How to Spot AI Washing in Agency Marketing

“AI-powered” appears in a lot of agency proposals. Sometimes it means a real system built around your data. Sometimes it means a chat subscription and new words for the same deliverables. From outside, the two read the same. This checklist gives you the questions that tell them apart.

We build AI integrations through our AI integration service, so weigh that. Every fact below about the FTC, the SEC, the UK ASA and the EU AI Act was checked against their own pages in October 2026.

Short answer: AI washing is claiming more about AI than the work supports: saying a service uses AI when it doesn’t, overstating what it does, or crediting it with results it didn’t produce. The FTC, the SEC and the UK’s advertising regulator have all acted on unsupported AI claims under existing law. To test an agency, ask it to walk through one deliverable step by step, show where a person reviews the AI’s output, explain what happens when the AI is wrong, and name the outcome metric it will be measured on, with a baseline. A real integration survives those questions. A label doesn’t.

What Regulators Have Said About AI Claims

As of October 2026, from the FTC, SEC, ASA and European Commission:

Sources: SEC press release, FTC Operation AI Comply, ASA/CAP report on “AI” as a marketing term, FTC on accessiBe, FTC on Workado, Workado final order, EU AI Act Article 50.

None of these involved a marketing agency. They show how regulators read an AI claim: as a fact that needs evidence.

What AI Washing Actually Means in an Agency Context

AI washing is presenting AI involvement as proof of value when the involvement is superficial, overstated or unrelated to the results you’re buying.

The clearest regulatory example is the SEC’s. On March 18, 2024, it settled charges against two investment advisers. It said Delphia claimed to use AI and machine learning on client data, capabilities the SEC found it didn’t have, and that Global Predictions called itself the “first regulated AI financial advisor” among other false and misleading claims. Both settled without admitting or denying the findings (source).

In agency marketing, it’s usually subtler: a repositioned proposal, a renamed workflow, a testimonial about “faster turnaround” with no word on whether the work performed.

The Difference Between Using AI Tools and Building AI Systems

An agency that uses a chat assistant to draft copy is using an AI tool. One that has built something that pulls your brand guidelines, past campaign data and product details into every brief, and checks the output against rules, is closer to a system. One is a subscription. The other takes engineering and testing.

Both are legitimate. The question is which one you’re being sold, and whether the promises match it.

When “AI-Powered” Is Just a Rebrand of Old Automation

Email sequences, social scheduling and A/B testing were automated long before generative AI. Renaming the same rules-based workflow “AI-powered” is a vocabulary change, and can be misleading: UK advertising guidance says not to claim a product uses AI or has AI features if it doesn’t (source).

The tell: ask what specifically changed in the workflow after the agency “added AI.” If the answer is vague, or describes a tool you could subscribe to yourself, you have your answer.

7 Things to Check in Agency Proposals and Pitches

None of these proves an agency is overstating its AI. Each one is a reason to ask a follow-up question before you sign.

Vague Efficiency Claims With No Baseline Data

“AI lets us work faster” means nothing without a baseline. Faster than what? If the agency can’t say what the timeline was, what it is now and how that shows up in your deliverable, it’s marketing copy, not a metric.

Real efficiency claims have units, a before-and-after and a record you can check.

“Human Oversight” Mentioned Once, Never Defined

Every agency with an AI workflow should be able to explain exactly where a person reviews AI output before it reaches you. “We always review everything” is a reassurance, not a process.

Ask: who reviews it, against what criteria, and is it tracked how often drafts get changed? If the answers aren’t quick and specific, oversight is an intention, not a step.

No Explanation of What Happens When AI Is Wrong

Language models can produce confident, plausible errors, and any agency using AI in client work will meet them. The question is whether it has a documented response: how errors are caught, who fixes them and what changes afterwards. If the proposal doesn’t mention failure modes, ask.

Testimonials About Speed, Not Results

“They turned around our content in half the time” is a testimonial about the agency’s operations. It tells you nothing about whether the content worked. Scan every case study and testimonial for outcome metrics: traffic, conversion rate, revenue, rankings, cost per lead.

Speed is a feature. Results are the product. If testimonials only mention features, ask why.

Ethics Language With No Enforcement Mechanism

“We’re committed to responsible AI use” appearing once in a proposal is decoration. What does it mean in this agency’s work? What happens if a tool skews targeting or produces something off-brand? Who is accountable? Principles without a named owner and a process are positioning, not policy.

Proprietary AI With No Technical Specifics

“Our proprietary AI system” is worth questioning. Which model does it run on? Was it fine-tuned, and on what data? What does it add to a commercial model from OpenAI, Anthropic or Google? What happens to data you share with it?

A real proprietary tool survives these questions. Deflection, or an NDA that stops you understanding what you’re paying for, is a red flag. The detail matters: the FTC alleged Workado said its detector was built on a wide range of content, while the model was trained or fine-tuned to classify only academic writing (source).

Case Studies With Nothing to Verify

Generic phrasing, round numbers, no client name and no attribution don’t prove a case study is invented, but they’re reasons to check. Ask for a client reference and the analytics behind a number.

Reviews deserve the same care. In Operation AI Comply, the FTC alleged Rytr’s AI service let subscribers generate detailed reviews that would almost certainly be false (source). It also alleged accessiBe formatted third-party articles and reviews to look independent without disclosing its connections to the reviewers (source).

The Five Questions to Ask Before You Sign

These work in a first meeting. You don’t need technical knowledge, only patience for a real answer.

Map the Claim to a Workflow Step

Ask the agency to walk you through one deliverable, say a campaign report or a content brief, and show exactly where AI is involved. Not “across our workflow.” One deliverable, step by step.

If they can’t do it for one example, the AI is a positioning layer, not an operational one.

Ask for the Failure Mode

“When your AI tool gives you bad output, what does that look like and what do you do?” A good answer includes a specific example, the correction process and the safeguard that now stops it reaching you. A claim that it has never happened suggests nobody is checking.

Demand Measurement of Incrementality, Not Activity

Activity metrics, such as pieces produced, posts scheduled or emails sent, are easy to inflate with AI without improving outcomes. Ask which outcomes the agency is responsible for, how they’ll be measured and what the baseline was before the engagement started.

Incrementality means what changed because of this work, compared with what would have happened anyway.

Ask Who Owns the Output and the Audit Trail

If an agency uses AI to produce copy, designs or strategy documents for you, who owns them? What happens to your data inside their tools? Can you see which parts were AI-generated and which were written or edited by a person?

This matters for intellectual property, for compliance and for carrying the work on after the engagement ends. Put it in the contract.

Find Out When They Last Updated Their AI Knowledge Sources

Every model has a training cutoff. Unless the agency feeds in current sources, fast-moving topics like ad platform rules can get confident, outdated answers. Ask how the agency keeps its tools current and how it checks AI output against up-to-date sources.

What Legitimate AI Integration Actually Looks Like

Red flags are half of it. The other half is knowing what a good answer sounds like.

Clear Boundary Between Automation and Human Judgment

A legitimate integration says what AI handles, what people handle and why. The boundary usually sits where AI errors are costly or hard to spot: brand tone, strategy, factual claims, compliance. There, a person catching a mistake early costs less than a client catching it late.

An agency should be able to describe that boundary for your project in plain words.

Measurable Baselines and Defined Improvement Targets

Before any AI-assisted work starts, there should be a written baseline: where you are now, which metric you want to move and what counts as success at agreed checkpoints. That’s how you find out whether the AI helped. Without it, any claim of “AI-driven results” can’t be tested either way.

The FTC sets a similar bar: Workado must now hold, and keep, competent and reliable evidence before making accuracy claims (source).

Transparent Tooling (Not Always “Proprietary”)

A trustworthy agency can tell you which tools it uses, what those tools do with your data and what happens to your outputs if the engagement ends. “Proprietary” is sometimes legitimate, but it should never mean “we can’t explain what it does.”

The same test applies to any build, from an AI workflow to a hand-coded WordPress site: you should be able to see the code, the tools and the settings, and take them with you.

Related Posts

• AI integration red flags in the sales process: more warning signs before you sign
• What to ask an AI agency before signing: contract and deliverable questions
• AI integration case studies: what they show: how to read published results
• Realistic AI productivity gains: what the research supports

Frequently Asked Questions

Is it AI washing if an agency uses ChatGPT for content but doesn’t disclose it?

It depends on what they claimed and what you paid for. If the agency sold “expert human-written content” and delivered lightly edited AI drafts, that’s misleading whatever the tool. If they said AI helps with drafts and people do the final edit, it’s a workflow choice. The UK ASA says there’s no blanket UK legal requirement to disclose AI use in ads (source). The EU AI Act’s Article 50 duty covers AI text published on matters of public interest, with an exception for human-reviewed text under editorial responsibility (source).

What’s the difference between AI washing and just exaggerating in a sales pitch?

AI washing credits AI with capabilities or results it doesn’t deliver, or claims AI is involved when it isn’t. The UK ASA draws a line: vague “productivity” claims may be treated as obvious exaggeration, while objective claims about what a product can do need evidence (source). “Our AI makes you more productive” is fluff. “Our AI is 98 percent accurate” is a claim someone has to prove.

Can I ask an agency to prove their AI claims before signing a contract?

Yes, and you should. Ask for a live walkthrough of a deliverable that uses their AI workflow, case study data with a baseline you can check, and a client whose results came from the AI work specifically. If the answers stay general, the claim probably is too.

Are there legal consequences for agencies that make false AI claims?

They can be. The FTC’s Chair said in 2024 that “there is no AI exemption from the laws on the books” (source), and its 2025 accessiBe and Workado orders were about unsupported AI claims in marketing. In the UK, the CAP Code applies to business-to-business ads too (source). Whether a specific claim crosses the line depends on the facts, so get legal advice for your case.

How do I evaluate an AI integration proposal without a technical background?

Ask the agency to explain one deliverable step by step without the words “AI-powered,” “intelligent” or “automated.” Ask for the failure example and the fix, and the metric they’ll be held to. Someone who has built something can describe it concretely. If the explanation only works at a high level, the work may only exist at a high level.

What should a legitimate AI proposal actually include?

The specific tools and their data policies, the workflow steps where AI is and isn’t involved, the human review process and who owns it, baseline metrics and targets, and ownership terms for AI-generated outputs. If any are missing, ask why before signing.