AI for marketing: useful tasks and required checks

Marketing looks like the easiest place to put AI to work: most of the job is words, images and numbers, and generative tools produce all three quickly. The catch is that marketing is also the part of a business that makes public promises, quotes customers, handles personal data and reports results to people who make budget decisions on them.

None of the rules that govern those activities were suspended when AI arrived. Regulators in the US, UK and EU have said so plainly, and their guidance is practical: it shows which checks to keep and where the hard lines are.

This guide maps six marketing tasks where AI helps, then gives the human check each one needs across five areas: claims substantiation, disclosure, brand and intellectual property, privacy, and measurement. If you are just setting up AI for your team, start with our guide to how to start using AI at work and come back here for the marketing-specific rules. This is an explainer, not legal advice; your sector and jurisdiction may add requirements.

The rule underneath every task: AI does not change who is responsible

Start with the principle that makes every other check necessary. The US Federal Trade Commission's long-standing policy on advertising substantiation requires a reasonable basis for objective claims before an ad runs.

“Advertisers and ad agencies have a reasonable basis for advertising claims before they are disseminated.”

— Federal Trade Commission, Policy Statement Regarding Advertising Substantiation (1984).[1]

The same statement says that when an ad expressly cites support, such as “tests prove” or “studies show”, the firm must have at least that level of support. It also covers implied claims: advertisers are expected to substantiate reasonable interpretations of their ads, not only the literal words.[1] A language model that writes “clinically proven” into a draft has created a claim the business must be able to back up. It has not supplied the backup.

The UK advertising regulators reach the same place from a different direction. The Committee of Advertising Practice (CAP), which writes the UK ad codes, describes them as media-neutral: they apply however the content was made. In its 2026 advice on AI-generated ads, CAP pointed to an ASA ruling about an ad assembled with Google's automated tools.

“Even though the ad was produced using Google’s automated tools, the advertiser was still primarily responsible for ensuring compliance with the CAP Code.”

— CAP News, AI and Deepfakes: Four Things Advertisers Need to Know Before They Hit “Run” (11 June 2026).[2]

The FTC's enforcement record points the same way. In its September 2024 sweep of AI-related cases, the FTC charged DoNotPay with claiming its AI service could substitute for a human lawyer. The complaint alleged the company had not tested whether the chatbot's output matched a lawyer's. The proposed order would bar the company from claiming it can substitute for a professional service without evidence.[3] The lesson for marketers is narrower than “AI is risky”. A claim about what AI does is still a performance claim, and it needs the same evidence as any other.

Six tasks where AI earns its place

These are the tasks where generative tools save real effort. For each, the useful question is not whether AI can do it, but which check catches the problem the tool cannot see.

AI Chronicle

Research and synthesis

AI is good at turning a pile of material into a first map: summarising competitor pages you have saved, clustering customer feedback, drafting questions for interviews, or explaining an unfamiliar category. The risk is that a summary reads as fact even where the model has filled a gap.

Google's own guidance for publishers puts it bluntly.

“Generative models don't retrieve facts, but predict a likely sequence of words based on their training data.”

— Google Search Central, Guidance on using generative AI content on your website (updated 1 October 2026).[4]

The check: open the source behind every fact you intend to reuse. If the tool cites a page, read the page. If it cannot point to one, treat the statement as a lead to verify, not a finding. Our explainer on why AI produces plausible but false answers covers the failure in detail. The second check is privacy: research prompts are where customer emails, support transcripts and CRM exports tend to get pasted in.

Briefs and outlines

A brief is where AI can save the most thinking time and cause the most downstream damage. Turning a product sheet, audience notes and last quarter's learnings into a structured brief is a sensible use. The danger is that a confident brief seeds claims into every asset written from it.

The check: keep an approved claims list, and require that every factual claim in a brief comes from it, with its evidence attached. If the brief will go to creators or influencers, it should also say how they must disclose the relationship (see disclosure, below). A brief is also the place to write down how success will be measured, before anyone sees results.

Drafts: copy, email and social posts

Drafting is the obvious use and usually a good one. A model can produce first versions of product descriptions, emails, landing-page sections and social posts in your house style, given good examples.

The check: substantiation first. Mark every objective statement in the draft: numbers, comparisons, “fastest”, “eco-friendly”, “AI-powered”, health or performance effects. Each needs evidence you hold today. Then the brand check: tone, accuracy of product details, and anything implied by the overall impression rather than the literal words. Under the FTC policy, implied claims count.[1]

Variants: headlines, ad versions and images

Generating twenty headlines or ten ad layouts for testing is where AI's speed compounds. Image tools add product mock-ups, lifestyle scenes and backgrounds. For a primer on how those images are made, see our explainer on how AI image generators work.

The check: brand and IP matter most here, because variants multiply quickly and images can depict people, places and results. Two questions catch most problems. Does any version imply more than the original, approved claim? And does any image show a real person, or something that could pass as a real scene? CAP's example is pointed: an AI-generated image showing a cosmetic effect that does not reflect real-world results is likely to mislead, even with an AI label attached.[5]

SEO support

AI can help with keyword research, outlines, internal-linking suggestions, title and description drafts, and structured data. Google says it does not penalise content for being AI-made as such.

“Using AI doesn't give content any special gains. It's just content.”

— Google Search Central Blog, Google Search's guidance about AI-generated content (8 February 2023).[6]

The line Google draws is about purpose and value. Its spam policies define “scaled content abuse” as generating many pages primarily to manipulate rankings rather than help users, and list “using generative AI tools or other similar tools to generate many pages without adding value for users” as an example.[7]

The check: accuracy, including the parts readers barely see. Google's guidance says fact-checking applies to metadata too: title elements, meta descriptions, structured data and image alt text.[4] The second check is value. Ask whether a page would still help a reader if search traffic never arrived.

Analysis and reporting

Models can draft commentary on campaign results, summarise survey responses, propose segments and help write queries. This is easy to over-trust. Language models are not reliable calculators, and fluent commentary invites causal language (“the new campaign drove sign-ups”) that the data may not support.

The check: every number in a report should trace to a query or a calculation done in code, and every causal claim should rest on a test designed to show cause. Our tutorial on how to build a growth reporting agent shows one way to enforce that lineage. Privacy is the other check: analysis often involves customer-level data, which brings in data protection law.

The five checks in detail

Claims substantiation

Substantiation is the check that applies most often, because almost every customer-facing output contains some claim. Three practical points follow from the sources.

First, the evidence must exist before publication. The FTC's policy is explicit that firms lacking a reasonable basis before an ad runs violate the FTC Act, and that later evidence is not a substitute.[1]

Second, claims about AI itself are claims. The FTC's 2024 cases involved businesses promising AI-powered results they could not show.[3] In the UK, CAP advises that the codes apply to AI claims like any other.[5]

Third, a disclosure does not rescue a misleading claim.

“Disclosure alone is very unlikely to mitigate the harm caused by a fundamentally misleading message.”

— CAP News, Disclosure of AI in Advertising (29 May 2025).[5]

A workable routine: keep a claims register with the wording, the evidence, its date and the owner; let AI draft freely but only from the register; and have a named person check every objective claim against it before anything ships.

Disclosure

Disclosure in marketing covers two different questions. Is there a commercial relationship the audience would want to know about? And does the audience need to know that content was made or altered by AI?

Commercial relationships. The FTC's Endorsement Guides say a connection between an endorser and the brand should be disclosed clearly and conspicuously when a significant minority of consumers would not expect it and it would affect how they weigh the endorsement. The FTC's staff guidance adds that an endorsement must reflect the endorser's honest opinion.[8] The 2024 reviews rule defines “clear and conspicuous” strictly: in an interactive medium the disclosure must be unavoidable, and it is not clear and conspicuous if a consumer must click a link or hover over an icon to see it.[9] AI does not change these rules. It does make it easier to produce influencer briefs, pre-written posts and review requests at scale, so the disclosure instructions belong in the template.

AI involvement. Here the rules differ by jurisdiction:

Two details in the European Commission's Article 50 guidance matter for marketers. First, the human-review exemption for text is not a formality: “Superficial, solely formal, or procedural checks (e.g. spell-checking or grammatical correction) are not considered to be human review or editorial control.” Second, a deployer cannot rely on the machine-readable marking a tool provider embeds in its outputs to meet the visible deepfake disclosure duty.[11] The Commission also notes that a deepfake can resemble something that “can plausibly exist”, so a realistic but invented scene may still qualify, depending on context and audience.[11] Article 50 has applied since 2 August 2026. The Commission says the only grace period, to 2 December 2026, covers the machine-readable marking duty for generative systems already on the market before that date; deployers' disclosure duties are not deferred.[11]

Reviews and testimonials: the hard line

This is the area where AI use is most clearly off limits for marketing teams. In the US, the FTC's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials (16 CFR Part 465) was announced in August 2024. The FTC's announcement said the rule addresses reviews by people who do not exist, “such as AI-generated fake reviews”, or by people with no actual experience of the product.[12]

The rule text makes it unlawful for a business to write, create or sell a review or testimonial that materially misrepresents that the reviewer exists, used the product, or had the experience described.[9] It also bans compensation or incentives conditioned on reviews expressing a particular sentiment, positive or negative, and it bars buying fake social media indicators such as bot-generated followers or views for commercial purposes.[9]

UK law reaches the same point. Since April 2025, the Digital Markets, Competition and Consumers Act 2024 has made fake reviews and concealed incentivised reviews a banned practice. CMA guidance defines a fake review as one that “purports to be, but is not, based on a person’s genuine experience,” and lists buying reviews generated by software such as bots as an example. It also expects publishers to correct review counts, ratings, rankings and “AI-generated review summaries” affected by banned reviews.[13]

What AI can legitimately do here is the unglamorous work: draft a neutral request for honest reviews sent to all purchasers, summarise genuine reviews for internal learning, or flag suspicious patterns for a person to investigate. What it cannot do is write a review, a testimonial or a “customer quote” that no customer gave. One enforcement footnote: the FTC's 2024 order against Rytr, a writing tool with a review-generation feature, was set aside in December 2025, and the FTC said the original complaint did not meet the FTC Act's requirements.[14] That decision concerned a tool provider's liability. It did not change the reviews rule, which applies to businesses that create or buy fake reviews.

Brand and intellectual property

Brand checks are familiar: voice, accuracy, consistency and taste. AI adds two IP questions.

Can you own what the tool made? The US Copyright Office's January 2025 report on copyrightability concluded that AI outputs are protected only where a human author has determined sufficient expressive elements. That can include perceptible human-authored material or creative arrangement or modification of the output, “but not the mere provision of prompts.”[15] For marketers, this means an asset generated from a prompt alone may not be protectable in the US. That matters if you plan to build a mascot, logo or campaign look around it. The Office also confirmed that using AI as an assistive tool, or including AI material in a larger human-authored work, does not bar protection for the human contribution.[15]

Are you using someone else's? The Office's AI report runs in three parts: digital replicas (2024), copyrightability (2025) and a pre-publication Part 3 on training AI on copyrighted works (May 2025).[16] CAP warns that a convincing AI “celebrity endorsement” is likely to breach the code, and that while personality rights and IP are outside the ASA's remit, advertisers should take legal advice before using a public figure's likeness.[2]

The check: keep records of how key assets were made and what human work went into them; do not promise clients exclusivity in purely AI-generated assets without legal advice; never generate a real person's likeness or voice without permission; and run brand-critical images through the same rights review as stock or commissioned work.

Privacy

Marketing data is personal data, and pasting it into an AI tool is processing. The UK Information Commissioner's Office (ICO) guidance on AI and data protection is a useful model even outside the UK. It says organisations must break down each distinct processing operation and identify a purpose and lawful basis for each one.[17]

The ICO has answered several questions that map directly onto marketing work:

• Drafting customer replies with AI. Legitimate interests may be a suitable basis if the organisation can show the use is necessary and proportionate. Special category data needs an additional condition.[18]
• Using an AI tool for existing tasks. If the purposes do not change, existing lawful bases can still apply, but privacy notices should explain how the tool is used. On accuracy, the ICO says: “When using an AI tool, organisations must check that the outputs are accurate, and take steps to correct any inaccurate information.”[18]
• Sharing data so a provider can train its model. That is data sharing in its own right, needing its own legal basis.[18]
• Marketing messages. The UK's electronic marketing rules (PECR) apply on top of data protection law, including to marketing sent by social media direct message.[18]

The EU AI Act adds one marketing-specific duty. Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them.[10] This matters if a vendor offers to read shoppers' faces or voices for “sentiment”.

The check: decide which data categories may go into which tools, in writing; prefer tools whose terms rule out training on your inputs, or check the data-sharing basis if they do not; update privacy notices; and keep special category data out of general-purpose tools unless the legal basis and condition are documented.

Measurement

Measurement is the check marketers most often skip, because it does not feel like compliance. It belongs on the list for two reasons.

First, reported results are themselves claims. If marketing tells customers, investors or clients that AI improved outcomes, the FTC substantiation principle applies to that statement too.[1]

Second, AI shifts work rather than removing it. Drafts get faster; review, fact-checking and approvals take time that rarely appears in the before-and-after story. Our guide on how to measure AI productivity without hiding review work sets out how to count the whole workflow.

The check: decide the metric and the comparison before the test; keep calculations in code or the analytics tool, not in model prose; avoid causal language without a test designed to show cause; and record review time alongside drafting time.

How much review each output needs

The five checks do not all apply with equal force to everything AI touches. A useful way to allocate review is by proximity: how close an output gets to the public, to factual claims and to real people.

AI Chronicle

Internal notes need only the author's source check; public copy needs an editor and evidence for every objective claim; endorsements and realistic depictions need disclosure, consent and rights clearance. Fake reviews are not a higher rung. They are excluded.

A worked example (hypothetical): launching a self-watering planter

This example is invented to show the checks in sequence. The company, product, claims and outputs are hypothetical. Nothing here describes a tested tool, a real campaign or a measured result.

Imagine a small online plant shop, which we will call Fernhill, selling in the UK and EU. It is launching a self-watering planter and wants AI help with the launch.

The “Outcome” column describes what the checks are designed to catch, not results from a real launch. The pattern matters more than the details: AI did most of the first-draft work, and each step had a named person checking a specific risk.

Getting started without a policy document

A team does not need a long AI policy to begin. It needs six decisions in writing: an approved claims register with evidence and an owner for each claim; a rule on which customer data may go into which tools; a hard line that AI never writes reviews or testimonials and incentives never depend on sentiment; a disclosure template for creators and affiliates, plus a labelling rule for realistic synthetic media shown to EU audiences; a named reviewer for each rung of the ladder; and a measurement plan set before each test, with review time counted.

Questions readers commonly ask

Do we have to label everything AI helped write?

Not as a blanket rule in the UK, according to CAP.[5] In the EU, Article 50 requires disclosure of deepfakes and of AI-generated text published to inform the public on matters of public interest, unless that text has had genuine human review and someone holds editorial responsibility.[10][11] Platforms can add their own rules, such as Google Merchant Center's metadata requirement for AI-generated product images.[4] Label where the audience would otherwise be misled, and never rely on a label to fix a misleading claim.

Can AI write review request emails?

Yes, if the request goes to purchasers generally, asks for honest opinions and does not tie any incentive to a positive or negative rating. What AI must not do is write the reviews or testimonials themselves.[9][13]

Can we use AI-generated people in ads?

Often, with care. CAP says disclosing that an influencer is AI-generated can help avoid a misleading impression.[5] In the EU, a realistic depiction that would falsely appear authentic may be a deepfake requiring a visible label.[11] A real person's likeness or voice needs their permission, and a fake endorsement from a recognisable person is likely to breach the UK code.[2]

Does Google penalise AI-written content?

Google says it rewards helpful content however it is produced, and that AI content gets no special gains or penalties as such.[6] Mass-producing pages mainly to manipulate rankings is a spam policy violation regardless of method.[7]

The next step

AI is useful at every stage of marketing, as long as each output has an owner who knows which check it needs. The part teams most often get wrong is the last one: numbers that do not trace back to a query, and commentary that claims cause without a test.

Our tutorial on how to build a growth reporting agent applies these principles to weekly marketing and growth reporting: read-only data access, calculations in code and a person who approves the report.

About these sources

All pages accessed 16 October 2026. Quotations were checked against the pages linked. Regulatory guidance is summarised for general information; it is not legal advice.

Related reading

• AI for ecommerce: practical starting workflows
• What workplace AI studies actually show