Most marketing teams adopted AI the same way. Someone tried it and it saved time. Then it spread. Very few sat down and decided which jobs it should do and which it should never touch. That missing decision is why so much marketing now sounds the same.
The research gives a better way to draw the line.
AI helps a lot, until it does not
In 2023, Boston Consulting Group ran a controlled experiment with more than 750 of its consultants. On creative product innovation tasks, participants using GPT-4 performed about 40% better than a control group. Around 90% of them improved.
On a different task, a business problem that required weighing evidence the model handled badly, the AI users did 23% worse than those working without it. The model produced confident answers that sounded right. People trusted them.
The same study found a quieter cost. The diversity of ideas across the AI group fell by 41%. BCG described the output as “individually better but collectively repetitive.”
Three findings, one lesson. AI is strong inside a certain zone and misleading outside it, and even inside that zone it pulls everyone toward the same answers.
Where the time savings really come from
Marketing leaders are seeing returns. Gartner’s 2025 survey of 402 CMOs found that generative AI was delivering value mainly through time efficiency, cited by 49%, and cost efficiency, cited by 40%. Just 27% pointed to the capacity to produce more content or handle more business.
That pattern matters. The safest gains are in doing existing work faster, not in letting AI decide what the work should be.
A practical split for marketing teams
We help clients draw a simple line between production and judgment.
Give AI the production work. Transcribing interviews and pulling quotes. First drafts of routine pieces from a clear brief. Resizing and reformatting assets for different platforms. Generating headline variations to test. Summarizing long research for a human to read. This is where the hours go, and where errors are easy to catch.
Keep judgment with people. Choosing what the brand believes and says. Deciding which story to tell and whose voice tells it. Interpreting data about your own customers. Final editing for voice and accuracy. Anything with legal or reputational risk.
The BCG results map neatly onto this. Production tasks sit inside the zone where AI shines. Strategy and interpretation sit closer to the kind of problem where the tool produced confident errors.
Protect the voice on purpose
Even good AI drafts drift toward a recognizable style. Polished, balanced, a little hollow. Readers are learning to spot it, and once they do, they trust it less. Edelman’s 2026 research on brands found that 88% of people say trusting a brand is important or critical when they buy. A brand that sounds machine-made spends that trust every time it publishes.
A few habits protect the voice.
Start from real material. Feed AI your transcripts and field notes instead of asking it to invent. The output inherits the specificity of what you gave it.
Keep a house style list. Write down phrases your brand never uses and patterns that signal generic writing. Check every draft against it. We keep ours long, and we add to it monthly.
Name an editor for every piece. One person is accountable for each published item and must add something the tool could not: a real example, a sharper claim, a cut.
Read it aloud. Generic writing sounds wrong spoken. If a sentence would feel strange coming out of your founder’s mouth, rewrite it.
A day in a team that gets this right
Picture a marketing lead on a Tuesday. A client interview from last week has been transcribed overnight, with the ten strongest quotes pulled out and timestamped. She reads them and picks two that capture something nobody else in the market is saying. AI drafts a first version of a case study from the transcript and the brief. She rewrites the opening, adds a detail from her own notes, cuts a paragraph that sounds like every other agency, and sends it to the client for approval.
The tool saved her most of a day. Every decision that made the piece worth reading was hers. That is the division of labor to aim for.
Set the rules before the habits set themselves
Teams that skip this step end up with an informal policy decided by whoever uses the tools most. Write a one-page AI policy for marketing that covers four things. Which tasks AI may do. Which tasks it may not. Who reviews the output. What must be disclosed, such as a synthetic voice.
Review it every quarter. The tools change quickly and the line will move. What should not move is that people own the decisions and the final word.
What to do this month
List your team’s recurring marketing tasks and mark each as production or judgment.
Move two or three production tasks fully onto AI with a human check, and measure the hours saved.
Write your house style list and one-page policy, and share them with everyone who publishes.
Assign a named editor to every piece going out under the brand’s name.
A limit to keep in view: the BCG experiment ran in 2023 with consultants on set tasks, and models have improved since. The pattern still holds in our experience. AI is a powerful assistant and a poor substitute for a point of view.
We believe story beats noise, and much of the noise is now automated. If you want help setting up an AI workflow that keeps your team fast and your voice unmistakable, start a consultation with us.