The pitch decks all say "AI-powered" now. The org charts do not. In a June 2026 survey of 101 enterprises, 71 percent said a quarter or fewer of their deployed "agents" were true multi-step orchestrated workflows. The rest were single-prompt chatbots with a new label (VentureBeat Pulse Research, 2026).

That same mislabel is what B2B marketing buyers now face in every agency pitch. Publicis Groupe just reported 87 percent of net revenue coming from "AI-powered marketing services" while Publicis Sapient, its transformation arm, declined mid-single-digits in the same quarter (Adweek, 2026). The pitch is outpacing the delivery. The category grew because it was easier to relabel services than to rebuild them.

The trap for buyers is that AI-washing looks great in a demo. A chatbot on a landing page. A prompt library screenshot. A slide that says "80 percent faster" with no visible mechanism. It is the same failure mode enterprise-tech buyers hit when Gartner counted the market: only about 130 of the thousands of self-described agentic-AI vendors are actually real (Gartner, 2025). Everyone else is a legacy assistant or an RPA workflow with a new label.

We built Moving Parade as an agentic firm from day one, and we have sat in enough evaluations now to see the pattern. The buyers who can tell the difference in a first call ask three questions. This is that diagnostic.

What is AI-washing in a B2B marketing agency?

The underlying delivery has not changed. That is the tell. AI-washing is "AI-powered" applied as a wrapper on the same shop you would have hired two years ago: chatbot demos labeled as agents, before-and-after slides that hide the mechanism, headcount that scales with revenue. Gartner estimates only about 130 of the thousands of self-described agentic-AI vendors are real.

The label moved before the architecture did. Every media agency now has an "AI practice." Every consultancy has an "agentic offering." Very few have rebuilt delivery so that agents author work a human used to author, with human review as the gate rather than the labor. The MIT NANDA study of 300 enterprise AI deployments found 95 percent of generative-AI pilots delivered no measurable P&L impact (Fortune, 2025). Not because the models were bad. Because the deployments were retrofits: an assistant sitting next to the same team, running the same tasks, on the same headcount curve.

The pattern reads the same at the holding-company level. When a holdco reports that 87 percent of its net revenue comes from AI-powered services while its actual transformation arm shrinks in the same quarter, the "AI-powered" category is a reporting construct. It is what the finance team can label without changing what the delivery team builds. The label expanded to cover work the label did not touch.

A pattern that plays out constantly in B2B marketing agency evaluations: the deck talks about agents, the SOW talks about hours, and the case studies talk about outcomes with no mechanism attached. The AI is in the marketing of the agency, not in the marketing the agency does.

The physical signature of the difference is what the agency can put on a screen.

What does an agentic marketing partner actually look like?

An agentic partner can show you the agents, the audit trail, and work a human never touches. The signature is verifiable output. Skills invoked, tasks routed, deliverables authored by an agent with human review as the gate rather than the labor. Without a working system to show, the AI has not moved from the pitch deck into delivery.

An agentic firm is legible on three surfaces. The org chart shows a small number of humans and a large library of skills or agents, each with a named domain. The delivery process shows work flowing from an agent-authored draft into a human reviewer, then out to the client, with the artifacts of each pass preserved. And the invoice math shows margin that comes from headcount staying flat while scope grows, not from billable hours that scale with revenue.

Moving Parade is four humans and more than eighty invocable skills. The account audit runs across three ad platforms and produces the same 145-check report a senior analyst would have spent two days writing. The optimization log writes itself while campaigns run. The comment engine drafts founder-voiced replies overnight. Every artifact carries the audit trail. The reviewer is human. The labor, in almost every case, is not.

This is what "AI as delivery" reads like in practice. It is not a Copilot bolted on to a marketing team; it is agents handling work the org chart used to demand more people for. If a firm cannot show you which agent handles which task, on which schedule, with which review gate, the AI-powered claim is a marketing decision, not an operating one. See our primer on building AI marketing workflows without replacing your team for the workflow architecture buyers should be looking for.

The three-question diagnostic below is the shortest path from claim to evidence.

What three questions should you ask an agency about its AI?

Three questions expose the difference in a first call. Can you show me an agent handling a task you would otherwise pay a human to complete? Where in your org chart does the AI live, delivery or marketing? What breaks if your headcount stays flat next year? Answers show you the architecture instead of the sales deck.

The first question forces a demo of the actual system. Not a landing-page chatbot, not a prompt library. An agent that produces an artifact a human on that team used to produce. The audit that senior analysts used to write. Ad copy variants a copywriter would have drafted overnight. A weekly report that once required a media manager's Monday morning. If the answer is "we use AI to help our team be faster," the AI is a Copilot. If the answer is "here is the skill, here is the invocation, here is the artifact, here is the reviewer's edits," the AI is the delivery layer.

The second question probes the org chart. AI-washers put the AI in the marketing function: a chatbot on the site, an AI-branded service line, a "prompt library" that sales shows in the demo. Agentic firms put the AI in the delivery function: the account team is smaller and the skill library is larger. Ask which team grew last year and which team stayed flat. The honest answer is a diagnostic on its own.

The third question tests the economics. An agentic firm's headcount curve decouples from its revenue curve. Ask the agency what breaks if it takes on twice the client load next year without hiring more delivery FTEs. If the answer is "we would hire," the AI has not changed the underlying economics of the business, which means it has not changed the underlying economics of your engagement either.

Buyers who ask these three questions get a very different call than buyers who ask about credentials or case studies. The next section is what the wrong answers usually look like.

What are the red flags of AI-washing when hiring a B2B marketing agency?

The red flags are structural. Chatbot demos framed as agents. Before-and-after screenshots that hide the mechanism. "AI-powered" applied as a wrapper on the same team you would have hired two years ago. When AI appears in the pitch but not in the invoice math or the org chart, the transformation is cosmetic. The headcount curve is the honest read.

The most common red flag is a case study without a mechanism. "We used AI to generate 40 percent more creative variants" is not a claim about delivery architecture. It is a claim about output volume. The follow-up question is: was the incremental volume authored by an agent, or by a copywriter using a Copilot to draft faster. If the answer is Copilot, the firm still needs the copywriter. The economics have not moved.

The second flag is the AI-branded service line without an AI-native delivery model. A firm sells you an "AI SEO audit" that is still executed by two SEO analysts in a shared doc. A firm sells you "AI-powered creative" that is still a creative director briefing junior designers who use Midjourney. The name is new; nothing in the underlying delivery changed. Gartner's finding on the roughly 130 real vendors (Gartner, 2025) is the wide-market version of this pattern. The AI-vendor market is roughly 3 percent real. The AI-agency market is not likely much higher.

The third flag is a linear headcount curve. If the firm's revenue doubled and its team doubled, the AI did not compound. AI that compounds is AI that lets a team of four do the work a team of forty used to be assigned. AI that scales with hires is a productivity tool for humans, and productivity tools are useful, but they are not the transformation the pitch is selling. We wrote about the underlying economics in Agencies Resist Productization. Meanwhile, You Can Buy $100K Software in One Call. It applies here.

The fourth flag is a substitution test failure. Read the "AI-powered" line in the deck and swap the agency's name for any competitor. If the sentence still makes sense, the claim carries no information. It is category language, not agency evidence.

The good news for buyers: none of these flags require a technical evaluation. They require asking to see the system rather than the pitch.

What should you look for in a B2B marketing partner in the AI era?

Look for delivery architecture, not credentials. Ask to see agents in production. Compare the AI story to the headcount curve, then test with a real problem on the first call. Match what they show you against what AI-search buyers now audit. The Day-One list only includes vendors whose evidence holds up under machine reading.

Two related patterns are worth watching. First, roughly 90 percent of B2B buyers purchase from a shortlist that forms before formal evaluation begins (Bain, 2026, origin HBR 2022). That list now includes vendors surfaced by AI search, and AI search reads the same evidence a human would look for: mechanism, specificity, verifiable claims. An AI-washer's category-language proof does not survive that reading. An agentic firm's mechanism-rich case studies do. If a firm's own site does not pass the substitution test, its inclusion in the Day-One list is at risk. Second, agentic AI is the layer buyers should evaluate the firm on, not the layer buyers should evaluate a specific tool on. Gartner projects that more than 40 percent of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025). A partner whose agentic model is real is more durable than one whose agentic tools are.

The comparison below is the shortest way to hold two candidates up to the same criteria at once.

One move: rewrite one line in your next RFP. "Show us an agent handling a task we would otherwise pay a human to complete, with the audit trail." Send it to three finalists. Watch what comes back. The gap between the demos will tell you more than a scoring matrix.

For a longer version of the buyer diagnostic, we published Eleven Questions Most Buyers Forget When Evaluating a Demand Gen Agency. For the architectural taxonomy that sits behind this article, see Holding Company vs Boutique vs AI-Native B2B Agencies in 2026. And for the AI-search side of the same story, How AI Search Builds Your B2B Shortlist covers the Day-One list mechanic buyers now inherit whether they choose to or not.

Frequently asked questions

What questions should I ask an agency about their AI?

Three questions cover most of the ground. Can you show me an agent handling a task you would otherwise pay a human to complete? Where in your org chart does the AI live, delivery or marketing? What breaks if your headcount stays flat next year? Any one answered vaguely is a signal.

How can I tell if an agency's "AI-powered" claim is real?

The claim is real if the firm can show you the mechanism: which agent handles which task, on which schedule, under which review gate, with which artifacts preserved. The claim is likely marketing language if the response is a screenshot of a chatbot, a prompt library, or an AI-branded service line with no underlying delivery change.

Is "AI-native" the same as "AI-powered"?

They point at different things. "AI-powered" describes tools used by a team of humans to work faster. "AI-native" describes a delivery model where agents author work a human used to author, with human review as the gate rather than the labor. A firm can be AI-powered without being AI-native. The difference shows up in headcount economics.

Can a large holding-company agency actually be agentic?

In principle, yes. In practice, the reporting shows the opposite pattern: holdcos are labeling more revenue as AI-powered while their transformation arms decline. That is not proof no holdco is agentic; it is proof the label is running ahead of the architecture. Test the claim with the same three questions above.

What are the biggest red flags of AI-washing?

Case studies without a named mechanism, AI-branded service lines with a non-AI-native delivery model, a linear headcount-to-revenue curve, and category language that survives the substitution test. Any single flag is worth a follow-up question. Two or more together is a firm to keep off the shortlist.