Every tool sold to your shop this year has AI in it. Most of the claims are noise. A few of the uses are genuinely good, and one rule separates them.
The rule is this. AI may rearrange things you said and things you did. It may never produce a fact, a price or a project that did not exist before you asked.
That is not a philosophy. It is the line where advertising law sits, and the gap between the two has already cost named businesses money.
What AI is actually good at in a small shop
Three jobs, all of them the same shape.
Drafting from your own transcript. You record a walkthrough, you talk for nine minutes about what you are proposing and why, and something turns that into a written summary a client can read. Every fact in it came out of your mouth.
Captioning your own photos. You took the photo, you know the stage, the software writes the sentence under it. The photo is the evidence.
Summarising your own record. The client asks what was decided about the hardware, and the answer is assembled from what is in the file rather than from your memory at half past six.
Notice what all three have in common. The material already exists. The model is a compressor, not an author. That is where it is reliable and it is also the only place it is safe.
What it must never do
Never a price. Pricing is arithmetic on your rates, and arithmetic is a thing software does correctly and predictably. A language model producing a number is producing a plausible number, which is a different object entirely and one you cannot defend when a client asks how you got there.
Never a project you did not build. Never a material you do not offer. Never a timeline nobody committed to. Never a testimonial.
And never a blank filled in because the blank looked bad empty. If the record does not contain the wood species, the correct output is an empty field and a flag, not a guess that reads well.
The law has no AI exemption
The US regulator was explicit when it announced a sweep of enforcement actions against companies using AI in deceptive schemes. The cases, it said, make clear that there is no AI exemption from the laws on the books (FTC, published 25 September 2024, read 2026-09-19). That announcement is now more than eighteen months old, and one of the cases in it was later reopened and set aside, so treat the sweep as context rather than as live precedent (FTC, published 22 December 2025, read 2026-09-19).
The baseline rule underneath it has not moved. The regulator's own guide for small business states that advertising must be truthful and non-deceptive, that advertisers must have evidence to back up their claims, and that they must have proof before the ad runs (FTC, page dated April 2001, edited January 2025, read 2026-09-19).
Two enforcement actions since then show what that means in practice. One company promoted its AI content detector as 98 percent accurate, while independent testing put the accuracy on general-purpose content at 53 percent; the order requires it to hold competent and reliable evidence before making efficacy claims (FTC, published 28 April 2025, read 2026-09-19). In a more recent case, three firms agreed to pay $930,000 over claims that an AI service listened to conversations from smart devices to target ads, when the service did not listen or use voice data at all (FTC, published 21 May 2026, read 2026-09-19).
Neither of those is a cabinet shop. Both are about the same thing: saying a thing your evidence does not support.
Fake reviews are named specifically
The US rule on consumer reviews and testimonials addresses reviews that misrepresent that they are by someone who does not exist, and names AI-generated fake reviews as an example (FTC, published 14 August 2024, read 2026-09-19). That announcement is dated but the rule is in force.
If an agency offers to seed your Google profile with reviews, the answer is no, and the reason is written down.
Canada and the UK say the same thing differently
In Canada, the test is the general impression a representation conveys, as well as its literal meaning, and all representations in any form whatsoever that are false or misleading in a material respect are covered. Under the civil regime, first-occurrence penalties run up to $750,000 for individuals and $10,000,000 for corporations (Competition Bureau Canada, page modified 19 February 2026, read 2026-09-19).
"In any form whatsoever" includes a page your website tool generated overnight.
The Canadian advertising self-regulator's current guidance is more pointed still. If an image showing the benefits or attributes of a product is generated by AI, or significantly altered using AI, it should be disclosed, because it can affect the truthfulness of the endorsement. And then the sentence every shop should write on the wall: images or videos created using filters or AI must remain truthful as to the benefits, results and attributes of the product, and a disclaimer cannot correct a misleading impression (Ad Standards, Influencer Marketing Disclosure Guidelines, update Fall 2025, read 2026-09-19).
A rendering of a kitchen you never built does not become honest because the caption says it is a concept.
The UK regulator puts the same point structurally: AI might change how ads are made, but it does not change how they are regulated, the code is media-neutral, and if an AI tool produces something problematic the advertiser remains responsible for any output used in their ads (ASA, published 11 June 2026, read 2026-09-19).
What it costs when nobody checks
Two documented cases, neither of them speculative.
A US metropolitan newspaper printed a syndicated summer reading list in which ten titles and descriptions were false or invented out of whole cloth. The freelancer had used an AI agent and sent the work in without checking it. The paper's chief executive published an apology, the company stopped taking editorial sections from that supplier, subscribers were not charged for the edition, and a review process was put in place (Chicago Sun-Times, published 29 May 2025, read 2026-09-19).
A global consulting firm's Australian arm agreed to a partial refund on a A$290,000 government report that contained references to non-existent academic papers and a fabricated quote from a court judgment. The revised version disclosed that a generative AI system was used in its creation (Fortune, published 7 October 2025, read 2026-09-19).
A firm that size has a review process. It still shipped invented citations in a paid deliverable.
The legal profession has the clearest record of all, because courts write down what happened. A public database of decisions in which a court addressed the use of AI in more than a passing reference listed 2,045 cases as of 19 September 2026, and it states that it tracks cases where a court found or implied that a party relied on hallucinated content, not the wider universe of all fake citations (AI Hallucination Cases Database, read 2026-09-19).
Two thousand professionals whose entire job is checking sources filed fabricated ones. This is not a problem carefulness alone solves.
The quiet cost: your own website
There is a less dramatic failure that hits shops more often.
Search guidance names scaled content abuse as generating many pages primarily to manipulate rankings rather than help users, and gives as its first example using generative AI tools to generate many pages without adding value. Sites that violate the policies may rank lower or not appear at all (Google Search Central, updated 28 August 2026, read 2026-09-19).
The agency that fills your site with forty AI-written pages about kitchen remodelling in your county is not buying you traffic. It is spending your domain's credibility.
What this means for a two-person shop
You are not behind. In a survey of 3,870 US small businesses with fewer than 250 employees, fielded in June 2025, 58 percent said they use generative AI, up from 40 percent in 2024. Among those not using it, the most cited reason was concern about the quality of the tools, at 33 percent (U.S. Chamber of Commerce, fourth edition, 2025, read 2026-09-19).
That concern is correct and it is also solvable. The answer is not to avoid the tools. It is to only let them touch material you produced.
Three tests for any tool a vendor shows you. Ask where each sentence came from, and if the answer is not "a record you created", walk. Ask whether anything reaches a client without you pressing approve, and if it does, walk. Ask whether the model ever touches a price, and if it does, walk.
And keep a rule for yourself: an empty field is a better outcome than a filled one you did not verify. Casework drafts only from your own project record and your own transcripts, leaves a field blank and flags it when the record does not hold the answer, and sends nothing client-facing until you approve it.
What is not public
There is no published figure for how many trades businesses have published AI-generated content containing false claims. The cases above are documented individually, not counted.
Two widely quoted regulator blog posts on AI claims from 2023 are no longer reachable at their published URLs and are not cited here for that reason.
The share of small businesses that have had to retract or correct AI-generated marketing content is not published anywhere that could be verified.
Where a number does not exist, the honest move is to say so. That is the same rule this post is about, applied to itself.