Marketing’s AI Gap Is Widening (and It Isn’t the Models’ Fault)

Summary

Why the distance between what AI can do and what GTM teams are allowed (and organized) to do is compounding in your competitors' favor, plus five ways to close it. The real constraint isn't model capability anymore, it's unresolved IT, legal and platform governance plus workflows that were never built with AI in mind, and closing that gap takes deliberate ownership and documentation, not another pilot.

The volume of chatter about AI among marketing executives and practitioners keeps growing, and more and more of it is about governance and permissions. This includes mandated systems, unapproved platforms and why the tools people use (or want to use) on their own time can’t touch the systems they work in every day.

The workarounds are increasing corporate risk and making AI’s true potential for your organization farther away from reality.

I’m writing this from day one (day zero?) of MAICON 2026 (The Marketing AI Conference) in Cleveland. More than 2,000 GTM professionals are here, including ~150 CMOs at this morning’s AI for CMOs Summit (recap of my panel on 2027 budgeting here).

Throughout the day it’s been fascinating to hear how companies are embracing AI…as well as those unfortunately falling farther behind (whether they realize it or not).

Some of the people in those conversations see the problem acutely. They’ve built real workflows themselves, then a security review or a single “approved” platform shut them down. They feel every month of delay compounding, because they know what the teams who did get connected are building in the meantime.

And then there are the ones who are blissfully unaware that their competitive advantage is vanishing.

For a couple of years the honest answer to “why isn’t AI doing more for us” was that the models weren’t good enough yet. That answer stopped being true months (if not quarters) ago.

Today’s constraint has moved into three gaps: between what the models can do and what GTM teams actually do with them, between how fast the technology moves and how fast your company’s governance adapts (or doesn’t) and between how work really gets done and how well it’s designed for AI to do any of it.

Marketing is using a fraction of what’s already there

Ethan Mollick recently called the first gap the capability overhang: the distance between what current models can do and what almost anyone is doing with them. His fix is largely individual, which is the half of the problem a motivated person can solve alone on a Saturday.

The marketing numbers show how wide that gap runs. Only 28 percent of organizations in Jim Lecinski’s research with Kellogg are rethinking their workflows to accommodate AI, and just 10 percent of marketing leaders say they’re recouping major value from their investments in it. OpenAI’s own enterprise data shows the same pattern inside a single license. Among monthly active users, 19 percent have never tried data analysis, 14 percent have never used reasoning and 12 percent have never used search.

From what I see, a lot of marketing teams are still using AI mostly to draft. First drafts, variants, summaries. That’s useful, and it’s why 85 percent of marketing and product users in the same OpenAI report say they’re executing campaigns faster. But that’s also the shallow end of the pool.

The deeper capability sits one layer down. Nearly every system your team runs on already exposes an API, and a growing number ship MCP servers (the emerging standard for plugging AI assistants into other software). That includes analytics, your CRM, marketing automation platforms, data and intent sources and more. A marketer cleared to connect those systems is running analysis, debugging campaigns and building new ones against live data, while the marketer who isn’t is pasting a CSV into a chat window and hoping the LLM can read their mind.

What holds that second marketer back is usually permission. OpenAI’s data hints at how common that is: roughly one in four enterprises still hasn’t enabled the connectors that give the model secure access to company data at all.

Governance hasn’t kept up

Getting connected, not individually but for your full team or organization, runs through three corporate constraints, and marketing hits all three.

IT policy.

Connecting a model to a production system means a security review, and security reviews are built (for good reasons) to slow down exactly the kind of change you’re asking for.

Legal.

Most marketing work worth automating touches customer data. Some of the caution around that is of course well justified, but a fair amount of it is absorbed from headlines rather than from the company’s own risk profile.

The mandated platform.

Somebody chose an enterprise LLM license, usually inside a companywide software bundle your team had no vote in. That choice is now the ceiling on what your team can build.

Legal’s caution is rational enough. There’s still no comprehensive federal AI law in the United States, and the 109 AI laws enacted across 29 states in the first half of 2026 make it even harder to understand. When the rules aren’t formalized and clarified, the defensible move for company counsel is to assume the strictest version of every rule and wait. Nobody gets fired for the campaign that didn’t get built, or for the data leak that never happened.

So people route around it. UpGuard surveyed 1,500 security leaders and employees across seven countries, and more than 80 percent reported using AI tools their company hadn’t approved (including nearly 90 percent of the security professionals). Seventy percent said they were aware of colleagues inappropriately sharing sensitive data with those tools.

A CMO this morning at the MAICON Summit said every one of his employees has a second laptop they use for real AI work. That’s terrifying.

The risk is real and the policies exist for a reason, even as I’d argue those same policies have become the binding constraint on marketing’s output.

The work itself was never designed for AI

The third gap is the least visible, and it was there long before AI showed up.

Many marketing organizations quietly run on institutional knowledge: who owns which step, where the current version of the plan lives, what “done” means for a campaign and how a new request gets weighed against what’s already committed. Shockingly little of it is written down (let alone followed by all), and for a long time that was survivable. Smart people filled the gaps. They asked around, reconstructed campaign status from Slack threads, meetings and memory, and guessed right often enough that little visibly broke (at least until the team got big enough that things started to).

Now we’re asking AI to “guess” at what we want inside processes that only exist in people’s heads. It guesses with a fraction of the context the person who used to fill that gap had, and at a speed and volume that turns small inconsistencies into big ones. Unclear ownership, inconsistent planning and no shared view of what’s in flight were already costing teams before AI, and AI magnifies every one of them.

We call this orchestration debt: the cumulative burden of misaligned processes, fragile or incomplete workflows, disconnected systems and inconsistent governance across your marketing and revenue engine. Like financial debt, it accrues interest, and AI raises the rate.

Gartner found 84 percent of marketers report high levels of “collaboration drag” in cross-functional work, and organizations with high drag are 37 percent less likely to hit their revenue goals. When almost everybody is dragging, dragging starts to feel like the job.

And the people at the top are often the last to feel it. We’ve been running our marketing orchestration self-assessment with a growing number of marketing leadership teams, and a pattern keeps showing up when the CMO and their key leaders score the same operation independently. Increasingly, CMOs rate their team’s collaboration materially higher than the leaders closest to the work do. On strategic alignment and planning, leaders on the same team often land more than two points apart on a five-point scale.

I don’t think either side is wrong. The CMO sees outcomes and a team that shows up, while the leaders below see the Slack threads, meetings and memory it took to produce them. That gap matters more with AI than it ever did, because the CMO is usually the one approving the AI-forward programs, based on a view of the operation the people doing the work don’t share.

Meanwhile, the gap is compounding in your competitors’ favor

The cost of waiting shows up as work that never happened rather than as a line item that went wrong.

In OpenAI’s data, frontier firms generate about twice as many messages per seat as the median enterprise and seven times as many messages to custom GPTs, the tools teams build to standardize their own workflows. McKinsey’s latest State of AI survey found 80 percent of people using AI in their roles say it has improved their productivity, yet only 37 percent of respondents attribute any bottom-line impact to it and just 6 percent qualify as high performers.

Individuals are getting real value, and a small group of companies have figured out how to turn that into business value. If your competitor is in that group and you aren’t, every month your request sits in the queue is a month they spend building on live data you still can’t touch. And I’d argue it compounds, because every system they connect and every process they write down makes the next approval easier and the next build faster.

What connected under real governance looks like

Five things that move it faster than another pilot:

1. Pick one system. I know the pace of innovation makes picking a single LLM hard right now. But you can’t afford to wait. Picking a system with rich capabilities (and that’s more than a couple LLMs right now) connected properly beats six tools bolted onto the old process, and it gives you a reference case for the next request.
2. Bring the governance with you. “We can’t use that” almost always means “we can’t put that data there,” and those are different problems with different fixes. Walk into legal with the scope already written: which objects and fields, read or write, which named users, what gets logged and retained, what happens to the data at the vendor. A request that arrives with controls attached gets a different answer than “can we use this?”
3. Write down how the work actually moves. Marketing orchestration is a disciplined approach to how marketing work gets done, from planning all the way through execution. Before you point an agent at a process, map who owns each step, where status lives and how a new request gets weighed against what’s already committed. Then have your CMO and key leaders score the operation independently and look at where they disagree before you look at the average. (Our free marketing orchestration self-assessment uses 12 realistic workplace scenarios and takes about eight minutes, if you want a starting point.)
4. Put a number on the wait. Nobody measures what the queue costs in campaigns not run or analysis not done. Bring that number, even a rough one, and frame it against what a connected competitor is doing in the meantime. It turns the conversation from a favor you’re asking into a tradeoff the business is making.
5. Give someone the job. In a lot of orgs, nothing gets approved because getting things approved is nobody’s role. Name a person, give them the time and let them own the relationship with your legal and security teams.

Change here is genuinely hard, mentally and structurally (let alone untangling the license bundles and data agreements already baked into everything you run). And remember, your colleagues writing the restrictive policies are doing their jobs, often with rules written for an earlier era of technology.

But this conversation can’t keep getting deferred to the next planning cycle, because your competitors are already pulling ahead.

This post originally appeared on Matt Heinz’s Substack.