AI in Marketing: What the 2026 Data Actually Shows

How many marketers are actually using AI, and for what?

The clearest, most recent picture comes from the Content Marketing Institute and MarketingProfs' 16th annual B2B survey: 1,015 B2B marketers, mostly in North America, surveyed between June and August 2025. Ninety-five percent say their organization uses AI-powered applications of some kind, and 89% specifically use AI tools to generate marketing copy — a level of adoption that makes "should we use AI for marketing" close to a solved question for most B2B teams.

What that adoption actually looks like task-by-task is clearer in HubSpot's own AI Trends for Marketers report, surveying over 1,000 marketing and advertising professionals worldwide. Content QA (spellcheck, accessibility checks) leads at 53%, followed closely by general text-based content creation at 52% and writing copy for blogs and emails specifically at 50%. Generating images with AI art tools comes in at 48%. The number that matters most for judging how "automated" marketing has actually become: only 4% of marketers say they use AI to write an entire piece of content outright. For the overwhelming majority, AI is a drafting and editing assistant bolted onto a human process — not a replacement for one.

Is this an enterprise story, or are small businesses actually using AI too?

Small businesses are in this data too, and the gap with enterprise adoption is smaller than the "AI is for big tech" narrative suggests. Constant Contact's State of Small Business Marketing report, based on a survey of 2,500 small-business decision-makers across Australia, Canada, the UK and the US run by Ascend2 Research in June 2025, found 48% of SMBs globally are already using AI in their marketing. The single most common use case, matching the enterprise data almost exactly, is writing copy for emails and social posts — with visual content creation and data analysis trailing behind it. Worth flagging: adoption in the US, one of the four countries surveyed, trailed the other three — this is a global-not-uniform trend, not an every-market-equally story.

Email specifically is moving fast on the image side. Litmus (a Validity brand) surveyed 692 marketing professionals across five countries in late December 2024 and found that AI-generated image usage in email had grown 340% in a single year — the sharpest adoption curve in their whole report. The comparable text-copy figure, 49% of marketers saying they'd use generative AI for static copy creation in 2025, was measured as a forward-looking intention at the start of the year rather than a confirmed, end-of-year usage number — a distinction worth keeping in mind whenever a stat is phrased as "will use" rather than "used."

Is all this adoption actually translating into better marketing?

Here's where the enthusiasm and the evidence pull apart, and it's the single most useful number in the whole CMI/MarketingProfs report: of the marketers who report actually using AI, only 39% say it's improving their performance. That's not a small caveat sitting next to a 95%-adoption headline — it's close to the whole story. Most teams that adopted AI can't yet say it moved a number they track.

Salesforce's tenth State of Marketing report — a double-blind survey of 4,450 marketing decision-makers across 26 countries, fielded in October and November 2025 — landed on almost the identical gap from a different angle. Its own headline finding: 75% of marketers have adopted AI, yet 84% admit they're still sending generic, one-way campaigns, and 61% say full AI integration into their workflow remains a work in progress. Adoption, in other words, is not the same question as "did this change what we send." For most marketing teams surveyed at the end of 2025, it hasn't yet.

Does AI actually save marketers time — or is that mostly self-reported?

The strongest evidence here isn't a marketing-vendor survey at all — it's a preregistered, peer-reviewed randomized controlled trial. MIT researchers Shakked Noy and Whitney Zhang gave 453 college-educated professionals (marketers, grant writers, consultants, data analysts, HR staff and managers) incentivized, occupation-specific writing tasks — cover letters, restructured emails, analysis plans — and randomly gave half of them access to ChatGPT. The group with ChatGPT finished 40% faster (11 minutes saved on average) and scored 18% higher on quality, graded by independent human raters. The most interesting finding wasn't the average, though: performance inequality between workers dropped, because the writers who scored lowest on the first, no-AI task gained the most from having AI on the second one. The researchers themselves flagged a real limitation worth repeating here: they graded writing quality, not factual accuracy, so the study says nothing about whether the AI-assisted output was actually correct.

Industry-reported time-savings claims point in the same direction but are considerably softer evidence. HubSpot's marketer survey reports an average of one to two hours saved per workday — a self-estimate with no disclosed survey dates or sampling detail beyond "1,000+ marketing and advertising professionals worldwide." Salesforce's report states that marketers it classifies as "high performers" say using AI agents lets them reclaim about 8 hours a week — a real and specific number, but one that applies to a self-identified high-performing subgroup, not the average marketer in the survey; reporting it as "marketers save 8 hours a week" without that qualifier overstates what was actually measured. We also looked specifically for a rigorous, independently disclosed study measuring the performance gap between purely AI-drafted content and AI-drafted-then-human-edited content — open rates, conversion, engagement — and did not find one that met a bar of disclosed methodology and independent verification. Several such percentages circulate on marketing blogs; none traced back to a real, checkable source.

What's actually claimed about AI marketing vs. what checks out?

What actually happens when AI-generated content gets it wrong?

This isn't hypothetical anymore, and the cases are specific enough to check yourself. In February 2024, Canada's BC Civil Resolution Tribunal ruled against Air Canada after its website chatbot told a customer he could apply for a bereavement discount retroactively — he couldn't, and the airline refused to honor it. The tribunal ordered Air Canada to pay $812.02 CAD and, more importantly, explicitly rejected the airline's argument that the chatbot was "a separate legal entity" responsible for its own words: a company is liable for what its AI says on its own website, the same as a static page.

Regulators have started attaching real dollar figures to AI marketing claims specifically, not just AI-generated errors. In September 2024 the FTC announced "Operation AI Comply," charging five companies with deceptive AI-related marketing, most notably DoNotPay — marketed as the "world's first robot lawyer" able to "replace the $200-billion dollar legal industry" — which had never tested whether its AI's output matched what a licensed lawyer would produce and hadn't hired or retained attorneys to test the AI's output. DoNotPay's finalized settlement (February 2025): $193,000 in monetary relief, a requirement to notify every 2021–2023 subscriber of the service's actual limitations, and a ban on future unsubstantiated claims that its AI performs like a licensed professional. Two smaller, less legally consequential incidents made the same point about brand risk rather than legal risk: UK courier DPD disabled part of its customer-service chatbot in January 2024 after a customer got it to swear and call the company "the worst delivery firm in the world," and in April 2025 the AI coding tool Cursor's own support bot invented a nonexistent "one device per subscription" policy in response to a real login bug, triggering public subscription cancellations before the company's co-founder corrected it publicly.

How often does AI actually hallucinate?

There's no single honest answer to "what percentage of the time does AI make things up," because the number depends entirely on the task, and most circulating figures don't say which task they're measuring. The most-cited ongoing tracker, Vectara's Hallucination Leaderboard, is transparent about exactly what it measures and doesn't: it gives models over 7,700 short source documents and asks them to summarize using only the information in the passage, then scores factual consistency between summary and source with a separate detection model. In the version we checked, the best-performing model scored around 1.8%; several others scored between 23% and 24% on the same narrow task. That's a real, disclosed methodology — and it's also a single vendor's self-published, continuously updated leaderboard, not an independently audited academic study, and it says nothing about hallucination rates in open-ended writing, marketing copy, or answering questions about your own business's policies (the exact failure mode behind the Cursor incident above).

What do you actually have to disclose when you use AI in marketing?

The honest answer is: it depends entirely on where your customers are, and most of the specific-sounding rules circulating online overstate what's actually required. In the US, the FTC's final rule against fake reviews and testimonials took effect October 21, 2024, with penalties up to $51,744 per violation at the time — a figure the FTC's own annual inflation adjustment has since raised to $53,088 (effective January 2025, and unchanged for 2026) — but its own published guidance is explicit that there's "no blanket prohibition on the use of AI-generated avatars in marketing." What's banned is a review or testimonial that misrepresents who's giving it or whether they actually used the product — AI involvement isn't itself the violation.

In the EU, the AI Act's transparency obligations (Article 50) became legally applicable on August 2, 2026 — worth noting given this article's own publish date, it's a matter of weeks old. From that date, providers of systems that generate synthetic audio, image, video or text must mark the output as AI-generated in a machine-readable format, and anyone deploying an AI system to create a "deep fake" must disclose that fact — with a further grace period on the machine-readable marking duty specifically running to December 2, 2026. The nuances are narrower than they're often summarized: an evidently artistic, satirical or fictional deepfake still has to be disclosed as AI-generated, just in a way that doesn't hamper the display or enjoyment of the work itself — it's an adapted disclosure, not an exemption. The genuine human-review carve-out is narrower still: it applies specifically to AI-generated or manipulated text published on a matter of public interest, and only where a named person or entity retains editorial responsibility for it.

Platform rules are moving on their own timeline, and they're not identical to each other or to any government rule. Meta announced in February 2025 that its own in-house generative-AI ad tools trigger an AI label when they produce a "significant edit," and that it will separately auto-detect and label some third-party-AI-edited ads — explicitly noting this "may vary in certain regions due to legal requirements," i.e. it isn't one uniform global mandate. A stricter, narrower rule applies specifically to ads about social issues, elections or politics: photorealistic AI-manipulated content there requires advertiser self-disclosure now, with Meta's own automatic detection for that category starting June 1, 2026. Google Ads, separately, rolled out an optional AI-content labeling tool starting July 2026 explicitly framed as a compliance aid for EU, Indian and New York State transparency rules — not a Google-imposed requirement in its own right, and Google says using it "doesn't guarantee compliance."

What about Israel specifically?

For an Israeli business, the honest summary is: guidance exists, binding law doesn't — yet. Israel's Ministry of Innovation, Science and Technology published a national AI policy document in December 2023, but it's explicitly non-binding: sector-by-sector principles aligned with OECD guidelines, meant to steer individual regulators rather than create one cross-sector AI law. It creates no marketing-disclosure duty on its own.

Israel's Privacy Protection Authority has a draft guidance document (dated April 2025) on how the Privacy Protection Law applies to AI systems — but its subject is personal-data processing and consent (for example, disclosing that a chatbot interaction is AI-driven when it's handling someone's personal data under Amendment 13), not a general rule that marketing images or ad copy must be labeled "AI-generated." It's also still a draft, reportedly facing pushback from the Finance Ministry and industry bodies, and hadn't been finalized as of this research.

The closest thing to an AI-content-labeling law sits in a private member's bill introduced in the Knesset in March 2025, which would require broadcast advertisements and sponsored messages containing AI-created or AI-edited content to carry a marking notice if it could be mistaken for an authentic recording. Two limits worth being precise about: it applies to broadcast TV and radio advertising specifically, not digital, WhatsApp or email marketing, and — like most private member's bills — there's no evidence it has passed even a preliminary Knesset reading. Separately, Israel's Consumer Protection and Fair Trade Authority does enforce a general "proper disclosure" duty for paid or sponsored endorsements (the same rule behind "paid partnership" labels on influencer posts) — but no guidance from that regulator specifically addresses AI-generated content, which is itself the useful finding: on AI-content labeling specifically, Israel's consumer-protection regulator has been silent.

So how should a business actually use AI in its marketing?

1. 1Treat every AI-drafted claim about your own business — policies, pricing, eligibility, deadlines — as unverified until a human checks it against the real source. The Air Canada and Cursor incidents above both trace back to exactly this gap: AI stating something confidently about the company's own rules that simply wasn't true.
2. 2Use AI for the first draft, not the final send. The strongest evidence for AI's productivity gain (the Noy & Zhang study above) measured a speed and quality boost on a first pass with human grading after the fact — not an unsupervised, end-to-end replacement for review.
3. 3If you market AI capabilities yourself — "AI-powered," "automated," "smart" — describe only what you've actually tested. The FTC's DoNotPay settlement turned specifically on claims about AI performance that were never verified against a real benchmark before being advertised.
4. 4Know which disclosure rules actually reach you: the FTC's fake-review ban if you have US customers, the EU AI Act's marking duties if you have EU customers, and Meta's or Google's platform-specific ad rules if that's where you advertise — and if you're an Israeli business, know that today it's mostly policy and draft guidance, not binding law, so revisit it rather than assuming it's settled.
5. 5Keep a human as the approval step before anything sends — this is the same principle oToK's own built-in AI tools are designed around: the AI drafts subject lines, full emails, images and even whole automations from a brief, but nothing activates or sends without a person reviewing it first.

Frequently asked questions

Is AI actually being used in small business marketing, or is this still mostly hype?

Both, in a specific way: adoption is real and large (89% of B2B marketers use AI for copy per CMI & MarketingProfs; 48% of SMBs globally per Constant Contact), but only 39% of those AI users in the CMI survey say it's actually improving their performance, and Salesforce found 75% of marketers have adopted AI while 84% still send generic, one-way campaigns. Adoption and results are two different, currently disconnected numbers.

Can a business actually get in legal trouble for what its AI-generated content says?

Yes. Canada's BC Civil Resolution Tribunal held Air Canada liable for its chatbot's false bereavement-discount claim in February 2024, ordering it to pay $812.02 CAD and explicitly rejecting the argument that the chatbot was "a separate legal entity." Separately, the FTC extracted a $193,000 settlement from DoNotPay in 2024–2025 over unsubstantiated "robot lawyer" marketing claims.

Do I have to disclose when my marketing content is AI-generated?

It depends on where your customers are. In the US, the FTC bans fake/AI-fabricated reviews specifically, not AI involvement generally. In the EU, the AI Act's Article 50 transparency duties (machine-readable marking of AI-generated media, deepfake disclosure) became legally applicable August 2, 2026. Meta's own tools auto-label ads with a "significant" AI edit; photorealistic political or social-issue ads already require advertiser self-disclosure now, with Meta's automatic detection for that category starting June 2026. In Israel, there's no binding AI-content-labeling law yet — only a non-binding 2023 policy document and an unpassed 2025 bill limited to broadcast advertising.

How often does AI actually make things up (hallucinate)?

There's no single reliable percentage, because it depends entirely on the task. The most-cited ongoing benchmark, Vectara's Hallucination Leaderboard, measures a narrow task — summarizing a source document faithfully — where top models score around 1.8% and weaker ones over 20%. That number doesn't transfer to open-ended writing or a chatbot answering questions about your business's own policies; treat any flat "AI lies X% of the time" claim with skepticism.

Does using AI for marketing copy actually save time?

The most rigorous evidence is a randomized controlled trial published in Science: 453 professionals given ChatGPT finished writing tasks 40% faster with 18% higher quality scores, with the biggest gains going to the weakest writers first. Vendor-surveyed industry figures point the same direction — HubSpot reports marketers self-estimating one to two hours saved per day, Salesforce reports self-identified "high performers" reclaiming about 8 hours a week using AI agents — but these are self-reported estimates, not independently measured outcomes, and the "8 hours" figure specifically applies only to that high-performing subgroup, not the average marketer.