AI in marketing: what the evidence supports, and how to run a pilot

This piece informs the decision of where to use AI in marketing, under what rules, and how to find out whether it pays. It keeps to evidence: controlled studies, standards and the law as of October 2026.

• Two kinds of AI Predictive models score and forecast (who is likely to buy or leave, when to send, what to bid). Generative models produce text, images, audio and code. Most teams now use both, often inside tools they already have.
• What the evidence shows In a controlled experiment published in Science, professionals given ChatGPT took 40% less time on writing tasks and produced work rated 18% higher in quality¹. Customer support agents with an AI assistant resolved 14% more issues per hour, with the biggest gains for novices^2, 3.
• What it does not show yet These studies measure time and rated quality on defined tasks. They do not measure sales, brand or long-term effects, and a field experiment with consultants found capability to be uneven across tasks that look similar⁴.
• Rules and risks NIST's generative AI profile names risks such as confabulation, harmful bias and intellectual property⁶. The EU AI Act's transparency duties, including deepfake disclosure, have applied since 2 August 2026⁷, and the FTC's rule bans fake reviews, including AI-generated ones⁹.

Predictive AI is not new in marketing: lead scoring, lookalike audiences and automated bidding have run on statistical models for years. What changed is generative AI, and now agents that take several steps on their own. Among 413 martech leaders Gartner surveyed in mid-2025, 81% were piloting or had fully implemented AI agents; 45% of those said vendor-offered agents fell short of the promised business performance, and half said their organisation lacked the technical and data readiness to deploy them⁵. The gap between promise and readiness is the first thing a pilot should test.

Writing tasks. Shakked Noy and Whitney Zhang ran a preregistered experiment with 453 college-educated professionals and randomly gave half of them access to ChatGPT for occupation-specific writing tasks. Time taken fell by 40% and rated output quality rose by 18%, and the gap between stronger and weaker writers narrowed¹.

Service conversations. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied customer support agents at a large software company who were given a generative AI assistant. In the working paper, issues resolved per hour rose by 14% on average, by 34% for novice and low-skilled workers, with minimal effect for the most experienced and skilled²; the Bureau of Labor Statistics' review notes that the gain came mostly from newer, less experienced agents³.

Uneven capability. In a field experiment with about 750 consultants at Boston Consulting Group using GPT-4, Fabrizio Dell'Acqua and colleagues described a "jagged technological frontier": tasks that look equally difficult to experts can fall on very different sides of what the AI can do well⁴. The practical lesson is that results from one task do not transfer to the next; each use needs its own test.

Read these findings for what they are: time and rated quality on defined tasks, in controlled conditions. None measured whether customers bought more or thought better of a brand. For that, use the methods in measuring effectiveness: an A/B test or a holdout.

NIST's Generative AI Profile (NIST AI 600-1, July 2024), a companion to its AI Risk Management Framework, lists twelve risks that generative AI creates or makes worse⁶. Five matter most to marketers.

• Confabulation, NIST's term for confidently stated but false content, often called hallucination. It is mostly a problem for text; invented facts, prices, claims or quotations are the marketing risk⁶.
• Harmful bias. NIST gives the example of text-to-image models that under-represent women, racial minorities and people with disabilities when asked to depict professions such as chief executives or judges⁶. Check every set of generated people against your audience.
• Intellectual property. NIST lists intellectual property among the risks, including infringement of third parties' rights⁶. Who owns generated material, and whether it may infringe, depends on the jurisdiction and the tool's terms; ask your legal adviser before generated assets become part of the brand.
• Brand safety and consistency. A model does not know your message house or your claims register. Ground it in approved material and review the output; see brand safety and consistency and adaptation.
• Data privacy. NIST lists data privacy too⁶. Customer personal data goes into an AI tool only under a contract and a lawful basis that cover it; see consent, identity and cookies.

The EU AI Act. Article 50's transparency obligations have applied since 2 August 2026. People must be told when they are interacting with an AI system, such as a chatbot, unless it is obvious; deployers must disclose deepfakes and certain AI-generated text published to inform the public on matters of public interest; providers must mark generated content in a machine-readable way, with a grace period until 2 December 2026 for systems already on the market before August, a rule the Commission describes as resting on the AI Omnibus proposal⁷. The machine-readable marking a provider embeds does not by itself meet the deployer's duty to disclose⁷. The Commission's voluntary Code of Practice on transparency of AI-generated content, published in June 2026, has a section for providers on marking and one for deployers on labelling; signing it is voluntary, and not signing does not in itself mean non-compliance⁸.

Reviews and testimonials. In the United States, the FTC's final rule of August 2024 bans fake reviews and testimonials, including reviews attributed to people who do not exist and AI-generated ones, and lets the agency seek civil penalties against knowing violators⁹. Enforcement priorities move: in December 2025 the FTC reopened and set aside its 2024 order against Rytr, an AI review-writing service, citing the administration's AI Action Plan¹⁰. The order is not the rule; never publish a review or testimonial that a real customer did not give.

Disclose by default when it would matter to the audience: a synthetic person in an advertisement, a chatbot that could be mistaken for staff, an influencer post made with AI (see influencer disclosure).

NIST organises its suggested actions under the AI Risk Management Framework's four functions: govern, map, measure and manage⁶. In a marketing team that means a short written policy, owned by someone with authority:

Put the policy beside the approval workflow in marketing operations and governance, name the roles in the integration RACI, and add AI incidents (a false claim, a biased image, a leaked prompt) to the crisis communications checklist.

Write the business case before the pilot, with the rule that will decide it ("scale if time per brief falls and error rates do not rise; stop otherwise"). Keep humans accountable throughout, and plan the effect on outcomes, not only on effort, with the measurement framework.

An illustrative business, not a real company: a consumer electronics brand with a small content team that writes product descriptions in English, Arabic and French.

• Task First drafts of product descriptions from the approved specification sheet, in three languages.
• Baseline For four weeks the team logs time per description, the number of factual corrections at review and a quality score from the brand lead.
• Pilot For the next four weeks, half the new products are drafted with an approved AI tool grounded in the specification sheet and the message house; the other half the usual way. Every AI draft is reviewed by a native-speaking editor.
• Rule Scale if time falls and corrections do not rise in any language; stop for any language where they rise.
• What the team does not do It does not generate customer reviews, synthetic spokespeople or claims that are not on the specification sheet.

• Starting from the tool, not the task A licence with no named task becomes shelfware, the same failure as in building a martech stack.
• Measuring effort only Faster drafts are not more sales; measure the outcome too.
• Trusting fluent output Confabulation sounds confident⁶; facts, prices and quotes need checking against a source.
• Assuming one success transfers The frontier is jagged⁴; test each new use.
• Taking vendor claims as results 45% of martech leaders with agents in use found them short of what was promised⁵.
• No disclosure policy until a regulator or a journalist asks for one.

1. List where AI is already used in your team, including features inside existing tools.
2. Write the one-page policy: approved tools, data rules, human review, claims checks and disclosure.
3. Choose one frequent, well-defined, low-risk task and record its baseline.
4. Run a pilot with a control group and written success and stop rules.
5. Review every output for errors, bias and brand fit, and log what you find.
6. Check disclosure duties in each market you serve, starting with the EU's Article 50 rules for deepfakes and chatbots.

• The evidence: Noy and Zhang in Science¹, Brynjolfsson, Li and Raymond² with the BLS review³, and Dell'Acqua and colleagues on the jagged frontier⁴.
• Adoption and readiness: Gartner's 2025 survey of martech leaders⁵.
• Risks and governance: NIST's Generative AI Profile⁶.
• The rules: the Commission's Article 50 FAQ⁷ and Code of Practice⁸, the FTC's fake reviews rule⁹ and its Rytr decision¹⁰.
• Related: building a martech stack, marketing operations and governance, the creative brief template and the content primer.