A CMO at a 240-person B2B SaaS forwarded me a research note last week — one of those slides going around about how marketing functions are seeing a productivity lift on the order of 20% from AI when the AI has the right context, and something closer to 4% when it doesn’t. The shape of the finding keeps appearing in 2026 industry research. She asked her Head of Brand where the 20-ish percent was showing up. Her Head of Brand thought about it and said, honestly, she wasn’t sure. She asked her Head of Content. Same answer. Her team had been using the agents every day for six months. The work wasn’t materially better.
She forwarded me the report with the subject line: “What am I supposed to be looking for?”
That’s the question every CMO I talk to is being asked right now by their CEO. And the honest answer is: most marketing leaders are measuring the wrong things. Utilization rate (are people using the agent?). Output volume (how many drafts are we producing?). Time saved on drafts (how many fewer minutes per email?). Those numbers go up, and the work that matters stays exactly where it was.
Before going further, a quick note for anyone whose company hasn’t deployed an agent yet: an AI agent, in this article, is software that uses a language model — Claude, ChatGPT, Gemini — to do real work on your behalf instead of just answering questions. You give it a goal; it takes several steps; it comes back with a draft or a decision. The simplest agent in marketing is the assistant that drafts a campaign email when you tell it the audience and the offer. The most useful ones do this with persistent access to the actual context your team works from — your past emails, your customer calls, your CRM, the decisions you made last week. The difference between “simplest” and “most useful” is the entire subject of this piece.
Here are the three places agentic value actually shows up. They are three different kinds of value, not three flavors of the same thing.
1. The agent makes a hard decision easier.
Drafting an email faster is not the value. Drafting an email faster is the easy part — and you can buy it from a $20/month writing assistant. The value is whether the agent helps you decide which email to send and what it should say.
A decision is the moment where you, the leader, have to choose between two paths and you don’t yet know which is right. Do we lead the V2 launch with the speed angle or the cost angle? Should the next campaign target our late-stage pipeline or our top-of-funnel? Does this customer story make the case better than the one we used last quarter, or are they too similar to ship in the same month?
An agent that has read your last six months of campaigns, your customer calls, your conversion data, and your competitive briefs can make these decisions easier — not by deciding for you, but by surfacing the patterns you don’t have time to find. The Head of Brand who used to spend Sunday night sketching the launch frame on a whiteboard now spends Sunday night reading the agent’s three-option memo and picking the one that’s right. The decision is still hers. The 90 minutes of pattern-finding before the decision is the agent’s.
If your agent isn’t making any decisions easier — if it’s only producing faster drafts of things you already knew how to write — the value is small.
2. The agent’s drafts arrive closer to what you’d actually ship.
Camille Trent runs content at dbt Labs. She told me last week about the change she noticed in her agent over the spring. Her team had given it access to the per-post outcome data for two years of dbt Labs content — pageviews, time on page, conversion, what got engagement, what bombed. The change wasn’t immediate; it took a few weeks for the drafts to settle into the new mode. But once they did, the agent had stopped feeling like a writing assistant. It was something closer to a strategy assistant.
The agent didn’t get smarter. It got opinionated in the direction her team’s outcomes had already pointed it. The drafts started carrying memory of which topics had worked, which framings had repeated, which headlines hadn’t. “I edit them differently now,” she said. “Less for taste, more for the things only I would catch.”
The metric here is sometimes called drift: a draft that’s far enough off your team’s voice and direction that an IC has to rewrite or throw out. When drift goes down, you know the agent is reading the right context. When it doesn’t, you know it isn’t. A Series-D fintech CMO told me at a Pavilion breakfast last month that she stopped tracking utilization a quarter ago and started watching how often her Head of Brand was sending drafts back for full rewrites. She couldn’t tell me an exact number; she said it was visibly less often than at the start of the year, and her Head of Brand had stopped flagging it as a problem at their 1:1s. “That’s the only signal I trust now,” she said.
If your agent’s drafts still need a full rewrite before they ship — if your Head of Brand is doing the same edits she was doing six months ago — the value is small.
3. The team’s time moves toward work that compounds.
This is the lens most marketing-leader articles skip, and it’s the one that matters most at the org level.
Some marketing work compounds. Positioning that’s right for a year keeps paying off. Customer insight that surfaces a real shift in the buyer means every campaign for the next two quarters lands better. A brand voice that holds across thirty pieces of content trains the audience to recognize you. This is the work that, done well, makes a CMO’s tenure visible in the company’s revenue line three years later.
Other marketing work is maintenance. Weekly newsletters. Campaign copy variants. Social repurposing. Status reports. This work has to happen, but it doesn’t compound — last week’s newsletter doesn’t make next week’s better.
An agent doing real work for you should be moving your team’s time from maintenance toward compounding. If your Head of Content used to spend 12 hours a week on the newsletter and now spends 4, the right question isn’t “great, we saved 8 hours.” The right question is: what is she doing with the 8 hours? If she’s spending them on customer interviews and the next quarter’s content thesis, the value is enormous. If she’s producing more weekly newsletters, you’ve bought yourself nothing — you have more maintenance, not more compounding.
If you can’t name what your team is doing with the time the agents are saving them, the value is invisible.
What all three lenses depend on
You’ll have noticed the same thing showing up in all three. The agent that makes decisions easier has read your campaigns, calls, and conversion data. The agent producing drafts closer-to-shippable has read your outcome history. The agent freeing your team for compounding work has done it by handling the maintenance well — which it can only do because it knows what your maintenance looks like.
The value comes from the context the agent has access to. Not the model. Not the prompt. The context. McKinsey’s 22%-vs-4% number isn’t about which model you bought — same models in both columns. It’s about whether the model can read your team’s actual context every time it works, or whether someone has to paste the context in.
This is also why most marketing teams are stuck. In the 2026 marketing-stack surveys I’ve seen, roughly four in ten orgs running agents have no integration between the agent and their CRM, marketing automation, or call recordings. Whatever the agent produces, it produced with the context of a brief and a prompt — not with the context of the team. That’s the low end of the productivity range. That’s the dashboard going green while the work stays the same.
The function head is the curator-owner. This is a gift.
A lot of marketing leaders read the above and think great, I need to centralize all of this into one place that I personally maintain. Don’t.
Your Head of Brand decides what voice rules go in. Your Head of Demand Gen decides what campaign learnings go in. Your Head of Content decides which past content gets fed in with which outcome data attached. Your Head of MarOps decides what data warehouse slices are exposed. Each function head curates their slice of the context the agents read from. You don’t curate everyone’s slice.
This is a gift, not a delegation. You stop being the bottleneck — every time a function head learns something, the team’s agents reflect it within the day, without an offsite slide and without you in the loop. The feedback you’d give, already in the draft — because the function head’s feedback is already in the substrate the draft was built from.
The risk of doing this wrong — and the discourse is full of leaders doing it wrong — is centralizing all the context to the CMO seat. That seat doesn’t have the bandwidth, and the function heads who got the seat by being good at curating their function feel correctly bypassed when the leader tries to do it for them. The function head curates; you set the direction.
How to start
One move this week. Pick one decision your team is currently making — a positioning call, a campaign choice, an editorial bet — and ask whether the AI tool the team is using has the context to help make that decision well. Not “could it draft something afterwards.” Could it help you decide?
If the answer is no, name what’s missing. Usually it’s one of three things: the past outcome data, the recent customer calls, or the decisions the team has already made and the agent hasn’t read. Whichever it is, that’s the first thing to wire up.
The ongoing cost of this is small but real: about five to ten minutes a day of curation per function head, on average. Hide that cost and the context decays; name it and it compounds.
The piece I’m building, Pomegranate, is one way to make the substrate practical — it’s the shared context layer that carries your team’s voice and current direction across every agent in marketing, product, sales, and CS. There are other ways to do this; the point isn’t the tool. The point is that until the agents can read what your team actually works from, the value question your CEO is asking you won’t have a satisfying answer.
The CMO who forwarded me the research note sent a follow-up about a month later. She’d picked one input — the past campaign outcome data — and worked with her Head of Demand Gen to pipe it into their agent setup. It took longer than she expected; there were schema questions, a permissions argument with the data team, a week where they thought it was hooked up and it wasn’t. But the drafts that came back from the next campaign were different. Not shippable as-is — her Head of Brand was still editing. But the edits were the kind a senior reviewer should be making: brand judgment, sequencing, calls only she could make. Not throw this out and start over.
She wrote: “I can see the shape of the value now. We’re not there. But I know what I’m building toward.”
That’s the answer to her original question. Not a clean number she could put in a deck for the CEO. Something more useful: a direction she could feel.