Most Vendors Selling "AI Agents" Are Not Selling Real Agents. Your Sales Team Needs to Ask 4 Questions

A Nylas survey from March covering 1,026 developers and tech leaders reported that a mere 4% permit an AI agent to take action with no human sign-off at all. In a separate count, Gartner places the number of true agentic vendors at roughly 130 among thousands. Some offerings sold as AI agents are in fact older chatbots or automation tools wearing a fresh name — a practice Gartner labels agent washing.

A Media Professor Just Put Every AI Agent Pitch on Trial

Shelly Palmer, a Syracuse University media professor who runs a consultancy, published an essay on November 1, 2026 called What Agentic Marketing Actually Means. His argument: the phrase has no settled meaning, and the companies selling it cannot agree on the riskiest question of all — whether a person signs off before an agent's work goes live.

Palmer holds the advanced media in residence chair at Syracuse University's Newhouse School of Public Communications. He also runs The Palmer Group, a consultancy that advises large companies on AI strategy, governance and implementation.

He opened with a line that should make any owner who has sat through a software demo shift in their chair. "'Agentic marketing' currently means whatever the vendor selling it needs it to mean," he wrote.

Two days later, Anthony Katsur, who runs IAB Tech Lab, the ad industry's technical standards body, praised the essay on X. He said his group also sees "APIs being relabeled as agents." He added: "A new name does not create agency."

The number that should worry you is not new. Gartner counted roughly 130 genuine vendors among the thousands claiming agentic capabilities. It also expects more than 40% of agentic AI projects to be scrapped before the end of 2027. Those figures trace back to a Gartner press release from June 25, 2025.

Palmer borrows Gartner's term for it: agent washing — the habit of selling chatbots, copilots and automation tools under the name agents.

So what do you actually do about it? He gives you four questions.

The Four Questions That Sort a Real Agent From a Renamed Chatbot

Palmer's test asks who sets the steps, who approves the actions, whether it learns, and who the agent works for. He says that first question draws the line between automation and real agency.

The approval question carries the most risk. Palmer says there are three honest answers: a person signs off on everything, a person signs off only on exceptions past a set threshold, or nothing needs approval at all. Each is a legitimate product, he argues, with a different risk profile, compliance load and staffing need.

The learning question is where I would push hardest. If the system never changes what it does, Palmer says, it is a script, not an agent. A real one acts, checks the result, and changes what it does next because of it.

The last question is the one I would put in writing. Palmer's warning: when the same business sells the ad space and runs the agent buying it, the agent works for the platform.

Now here is the part that should make you slow down before your next renewal.

The Vendors Selling You This Stuff Do Not Agree With Each Other

Palmer compared how seven platforms define the same phrase, and the definitions do not line up.

Palmer describes Salesforce as marketing its stack under Agentforce Marketing, centered on campaign agents that handle planning, creation, personalization and optimization. Adobe's Agent Orchestrator, he says, is a reasoning layer that orders purpose-built marketing agents within the Experience Platform. HubSpot's newest release cycle centered on content and nurture agents that run a plan-create-measure cycle. In September, Amazon rebranded its consolidated advertising platform as Amazon Ads Agent.

Then the split widens. Bloomreach calls it handing a system a goal rather than a script. Klaviyo's version keeps a person signing off on everything that goes live. Amazon's guide leaves continuous human oversight out. Profound's definition gives the agent every step, even though its own launch copy says marketers control every approval.

Palmer's conclusion is blunt: the industry cannot agree on whether a person signs off before work goes live, and some vendors cannot agree even with themselves.

That is not a marketing problem. That is a risk problem, and it lands on whoever owns the phone number your leads call.

Why This Is Landing on Your Desk Right Now

The companies building this stuff are already contradicting themselves in public. When Amazon launched Ads Agent at its unBoxed 2025 event on November 11, 2025, campaigns went live only after advertisers signed off on the recommendations. On September 29, 2026, Amazon put its sponsored ads and its demand-side platform under one brand, Amazon Ads Agent, and dropped the Amazon DSP name. A PPC Land report on October 29 said the conversational agent was still in closed beta, open only to eligible advertisers in 25 countries.

Buyers are pulling back. A survey of 2,145 senior leaders by KPMG, reported by PPC Land in August 2026, found that 49% had paused, delayed or narrowed their agent rollouts once costs outweighed the value. Only 7% could point to a return on investment.

Approval is still the norm. A survey by Nylas, reported in March, asked 1,026 developers and technology leaders and found that only 4% let agents act with no sign-off at all.

Accountability is thin. A study of 801 executives by WSJ Intelligence and Code and Theory, publicised in October, found that at one in four big US companies, nobody is accountable for coordinating AI. Only 18% hold decision-making AI agents to the same accountability and reporting rules as human leaders.

And the rulebook is still open. IAB Tech Lab named its agentic initiative AAMP on February 26, 2026, and on October 23 it extended the comment period on the AAMP 3.0 draft to November 10, 2026.

What This Means for an Owner Who Sells (My Read)

Here is my read, and it is opinion, not reporting.

Picture a pool builder with three crews and two salespeople. Leads arrive by web form, Facebook message and text. A vendor demos an AI agent that answers inquiries in seconds and books estimates. The demo is smooth. The question you have to ask is the second one: who approves what it sends? If the answer is nobody, you just handed your brand to a script that can quote a price, promise a start date, or argue with a homeowner about a permit. If the answer is that a human approves everything, you bought a faster inbox, not an employee. Both are fine purchases. They are not the same purchase.

Picture a fitness studio selling trial memberships. Your front desk is slammed at 6 p.m. and the phone rings out. An agent that texts back within a minute and books a tour is worth real money. An agent that invents a membership rate is a problem with a logo. Same word on the proposal. Completely different business.

The third question is the one I would press hardest on. Does it learn? If the tool cannot tell you what it changed and why, you are not buying an agent. You are buying a script that never learns, and you will be the one explaining it to a customer.

How I Would Use This This Week

1. Open the last three AI or automation proposals you were sent. Find the section on approvals. If there is no such section, that is your answer.

1. Write down your own definition of agent before you take another demo. Palmer's advice: write your own definition down and require anyone using the word — vendors included — to stick to it.

1. Ask the four questions in writing, and make the rep answer in writing. Who sets the steps. Who approves. Does it learn. Who does it work for.

1. Pick your approval tier on purpose. If you sell anything with a price, a schedule or a legal commitment, put a human on exceptions at minimum. Only 4% of the people surveyed let agents act with no sign-off at all.

1. Test it on your worst lead, not your best. Take a real inquiry from last month that went nowhere and run it through. Watch what the tool sends before it sends it.

1. Write the follow-up yourself first, then see if the tool can beat it. Here are three openers I would steal.

• Pool builder, first text back: greet them by name, confirm you got their note about the pool, ask the rough yard size, and ask whether they want to swim this summer or next. Promise a real number this week.
• Fitness studio, missed call: apologize for missing them, say the front desk gets busy at 6, and offer two specific class times so you can hold a spot.
• Any trade, second follow-up: check back once, tell them to say so if the timing is wrong and you will stop, and ask for the one thing holding them up.

1. Put a name on it. That same study found that a quarter of big US companies have nobody whose job is to coordinate AI. Do not be that company at any size. One person owns the tool, its output and its kill switch.

Mistakes I See Owners Make

• Buying the word instead of the behavior. Agentic on a slide tells you nothing. The approval line tells you everything.
• Letting the vendor define the term. If you do not write your own definition, theirs wins by default.
• Skipping the learning question. If nobody can show you what the system changed and why, you own the mistakes.
• Treating a renamed chatbot as headcount. A tool that runs the same play forever does not get better at your customers.