The FTC Is Now Fining Companies Over the AI Claims in Their B2B Marketing

On May 21 the Federal Trade Commission announced proposed settlements with three companies over an advertising product called Active Listening. CMG Media Corporation, which does business as Cox Media Group, sold it to small businesses as an AI service that listened through the microphones in phones, televisions and smart speakers, worked out who was shopping for what, and served those people local ads.

The FTC says none of it was real. No voice data was ever collected. The geographic targeting missed the areas customers were paying to reach. What buyers got were email lists CMG had purchased from data brokers and resold at a heavy markup. All three companies also told prospects that consumers had opted into the arrangement by accepting app terms of service.

CMG pays $880,000. MindSift LLC of New Hampshire and 1010 Digital Works LLC of Wisconsin pay $25,000 apiece, and the money goes back to CMG's customers as redress. Going forward all three are barred from misrepresenting what their advertising services do, how they collect and use voice data, whether consumers consented, and how accurately they target by geography. The Commission voted 2-0.

Under a million dollars will not frighten anyone. The second count in the complaints is another matter.

MindSift and 1010 were charged for the materials they handed CMG

Alongside the deception claims, the FTC charged MindSift and 1010 Digital Works with giving CMG the means and instrumentalities to deceive its customers. What they gave it was marketing collateral and sales presentations.

Two firms wrote the materials. A third put them in front of buyers. All three landed in the complaints, and the two that never sold anything to the deceived customers were charged over the documents.

That chain runs through most B2B software companies. You put your vendor's capability claim into your own deck. Your agency writes a product page from a briefing your engineers gave them. Your resellers close deals using a one-pager your team produced. Any advertising lawyer will tell you the means and instrumentalities theory has been around for decades, and they are right about that. Until now it had not been pointed at AI capability claims moving from one business to another.

Once you publish your vendor's claim, it stops being your vendor's claim.

Every buyer in these cases was a business

Read across the docket and the same buyer keeps turning up.

Active Listening went to small businesses buying advertising. Air AI, settled in March, involved claims about business growth, earnings potential and refund guarantees pitched at entrepreneurs and small firms. Growth Cave, resolved in January, sold software the FTC says was marketed as automating nearly the whole job of running an online education business while users were left uploading their own ads, booking their own appointments and writing their own messages. Go back further and you reach Evolv Technologies, which sold AI weapons detection to schools and public venues, and IntelliVision, pursued over the accuracy and bias claims it made for its facial recognition software.

Nothing in Section 5 exempts a buyer who should have known better. B2B marketing has spent years assuming consumer protection law belongs to consumer companies, and that a procurement committee counts as a sophisticated party running its own diligence. Eight months of enforcement say otherwise.

DLA Piper flagged as much in its note on the CMG action, reading the case as enforcement in a business-to-business context.

Read the Workado order if you write product copy

Workado sold an AI content detector and advertised that it could separate human writing from AI writing with better than 98% accuracy. Testing came in around 53%.

How that gap opened up is the part marketers should sit with. Workado had trained the model on academic writing. It sold the accuracy number as a general one.

The final order bars the company from advertising accuracy or efficacy without competent and reliable evidence behind it, makes it hold on to whatever evidence it used, and requires it to tell subscribers about the settlement.

Two things follow for anyone writing product copy. Evidence has to exist before the claim goes live, and somebody has to be able to produce it a year later, which means a named owner and a file instead of a Slack thread. The harder one is that a number cannot outrun the conditions that produced it. Benchmark your product on one document type and what you have is a claim about that document type. A great many live AI performance claims in B2B software fail that distinction today.

Where this gets overstated

Start with the money, which is mostly theater. Air AI's $18 million judgment was suspended almost entirely once the FTC accepted that the operators could not pay it, leaving $50,000 in consumer relief. What actually bit was the permanent ban on marketing business opportunities. CMG's settlement barely registers against the size of the company. Nobody has been ruined by an AI-washing action yet.

The case count is contested as well. Holland & Knight, writing at two years of Operation AI Comply, puts it at more than a dozen AI-washing cases last year. DLA Piper counts three since April 2025 on a tighter definition limited to AI-specific deceptive marketing. Both hold up depending on where the line falls, so anyone quoting "a dozen cases" owes readers the definition they are using.

The Commission reverses itself too. It reopened and set aside its 2024 consent order against Rytr after deciding the complaint had not met its own requirements. Chairman Andrew Ferguson has described the program to Congress as protecting growth in the AI market by going after bad actors, which sounds like an agency choosing its targets. Choosing targets is a long way from sweeping a category, and there is no reason to read enforcement risk into every adjective on a homepage.

What holds up through all of that is narrow and unpleasant. Liability now reaches the firm that wrote the deck, and selling to businesses buys no protection.

The audit

Read your own site the way an FTC staff attorney would.

Every performance number needs an owner and a source. Accuracy rates, hours saved, percentage lift. Whoever holds the evidence should be able to say what conditions the test ran under, and if those conditions are narrower than the claim on the page, the claim is wrong.

Absolutes draw the most attention. Fully autonomous, bias-free, no human in the loop, zero hallucination. Enforcement has landed on these exact formulations and they are close to impossible to substantiate.

Then go find the claims you inherited. Anything on your site that started life in a vendor datasheet, a partner one-pager or a model provider's press materials belongs to you the moment you publish it. Provenance language deserves the hardest look of all, since "trained only on licensed data" is an assertion about somebody else's operation that you almost certainly cannot verify.

Last, look at what you hand your partners. The materials your resellers and agencies sell with are the same category of document MindSift and 1010 Digital Works were charged over.

You can keep AI in your positioning. What has to change is the order of operations. Test the thing, keep what the test produced, then write the page.