Two reports landed on enterprise IT and marketing desks this week that, read in sequence, describe the same problem from opposite ends. Harmonic Security analysed roughly two million classified AI session minutes across major providers and found that, on accounts users created for themselves, most of the activity was work. Two-thirds of what staff are doing on free personal ChatGPT, Gemini and Claude logins is corporate use: drafting work emails, summarising meeting notes, debugging code. The companies paying for the enterprise licences have, at best, partial visibility into the other half of the usage.
A day earlier in Aurora, Colorado, Gartner used its Marketing Symposium keynote to tell CMOs that pilot-stage worry about whether to use AI is over. According to the analyst, 98 percent of CMOs are using or piloting AI tools, and 15 percent of marketing budgets are now going to AI initiatives. The new diagnosis is that most marketing teams have settled into a stage Gartner calls "AI competent," where the tools work, the workflows are humming, and only one in three teams is seeing the returns they hoped for. Gartner's term for it is the competency trap.
The two findings line up uncomfortably well. Harmonic's vice president of marketing, Michael Marriott, described the shadow-AI version of it bluntly. "The CEOs view they've got an AI strategy they can force down on people, while the rest of the business has a completely different idea of what it is, and they're just using whatever tools they want to do their jobs." His most quotable line for the leaderboard of dangerous personas: "the worst defenders were the sales and marketing teams, who were using better tools to do their work." Pipeline data, customer records and M&A material are flowing through accounts that the employer cannot audit, and if the salesperson leaves, the dataset goes with them.
Gartner's framing, presented by analysts Kristina LaRocca-Cerrone and Jay Wilson, sits above that operational picture. AI, they argued, is "a magnifier of disruption and organizational fault lines." Weak collaboration, reactive strategy and order-taker marketing models do not improve when you bolt copilots onto them. They break faster. Eighty percent of CEOs surveyed told Gartner they want AI to deliver "transformative change," not operational improvement. The average marketing function, by Gartner's count, has invested in four major AI use cases and projects AI-driven automation of marketing work to double, to 36 percent, by 2028. Activity is rising. Advantage is not.
Put the two pictures together and you can see why. Buyers in the competence trap are running more variations through more channels, faster, on the same fragile economics they had before. Meanwhile the people actually getting productivity out of AI in the building are doing it on personal accounts, with tools the CMO has not approved, in ways nobody is measuring. The "AI strategy" the boardroom is buying is mostly licences. The "AI work" the company is doing is mostly invisible, and a lot of it is leaving the building when employees do.
Marriott's prescription is unglamorous but worth quoting. The teams getting ahead of shadow AI, he said, are the ones "eating a bit of humble pie" and asking employees what tools they actually want. The same teams use that intelligence to consolidate licences and reclaim spend; a Copilot seat is 39 US dollars per user per month, and a lot of those seats sit idle while the work happens elsewhere. Gartner's prescription is almost a mirror image: stop using AI to do more of the same thing, and use it instead to "change the altitude" of the work. Reallocate budgets in real time, orchestrate full campaigns, attack problems the team would not have tried by hand.
Neither side calls out the awkward implication out loud. The CMO who is busy hitting Gartner's AI-confident tier is, very often, the same CMO whose sales team is quietly running the actual pipeline through a personal Claude account. The competence trap and the shadow AI trap are not separate problems. They are what happens when the people procuring AI and the people using it stop talking to each other.