I have watched marketing teams begin their AI journey the same way. Content generation, usually with a brand voice project attached: define how the company sounds, then generate new content that sounds like it.
It is a reasonable experiment, and I would not talk a team out of it. But the same story keeps playing out afterwards. The output does not work the way people hoped. They use the tool a little, rewrite what it gives them, and drift back to plain ChatGPT or Claude, or to writing things themselves. Meanwhile the team keeps talking about its AI projects. Leadership hears that AI is in use. Nobody finds out how few people are using it.
And nothing surfaces it. Someone submits a draft. Was it AI? Written by hand? An AI draft somebody fixed? Nobody can tell, so nobody asks, and the adoption claim can never be tested.
Content generation captures none of your corrections
Content generation has no improvement loop built in. Someone prompts the tool and does not like what comes back, so they rewrite it themselves, or abandon it and write from scratch, or keep re-prompting until it gets closer. Whichever one happens, the editorial input is never captured. Next time, it starts all over again.
Compare that with a report. You generate it, you distribute it, and people respond. I don’t like this report. It is missing this insight. I want this data. The table should be presented differently. Every one of those comments can go back into the system to improve the next run. That cycle is what content generation misses, and it is the difference between a tool that gets better and a tool that gets quietly abandoned.
Teams manage AI like a purchase, not like a junior
Teams treat AI the way they treat every other piece of marketing software. Buy it, set it up, use it. A purchase has that lifecycle. So the AI gets bought, set up, used, disliked and dropped, and the conclusion is that AI does not work for this team.
Marketing teams do this with software constantly, and the reasons are usually real. What gets delivered is not what was demoed, nobody changes how they work to suit it, some requirement nobody checked turns out to be missing, or the manager who championed it leaves. Any of those and the tool quietly stops being opened. For software that is survivable. You lose the licence fee and go back to what you were doing.
A better model is that an AI system is a junior teammate. Very intelligent, very inexperienced. Someone has to manage it: review its work, feed corrections back, be answerable for whether it is improving. No team would hire a bright graduate, give them no manager, no feedback and no way to tell whether their work was being used, and then conclude after three months that graduates do not work. A content rollout with no loop does roughly that.
This is a claim about accountability, not a cute way of talking about software. A junior has a manager who answers for their development. Most AI rollouts have a project sponsor who answered for the launch.
There is a bigger version of this argument. What the loop really builds is judgment that outlasts the tool it was learned on, so dropping the tool forfeits that as well. It deserves more room than it gets here.
Let the content project happen, then pick work where failure is loud
The obvious advice is useless. Telling marketing teams not to start with content will not help, because they are going to start with content anyway. It is the visible, exciting first project, and it is a fine experiment.
The advice is aimed at management. Let the team do content quickly, then push beyond it. The failure is stopping there. Announcing that the content workflow is live and the team is now AI-powered does nothing for the team and nothing for the people in it.
The next project is the one that decides whether AI adoption is real, and the selection criterion is simple: pick work where failure is loud. The best candidates combine multiple teams.
- The AI produces output for a different team.
- The AI produces analysis that has to be presented.
- The AI produces reports that have to be distributed.
Two properties follow. Failure is apparent, because there is no quietly-fixing-it-yourself option when the output crosses a team boundary. And feedback is easy to give, and has somewhere to go.
Notice that the generic advice points the other way. Start small, pick a low-risk use case. Low-risk usually means low-visibility, and low visibility is what killed the content project. The risk worth managing is mediocre work in private that nobody notices for two quarters, not a public mistake in week one.
What the loud project needs in place
- One named person is accountable for the system’s performance and improvement, the way a manager is accountable for a junior’s.
- Corrections have a route into the system, not just into this week’s output.
- You can tell whether the output was used. If nobody can distinguish an AI draft from a human one, no adoption claim is falsifiable.
- The work already happens manually and consistently. Loud failure is a selection criterion, not a substitute for having a real process to automate.
Review output weekly for the first quarter, then monthly. Check quarterly whether usage is broadening or concentrating, because concentration is the early sign of the quiet death: a tool used enthusiastically by two people and ignored by six is the failure pattern in progress, not a partial success.
Signals the adoption is real
Early signals:
- Is anyone correcting it, and do the corrections reach the system? Silence is not success. It usually means nobody is reading the output.
- Can you say how many pieces of work went through the AI workflow this month? If not, you have an announcement, not adoption data.
Later signals:
- Six months in, usage is broadening across the team rather than concentrating in the two people who liked it on day one.
- The output has improved in a way someone can point to, which is the only evidence that anyone is managing it.
- If the system were switched off, somebody would notice within days.
That last one is the honest test of the whole thing. For most content-and-brand-voice rollouts, the answer to “who would notice?” is nobody, and that answer is the diagnosis.
Where to start
-
Ask the switch-off question tomorrow. If your AI tooling disappeared overnight, who would notice, and how quickly? It takes a minute and it tells you whether you have adoption or an announcement.
-
Pick a next project whose output goes to an audience on a schedule. The strongest shape I know: AI-drafted campaign analysis, published regularly to people outside the team. Within two cycles you will know whether it is any good, because the audience will tell you.
-
Name a manager, not an owner. One person accountable for that system’s performance and improvement, with the same expectations you would have of someone managing a bright new graduate.
-
Make usage visible without making it a compliance metric. You need to know whether the output is used and trusted. You do not need a dashboard of who complied.
None of this asks the team to skip the content experiment. Run it, learn from it, and then move to a project you cannot hide from. That second project, not the first, is where AI adoption starts.