Marketing teams are using only a fraction of AI’s capability

What is capability overhang?

In his newsletter One Useful Thing , Mollick uses the term capability overhang to describe the gap between what AI models can do and how they\u2019re actually used. He argues that we only use a small part of the models\u2019 abilities, and that we often lack a clear picture of what they can actually do.

He illustrates this overhang with his own experiments. In one, he had an AI model recreate Italian author Umberto Eco\u2019s library in Milan in 3D, including its 27,000 books. The model analyzed films and photographs frame by frame, identified around 5,000 book titles, and stated how confident it was in each identification.

In another experiment, an AI model created an animated trailer for Mollick\u2019s upcoming book. It produced the script, voices, music and sound effects – work that a human team would have needed several weeks to complete.

On a marketing team, the overhang looks more mundane. The same models that write your posts can also review a competitor\u2019s ads, compare their messaging to your own, and map out where you stand out. But not many teams make use of this.

The studies show the exact same pattern

In March, Supermetrics, a company that sells marketing data tools, published a report based on responses from 435 marketers in five countries. 80 percent of them feel pressure to use AI, but only 6 percent have fully embedded AI into their workflows, meaning AI is a fixed part of their daily work and team processes. 87 percent use AI mainly for content and copy.

In Sweden, AI adoption is growing fast and spreading to more and more companies. According to Statistics Sweden (SCB), 35 percent of companies used AI in 2025, up from 25 percent the year before. Among larger companies, the share was 72 percent. The figures show that more companies have access to AI tools, but say nothing about how extensive or advanced that use actually is.

Which human abilities make your team better at using AI?

Mollick\u2019s answer to the overhang is to stop competing with AI in production itself. Instead, the collaboration should build on four human strengths, so that people and AI together can achieve what neither can do alone.

• Expertise. Someone who has mastered their field quickly sees when AI gets it right, and when it gets it wrong. Mollick also argues that an expert\u2019s knowledge affects the quality of what AI produces.
• Versatility. Knowledge from multiple areas helps you understand what to ask AI for. Mollick draws his example from design, but the principle applies just as much in marketing. Someone who understands advertising knows that an AI-generated ad copy also has to fit the platform\u2019s format and stay within character limits.
• Taste. The faster and cheaper production becomes, the more valuable the ability to pick the best result becomes. Much of the time therefore goes into correcting, fine-tuning and judging when something is good enough.
• Agency. Mollick uses the term agency to describe being curious and willing to explore. If you dare to experiment with AI, you can find new opportunities before your competitors do.

Experience using AI tools naturally affects the outcome too. Anthropic, the company behind Claude, reported in March that users with at least six months of experience more often get the result they were after, even when solving the same types of tasks as newer users. The difference was three to five percentage points.

What does this mean for your marketing team?

A team with expertise, versatility, taste and agency already has a strong foundation to build on.

The next step for such a team is to let specialists explore AI within the area where they have deep knowledge. An SEO specialist can tell immediately whether an AI-generated keyword analysis is good enough, and a copywriter notices when the language feels off.

Supermetrics\u2019 report shows that 37 percent of marketers lack a clear AI strategy from management. Agency requires room to experiment, and management needs to create that room. That means you need to:

• Set aside time in the schedule to test AI on real work tasks.
• Decide how the team shares what it learns, for example in a recurring meeting or a shared channel.
• Collect examples of text, images and analysis that meet your standard, and use them as a benchmark when evaluating what AI delivers.

At BBO, specialists in advertising, SEO, tracking and analytics, AI and content work together. That means deliverables rarely rest on a single skill, but on a combination of depth and breadth, just as Mollick argues.

Who shapes the change?

Mollick concludes that change is coming, regardless of how fast the technology develops. What it will look like, however, is an open question. The teams that address the overhang now get to help shape how AI and people work together within their own organization.

A good place to start is an internal question: which of your specialists currently have time to discover how you can make even better use of the tools you already use?