McKinsey’s vision of the future has one marketer supervising a network of AI agents, each one handling a slice of the work while the human sets direction. Sounds Lovely on a slide. In a smal team environment, especially when you are moving at pace, the real risk isn’t that the vision is wrong. It’s that you automate a workflow before it’s actually good, and now you’re producing bad output at ten times the volume, with less time than ever to notice.
The slide vs the lean team
Automation doesn’t know the difference between a good process and a bad one. It just does the process faster, more consistently, and at greater scale. If the process was sound, it becomes a big win. If it wasn’t, you’ve just built a machine for manufacturing the same mistake, again and again, and shipping it faster, before anyone checks.
A near miss
A three-person marketing team I know well nearly did exactly this. They were drowning in monthly reporting, pulling numbers from five different platforms by hand, so they built an AI workflow to assemble the report automatically: pull the data, summarise the trends, flag the wins and misses.
The dashboard they built worked. The report looked more legit than anything they’d produced manually, clean charts, confident recommendations and all out the door in twenty minutes instead of two days.
The problem surfaced a month later. Sales and marketing had never actually agreed on what counted as a “qualified lead.” Marketing’s definition was looser than sales’. Manually, someone used to catch the discrepancy while assembling the report by hand, because they’d eyeball a number and think “that doesn’t look right.” The automated version had no reason to eyeball anything. It just reported the number, confidently, in a nicer chart than before.
They caught it before it reached the board in the past, because someone happened to cross-check a figure against CRM data out of habit.
Fix it on paper first
The rule that came out of it, and the one I’d give any lean team before they automate anything: fix the process on paper first, then automate it. Write down exactly what should happen, get agreement from everyone who touches the process, and only then hand it to an agent to run at scale.
That’s slower up front. It’s also the only way to be sure you’re scaling something worth scaling, rather than scaling the exact inconsistency nobody had gotten around to fixing.
One catch here, is that when you write it on paper, you may do it a certain way due to constraints or limitations of skills, technology or data. Once you are ready to automate, don’t just automate the manual way, think about if it can be done differently, better or more accurately with the new tools at hand.
Where this actually applies
This isn’t an argument against automation, and it isn’t an argument against AI agents doing real work in a lean team. What’s i’m warning against is going in blindly. Automate a good process and you get the McKinsey result: more capacity, same quality, faster. Automate a broken one and you get the same mess, produced faster, wrapped in a more convincing chart, with less time built into anyone’s week to catch it before it’s sent.
Before you hand a workflow to an agent, ask the boring question first: does everyone actually agree on what this process should produce? If the answer’s no, work that out first.
The questions worth asking before you automate anything
In practice, I run any workflow through three questions before I’ll let a lean team automate it.
Would two people doing this by hand get the same answer? If the process still involves judgement calls that vary person to person (what counts as “qualified,” what counts as “on-brand,” what counts as “done”) automating it just picks whichever person’s judgement happened to train the prompt, and locks it in as the default.
Who currently catches the mistakes, and why? Somebody usually is, even if it’s informal. Understand what they’re actually checking for before you remove them from the loop, because that check needs to happen somewhere else afterwards, not disappear.
What does the output look like when it’s wrong? If a wrong answer looks obviously wrong, an automated version is low risk, because someone will catch it fast. If a wrong answer looks exactly like a right one (a monthly report, a lead score, a forecast) that’s the workflow that needs the process fixed first, because nobody’s going to catch it by eye once it’s automated.
The honest version of the AI-agents pitch
None of this is an argument against where the industry’s heading. A lean marketer directing a handful of agents through real work is a genuinely good use of a small team’s time, and I’d rather be building towards that than ignoring it. The honest version of the pitch just includes the bit that gets left out of the keynote: the agents are only as good as the process you hand them, and fixing that process is still a human job, done before the automation starts, not after it’s already shipped a hundred versions of the same mistake.
If you’re short on time, start small. Pick one workflow, run it through the three questions above, fix the one thing that’s actually inconsistent, and only then automate it. You’ll learn more from getting that single workflow right than from automating five at once and hoping the process was fine all along.
If you still aren’t convinced, automate it and run it manually for a few cycles till you are comfortable with the output.
Dom O'Brien
CMO at MATE and author of The Startup Marketing Playbook. Fifteen years building lean marketing teams that punch above their weight.
Work with Dom →