BCG surveyed 300 CMOs this year, and e27 published a piece on the results on 10 November 2026. The piece describes a room of marketing leaders where nearly every hand goes up when asked if AI has transformed their function. Ask what their agents did on their own last week, and the room goes quiet. e27 calls that gap the headline of the survey and builds a framework called DEFY around where agentic AI belongs in the marketing funnel.

I could only read the feed summary of the article, not the full text. So I’ll stick to that gap and say nothing about the framework’s details.

The claim is easy to make

Saying AI has changed your marketing costs nothing. The AI could be drafting captions, summarising reports or suggesting subject lines, and all of that counts as “using AI”. It is also a different thing from an agent that does a piece of the job without anyone prompting it.

“Agentic” has become the word that covers both. A vendor adds a chat window to a dashboard and calls it an agent. A team starts using a writing assistant and tells the board it has adopted agentic AI. The label spreads because it is flattering and hard to disprove.

What I ask first

I ask whether there is a real need. Not “what can it do?” but “what problem does this solve that I couldn’t already solve, and is it worth adding?”

Then I ask what actually changed in a normal week. If the answer is that people now open an AI tool a few times a day, that is a change in habit. It is a fair thing to report, but it isn’t an agent running part of your funnel.

The BCG gap tells me many leaders can answer the first question and not the second. They know AI is in the building. They can’t point to a task that now happens without a person starting it.

My test

Name one workflow the agent runs end to end in your systems.

End to end means it starts, does the work and finishes, and the result shows up in a system you own. A lead comes in, the agent checks it against your records, sends the right reply, books the slot and updates the CRM. Or a customer asks where their order is, the agent looks it up in the store and answers with the real status.

If you can name one, write down what it touches, what it is allowed to commit to, and who sees the result. If you can’t name one, you have bought a label. That isn’t a disaster. It just means you have a tool that helps people work, and you should judge it as that.

What this looks like for a normal business

Take a dental clinic with a front desk of two people. The clinic buys a tool sold as an “AI agent for patient engagement”. Six weeks later the owner is asked what it does. The honest answer might be that it suggests replies and the front desk still copies them into WhatsApp and the booking system by hand. That is a drafting aid. It may save a few minutes a day, and that has some value, but nobody should call it an agent.

Now take the same clinic with one workflow that really runs. A patient messages asking for a cleaning next week. The system checks open slots, offers two, takes the choice and books it, and the front desk sees the booking on their calendar. They didn’t touch it. You can point to that and say what it did last week.

The second one is much narrower than the pitch from the first vendor. It is also the only one a business can check, because there is a booking in the calendar or there isn’t.

Where this ties to what I build

This is the same reason I describe AeroChat the way I do. It is chat infrastructure that connects to your own data systems and backend processes, so it can perform actions, not just answer questions. The question I want an operator to be able to answer about any tool, mine included, is what it did in their systems last week.

What to do

Before you buy another tool with “agent” in the pitch, do three things.

  1. Pick one workflow in your funnel that is repetitive and has a clear finish, such as lead follow-up, order status questions or booking confirmations.
  2. Ask the vendor to show that workflow running in your systems, with your data, and not a demo account. If they can’t, treat the tool as an assistant and price it that way.
  3. Set the limits in the system before it goes live. Decide what the agent can commit to on money, promises and dates, and make sure the platform enforces those limits instead of leaving them to a prompt.

Then check it after a week. Count what it did on its own. If you can’t count anything, you know where you stand.

Fewer tools used deeply will tell you more than many tools used shallowly. One workflow that runs is worth more than a deck full of agents.

Source: e27: The DEFY framework: Where agentic AI actually earns its place in the marketing funnel