The costly ones are publishing drafts unedited, giving AI no business facts so it invents them, altering reviews, automating replies with no hand-off, letting ad platforms' AI spend without conversion tracking, and pasting customer data into consumer AI accounts. Most share one fix: a fact sheet, a human approval step, and measuring one change at a time.
Very few of these mistakes come from picking the wrong tool. They come from process: nobody decided what the AI may say, who checks it, or how anyone will know whether it helped. That is good news, because process is cheap to fix. Each mistake below comes with how it typically shows up, why it happens, the fix and a quick way to check the fix worked. The business examples are illustrative.
Mistakes in what you publish
1. Publishing drafts without an edit pass
How it shows up. A personal trainer schedules a month of AI-drafted posts in one sitting. Scrolling his own grid a week later, he notices six of the eight captions start with a question ("Ready to transform your mornings?") and all end with three emojis. Followers notice too: likes drop, and a regular client jokes that his posts "sound like a robot now".
Why it happens. Language models produce the most likely wording for a prompt, and for "fitness post" that's the average of thousands of fitness posts. Unedited, it reads like everyone else's.
The fix. A five-minute edit pass on every draft: delete the first sentence (it's usually throat-clearing), add one detail only you could know (a client's result shared with permission, a specific exercise, a time of day), cut any line you wouldn't say out loud, and read it aloud once. Why AI marketing copy sounds generic goes deeper on the causes and the prompts that prevent it.
An illustrative before and after. The draft: "Ready to transform your mornings? Our sessions are designed to help you feel stronger, fitter and more confident. Book now!" After the edit pass: "6.30am, three clients, one kettlebell each. Twenty minutes later they'd done more than most people manage before lunch. Two spaces left in the Wednesday early group." Same length, but only this trainer could have written the second one.
The check. The swap test: replace your business name with a competitor's. If the post still works, it isn't specific enough.
2. Giving AI no facts, so it invents them
How it shows up. A wedding planner asks AI to refresh her packages page. The new version is lovely and promises "unlimited supplier meetings" and "a dedicated coordinator for your rehearsal dinner". She offers six supplier meetings and doesn't cover rehearsal dinners. A couple quotes the page back to her in their first call.
Why it happens. When a prompt leaves gaps, the model fills them with whatever is typical for the category. It isn't lying; it's completing a pattern.
The fix. One page of facts, pasted into every marketing prompt or saved as the instructions of a shared project in your AI tool:
BUSINESS FACTS - use only these; ask if something is missing
Services: Full planning (from $X), partial planning (from $Y), on-the-day
coordination (from $Z).
Included in full planning: 6 supplier meetings, 2 venue visits, timeline,
on-the-day team of 2 for 12 hours.
Not offered: rehearsal dinners, destination weddings, stationery design.
Booking: $500 deposit, balance 8 weeks before.
Never claim: awards, "best", guarantees, anything not listed here.The last line of the sheet does quiet work. Told to ask when something is missing, the assistant will often come back with a question ("Do you offer a payment plan? The brief doesn't say") instead of guessing, which is exactly the behaviour you want.
The check. Highlight every factual claim in a draft and tick each against the sheet. Anything unticked is cut or confirmed.
3. Polishing reviews and writing testimonials
How it shows up. A yoga studio asks AI to "tidy up" three member reviews for the website. The polished versions read better and now include phrases the members never wrote: "the best studio in the area", "cured my back pain". One member sees her "review" and emails, upset, asking for it to be taken down.
Why it happens. Rewriting feels like editing, but a testimonial is someone else's statement. Changing it, or generating one, attributes words to a person who didn't say them, which is misleading and in many places unlawful in advertising.
The fix. Quote reviews word for word, with permission, and let AI help only with selection: "From these 40 reviews, pick five short excerpts that mention beginners, quoted exactly, with the reviewer's first name." Never let it rewrite them. Where the legal line sits on AI and testimonials covers the detail.
Selection alone does most of what the owner wanted. From the studio's reviews, an exact excerpt such as "I'd never done yoga and was terrified of being the worst in the room. Nobody cared, and by week three I'd stopped worrying" does more for nervous beginners than any polished line, precisely because it sounds like a real person. Ask the member before using it on the website, and note the date they agreed.
The check. Paste the website quote and the original review side by side; they should be identical apart from trimming marked with an ellipsis.
4. AI images that misrepresent the work
How it shows up. A tattoo studio posts AI-generated images of "healed fine-line tattoos" because its own photos are inconsistent. A follower spots the tell-tale mangled fingers and comments "these aren't real tattoos". The comment gets more likes than the post.
Why it happens. For businesses whose product is their craft (tattooists, photographers, hairdressers, makers), images are the evidence. Replacing evidence with generated images breaks the one thing the image was for.
The fix. Real work only in anything that looks like a portfolio. Use AI for backgrounds, textures, layout and editing your own photos, and label realistic AI images where they could be mistaken for real people or work. Ad platforms apply their own labels in some cases. Whether to tell customers you used AI sets out when a label is needed.
For the tattoo studio, the practical fix was a monthly photo session: every healed piece clients came back to show got photographed in the same corner, same light, same angle, with permission. Within two months the studio had more consistent real images than it needed, and the AI images came down.
The check. For every image, ask: would a customer who saw this expect to get exactly this? If the honest answer is no, it doesn't go in.
Mistakes in what you automate
5. Auto-replies with no limits and no hand-off
How it shows up. A photography studio switches on an AI assistant for Instagram and website chat. A parent asks about Saturday availability for a newborn session; the assistant replies "Yes, we have Saturday slots available, see you then!" There were no Saturday slots, and nothing was booked.
Why it happens. Chat assistants are built to be helpful and agreeable. Without a clear list of what they can and can't say, they'll confirm things they have no way of knowing.
The fix. Three rules. The assistant answers only from a written FAQ and price list. It never confirms availability, bookings, prices outside the list or anything medical; instead it links the booking page or says a person will reply. And certain words (complaint, refund, allergy, injury, urgent, cancel) trigger an immediate hand-off to a human. If you sell to customers in the EU, the EU AI Act's transparency rules have applied since 2 August 2026, so the chat must make clear the customer is talking to AI. Chatbot guardrails that stop AI promising what you don't offer has the full setup.
The hand-off reply should sound like a person, not an error message. Something like: "That's one for the photographer herself. I've passed your message on and she'll reply today. If it's about a date, you can see open slots here: [booking link]." Then make sure someone actually does reply that day, because a promised human who never appears is worse than no chat at all.
The check. Before launch, ask it twenty awkward questions yourself, including three it should refuse. Repeat monthly.
6. Letting ad platforms' AI spend without conversion tracking
How it shows up. A driving school launches an automated campaign on Meta, optimised for link clicks because that was the default it accepted. Clicks are cheap and plentiful. Lesson bookings don't move. After a month and a real chunk of budget, the owner realises the algorithm found exactly what it was asked for: people who click, not people who book.
Why it happens. Automated campaign types (Meta's Advantage+, Google's Performance Max and AI Max for Search) optimise towards whatever signal you give them. With no conversion data, they optimise towards the nearest thing they can measure.
The fix. Before handing budget to any AI campaign, set up tracking for the action that matters (booking made, enquiry form submitted, call from the ad), check it fires by doing it yourself, then optimise the campaign for that event and set a daily budget cap. Setting up conversion tracking before AI spends your budget takes you through it.
A quick sum shows how fast this adds up. At an illustrative $15 a day, a month of click-optimised spend is $450. If that brought 900 clicks and two lesson bookings, each booking cost $225, far more than a first lesson block is worth. The same budget optimised for completed enquiry forms might bring fewer, pricier clicks and many more enquiries; you only find out with tracking in place.
The check. Your ad dashboard's conversions should roughly match the bookings your own system recorded from ads that week. If they're wildly different, tracking is broken.
7. Pasting customer data into consumer AI accounts
How it shows up. A photography studio owner uploads her full client spreadsheet (names, emails, phone numbers, children's names and birthdays) to a free AI account to "find who's due a birthday session". It works. Later she realises the account's model-training setting was on, and she has no clear idea where the data went or how long it is kept.
Why it happens. Consumer AI plans are designed for individuals, and many keep chats for training unless you switch that off. Business plans (ChatGPT Business, Claude Team, Gemini in Workspace, Microsoft 365 Copilot) don't train on business content by default, but plenty of small businesses use personal accounts for work.
The fix. Decide what never goes into an AI tool (children's details, health information, payment details). For everything else, strip names and contact details and replace them with a customer number before uploading; the analysis rarely needs to know who anyone is. Use a business plan, or switch off the model-training setting in a consumer account's privacy settings.
A short "never goes in" list, pinned where everyone can see it, works better than a long policy. For the studio it read: children's names and birthdays, home addresses, payment details, anything a client told us in confidence, and any photo of a client that hasn't been cleared for marketing. Everything else can go in once names and contact details are swapped for a client number.
The check. Open the file you're about to upload and search it for "@". If email addresses are there, it isn't ready.
Mistakes in where you spend time and money
8. Using AI to post more instead of better
How it shows up. A yoga studio goes from three posts a week to daily, because AI makes drafting easy. Three months later, total reach is about the same, reach per post has roughly halved, and intro-offer bookings haven't changed. The owner is spending more time than before approving posts nobody engages with.
Why it happens. Cheap drafting removes the natural limit on volume, but audience attention didn't grow.
The fix. Give every post a job (announce, show, teach, prove, invite) and cut any post that has none. Many small businesses do better with three well-made posts a week than seven thin ones. Use the time AI saves to make each post more specific: a better photo, a real member story, a clearer offer.
The time maths is worth doing too. Seven posts a week at 15 minutes each (drafting, editing, image, approval) is almost two hours a week. Three better posts at 25 minutes each is 75 minutes, and each one has a real photo and a specific story. The owner got time back and better posts by doing less.
The check. Compare average engagement per post and bookings from social over the same period before and after the change, not total post count.
9. One prompt, every channel
How it shows up. A wedding planner writes one AI caption and pastes it everywhere. Her email subject line ends up with hashtags. Her Google Business Profile post says "link in bio", which means nothing on Google. Her LinkedIn post, aimed at venues she'd like to partner with, uses the same "hey lovebirds" opening as her Instagram.
Why it happens. Repurposing is one of AI's best uses, but only if each version is written for its channel.
The fix. Keep channel rules in the prompt:
| Channel | Rules to give the AI |
|---|---|
| Under 60 words, one idea, hashtags at the end, audience is couples | |
| Subject under 45 characters, no hashtags, one link, written as a note | |
| Google Business Profile | Under 100 words, a clear offer or update, use the button for the link |
| Professional tone, audience is venues and suppliers, no emojis |
The check. Read each version on the device and platform it'll appear on before scheduling.
10. Chasing AI search with tricks
How it shows up. A tattoo studio pays for an "AI search optimisation" package that adds an llms.txt file and stuffs every page with question-and-answer blocks full of keywords. Nothing changes, except that the website now reads badly.
Why it happens. AI search is new and confusing, which makes it easy to sell shortcuts.
The fix. Google says its AI Overviews and AI Mode need no special optimisation beyond normal good SEO, and that it doesn't use llms.txt files or special markup for them. What helps is plain: accurate, consistent facts (hours, prices from, services, policies) on your website and business profiles, a genuinely useful FAQ page and a steady stream of honest reviews. One technical check is worth doing: if your site blocks OpenAI's OAI-SearchBot crawler, it can't appear in ChatGPT search answers (its GPTBot crawler is for training only, which you may block separately if you prefer).
When the studio finally ran that check, the problems were ordinary: one assistant gave old opening hours taken from a directory listing, and another said the studio didn't do walk-ins because an old web page said so. Updating those two sources did more than the entire paid package.
The check. Ask two or three AI assistants the questions your customers ask about you, monthly, and note whether the answers are accurate.
11. Buying tools before deciding the jobs
How it shows up. A sole-trader personal trainer signs up, over two months, for a chat assistant ($20 a month), Zapier Professional ($29.99 a month on monthly billing), a paid social scheduler and an AI copywriting subscription at around $40 to $50. That's well over $100 a month. Six weeks later he uses only the chat assistant regularly.
Why it happens. Every tool demo shows a job being done beautifully. It's easy to buy the demo rather than the job.
The fix. List the three marketing jobs that cost you the most time. For 60 days, do them with one general chat assistant and the free tiers of what you already have (Meta Business Suite schedules Facebook and Instagram posts for free; Buffer's free plan covers three channels with ten scheduled posts each; Zapier's free plan runs 100 tasks a month on two-step automations). Pay for a specialist tool only when a specific limit blocks a specific job.
The check. Once a quarter, list every subscription and the last date you used it. Cancel anything unused for a month.
Mistakes in judging whether it worked
12. No baseline, so no idea whether AI helped
How it shows up. The same personal trainer tells a friend AI has "saved him loads of time". Asked how many enquiries he gets now compared with before, he doesn't know. When he checks his inbox, enquiries have actually dipped since his posts became more generic.
Why it happens. AI makes work feel faster, and that feeling is easy to mistake for results.
The fix. Before changing anything, write down three numbers for the previous month: hours spent on marketing each week, enquiries or bookings from marketing, and marketing spend. Then change one thing at a time (AI-drafted posts, say) and compare after a month. If hours fell and enquiries held steady or rose, keep it. If enquiries fell, the time saving cost you more than it saved.
Filled in for the trainer, three months of notes might look like this (illustrative):
| Month | Marketing hours a week | Enquiries | Spend | What changed |
|---|---|---|---|---|
| Before AI | 4 | 14 | $0 | Nothing |
| Month 1 | 1.5 | 9 | $20 | AI-drafted posts, unedited |
| Month 2 | 2 | 15 | $20 | Same, plus the five-minute edit pass |
Month 1 alone would have looked like a success on time saved. The enquiry column shows it wasn't, and month 2 shows the fix: half an hour more a week, and enquiries back above where they started.
The check. A three-line note in your calendar on the first of each month: hours, enquiries, spend.
A 20-minute monthly review that catches most of these
Most of the twelve are caught by the same short routine. Once a month, with a cup of tea:
- Read five random published pieces from the month. Run the swap test and tick the facts against your fact sheet. (Mistakes 1, 2, 9)
- Scroll your image grid and ask whether each image shows something a customer would actually get. (Mistake 4)
- Ask your chat assistant or DM bot five awkward questions. (Mistake 5)
- Compare ad-reported conversions with your own bookings from ads. (Mistake 6)
- Check which AI accounts hold customer data and whether training is off. (Mistake 7)
- Write down hours, enquiries and spend, and list subscriptions not used this month. (Mistakes 8, 11, 12)
- Ask two AI assistants about your business and note any wrong answers. (Mistake 10)
Keep the notes in one place. Over a few months, the same one or two problems usually turn out to cause most of the trouble, and that tells you where a better prompt, a clearer rule or a different tool would pay off. Reviews and testimonials (mistake 3) only need checking when you add new ones, which is a good reason to make "quote exactly, with permission" part of how you collect them.
Questions after spotting a mistake
Which of these mistakes is most expensive?
Usually letting an ad platform's AI spend without conversion tracking, because the cost is cash rather than reputation and it compounds daily. The most damaging to trust is publishing invented claims or altered reviews, since customers who spot one doubt everything else. Fix tracking first if you run ads; otherwise start with a fact sheet and an approval step.
We've already published AI content with errors. What now?
Correct the facts first, starting with prices, availability, qualifications and anything a customer might rely on. Then check older posts and pages for the same error, because a wrong detail in a fact sheet or prompt tends to repeat. If a customer acted on the wrong information, deal with them directly and honestly. Add the error to a simple log so the prompt or checklist gets fixed.
Do I need an AI policy to avoid these?
A one-page set of rules helps once more than one person uses AI for marketing: what may be generated, what must be checked, what data never goes into AI tools, who approves before publishing, and when AI use is disclosed. A sole trader can keep the same rules as a short checklist. The rules matter more than the document.
Further reads
- How Small Businesses Actually Use AI in Marketing: 15 Examples — Fifteen examples of the same jobs done well.
- Which Marketing Tasks Should a Small Business Never Hand to AI? — The marketing jobs to keep fully human.
- Social Media Pre-Publish Checklist for AI-Drafted Posts — A pre-publish checklist for every AI-drafted post.
- How to Back Up Every Claim AI Writes in Your Marketing — How to source or cut every claim in AI copy.
- Does Google Penalise AI-Written Content on a Business Website? — What Google actually says about AI-written pages.
- How to Write an AI Content Policy for Your Marketing — Turn these fixes into a short policy for your team.
- Performance Max for Small Businesses: 9 Costly Mistakes — The costly mistakes specific to Google's AI campaigns.
- Common AI Mistakes Small Businesses Make and How to Avoid Them — Twelve everyday AI mistakes in small firms, grouped by data, output, money and ownership, each with how to spot it and fix it in under an hour.
- How Meta and Google Use AI to Run Your Ads: A Plain-English Guide — How Advantage+, Smart Bidding, AI Max and Performance Max decide who sees your ads, what they learn from, and the settings a small business still controls.
- 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: Google Search Central guidance on AI features; OpenAI crawler documentation (OAI-SearchBot and GPTBot); EU AI Act Article 50 transparency duties; vendor pricing pages for the tools mentioned (checked September 2026).