracking before AI spends your ad budget walks through that first step.

What is safe to hand over

For balance, here's what most small businesses can let AI run with little or no review once it's set up, because an error is cheap and easy to spot:

• Resizing and reformatting images and videos for different platforms.
• Captions and transcripts, after a quick proofread for names and technical terms.
• Scheduling approved posts and emails.
• Subject-line and headline variations for testing, from approved copy.
• Alt text for images, checked occasionally.
• Summaries of reports and analytics for your own use.
• Sorting enquiries into categories for a person to act on.
• First drafts of anything, which is where most of the time saving is anyway.

The pattern: AI can run tasks where the output is private, reversible or checked before it matters. It drafts tasks where the output is public or binding. It stays out of decisions about people. For the specific risks of putting social posting fully on autopilot, see what breaks when AI runs your social media on autopilot.

Sorting your own task list in fifteen minutes

Your marketing will include tasks this list doesn't name. AI can do a useful first sort, as long as you make the final call. List everything your business does in marketing in a typical month, then use a prompt like this:

Here is every marketing task my [business] does in a month. Sort each
into: RUN (AI can do it with no review), DRAFT (AI drafts, a named
person approves) or KEEP (a person does it). For each, give a one-line
reason based on: does it make a claim or promise, does it speak for a
real person, is it public, can it be undone, does it involve personal
data. Flag any you're unsure about rather than guessing.

An illustrative extract of the result for a pharmacy: "Weekly Google profile post about opening hours: RUN, low risk, easily corrected. Monthly email about the flu clinic: DRAFT, contains dates, prices and eligibility; approve: pharmacist. Instagram reply to a customer asking which cold remedy to take: KEEP, a medicine question needing professional advice. Choosing which patients to invite for a blood pressure check: KEEP, uses health data. Flagged: resharing a local charity's post (unclear who approves; decide)." The flagged item is the useful one. It's a decision you hadn't made, and now you will.

Check the sort against the fourteen tasks above. If the AI put anything involving claims, prices, real people or personal data in RUN, move it, and treat that as a reminder of why it doesn't get the final word.

One dental practice's month of marketing, audited

An illustration. A two-dentist practice's receptionist handles marketing for about three hours a week, with AI drafting most of it. The owner audited the previous month's output against the three-way split:

• 18 social posts: 15 were routine (hours, team birthdays, a new hygienist) and fine to run with a quick glance. Three mentioned whitening results or "pain-free" treatment and had gone out unapproved. One claimed "pain-free root canals". Moved to DRAFT with dentist sign-off.
• 4 review replies: all AI-drafted, all polite. One confirmed the reviewer had "completed their treatment plan with us", which disclosed that she was a patient. Rule added: reply from the review text only.
• 2 emails: a recall reminder (fine) and a "refer a friend" offer whose AI-written subject line said "$50 off for you and a friend" when the terms offered the $50 only to the new patient. Caught by a patient who asked. Offer terms now approved by the manager, subject line included.
• 1 ad campaign: running within a cap, with bookings tracked. No change.

Four problems in 25 published items, all in the categories this tutorial calls DRAFT, and all from approval being skipped rather than AI being used. After the audit, the receptionist kept using AI for every draft, the dentist spent about ten minutes a week approving clinical claims, and the next month's audit found none. The time saving barely moved; the risk did.

Write the sign-off rule down

A rule that lives in the owner's head doesn't survive a busy week or a new staff member. One page is enough. An illustrative version for a small osteopathy clinic, ready to adapt:

AI IN OUR MARKETING: WHO APPROVES WHAT (reviewed every 6 months)

AI may do on its own:
- resize images, write captions and alt text, schedule approved posts,
  summarise our analytics, suggest subject lines

AI may draft; a named person approves before anything is published:
- any claim about what treatment does or for whom   -> practitioner
- any price, offer, date or condition                 -> practice manager
- replies to reviews and complaints (from the review
  text only, never from patient records)              -> practice manager
- blog, email and social copy                          -> practice manager

AI does not do:
- testimonials (real patients' words only, with written approval)
- images of people or of our premises (real photos only)
- condolence, apology or personal thank-you messages
- statements when something has gone wrong
- choosing who we contact, or uploading patient data

First step in any incident: pause all scheduled posts and emails.
Owner of this page: [name]. Questions: ask before you publish.

Print it, put it where the person who does the marketing will see it, and walk a new team member through it on their first day.

How to tell if the line has slipped

Lines drift, usually because the AI's drafts are good and approval becomes a glance. Warning signs:

• Posts or emails go out that nobody remembers approving.
• Customers quote an offer or claim that staff don't recognise.
• Comments under posts ask "Is this AI?" or point out something untrue.
• Replies to reviews all start with the same sentence.
• Someone uploads a customer export to a chat tool "just to tidy it up".

Once a month, take ten minutes to open the last 20 things you published and ask of each one: which category was it in, and did the right person approve it? If two or more slipped through, tighten the process before the next campaign, not after the next complaint.

Further reads

• Which Agency Tasks AI Handles Well, and Which It Doesn't — The agency view of what AI does well and badly.
• 12 AI Marketing Mistakes Small Businesses Make (and the Fixes) — Twelve common AI marketing mistakes and their fixes.
• How to Write an AI Content Policy for Your Marketing — Turn the sign-off rule into a full content policy.
• How to Set Up Human Review for AI Work Without Slowing Down — Set up review steps that don't slow everything down.
• Which Pharmacy Tasks Should Never Be Handed to AI? — The same question for a regulated health business.
• How to Keep Your Brand Safe When AI Places Your Ads — Guard rails for ads placed by automated systems.
• AI Marketing Tools vs Hiring a Marketer: A Cost Comparison — Tools cost far less than a salary, but only replace the drafting. A task audit, three set-ups priced over a year, and signs you need a person.
• How to Write a Press Release With AI That Journalists Will Read — Test the angle, draft from a fact sheet, strip out the tells that get AI releases binned, and write the short personal pitch that actually gets coverage.
• How Small Businesses Actually Use AI in Marketing: 15 Examples — Fifteen specific marketing jobs small businesses hand to AI, each with the steps, a sample, the time it takes and the catch to watch for.
• Can AI Write a Marketing Plan for a Small Business? — The input pack AI needs, a three-prompt sequence, how to red-pen the first draft, and a veterinary practice's 90-day plan with budget and owners.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Meta and Google advertising policies on sensitive targeting; Google Business Profile Help on review policies; general guidance on consumer-protection and data-protection rules for marketing.