"Do they actually sound like you?"
That's the first question I get when I tell people I have 18 AI agents running my marketing every night. It's a good question. And the honest answer is: at first, no. Not even close.
The Problem Nobody Warns You About
AI brand drift is real, and it's subtle enough to be dangerous.
The scale of it across the industry is larger than most people realize. Salesforce's 2024 State of Marketing report found that 75% of marketers are already using AI in their work, but only 36% say they have formal guidelines for how AI should represent their brand. That means two out of three companies deploying AI marketing content are doing so without guardrails. They're not managing brand consistency. They're hoping for it.
Without explicit brand input, an AI model doesn't construct your voice. It constructs an average of everything it's ever read. And that average is always the same thing: professional, cautious, slightly boring. Generic consultancy language. Technically coherent. Completely wrong tone.
In my case, it was worse than just tone.
In the early weeks, agents were generating outreach messages and posts with lines like "82% of companies report X", plausible-sounding statistics, completely fabricated. They were using titles like "freelance Head of Marketing" instead of "Fractional CMO." That's not a minor wording preference. That positioning difference is worth tens of thousands in revenue over the course of a year. "Freelance" signals tactical, hourly, disposable. "Fractional CMO" signals strategic, senior, embedded.
The agents weren't trying to undermine me. They were just averaging the internet.
I needed to fix that.
Step 1: A Brand Constitution (Not a Brand Book)
The first thing I built was a reference document I call `brand-guidelines-quick.md`. Two pages. Every agent loads it before generating output.
No 40-page brand bible. No mood boards. Two pages that answer the questions that actually matter when you're generating content at 1 AM:
- What is the official title? ("Fractional CMO" : never "freelance", never "fractional Head of Marketing", never any variant)
- What are the approved stats, with exact notation? (18 years of experience. 200+ companies. €200M+ in media budget allocated over 18 years. CAC reduced by up to 40%. Pipeline growth up to 65%.)
- What's the tone per market? (NL: informal, direct, je/jij. EN: casual-professional, first person. DE: formal, Sie-form, thorough.)
- What are the forbidden phrases? (An explicit list, updated every time a new violation is caught.)
- What are the approved opening formats per language?
The document is not aspirational. It's operational. Every line in it exists because something went wrong once and I didn't want it to happen again.
Step 2: A Branding Manager Agent
Having a reference document helps. Having an agent whose only job is to check against that document is what actually makes it stick.
The Branding Manager runs at 4:00 AM, after every other agent has finished producing output for the night. It reads everything, outreach drafts, blog posts, LinkedIn content, SEO recommendations, and checks each piece against the brand constitution.
The checks are specific:
- Is the title correct?
- Do the stats match the approved list, with the right notation?
- Is the tone right for this market? (A message written for a German B2B audience should read very differently from one written for a Dutch startup founder.)
- Are there any invented statistics?
In the first week, the Branding Manager blocked 5 violations in a lead-gen strategy document, caught 1 in a cold email draft, and flagged an error in the brand guidelines themselves, a stat I had written incorrectly. The AI found my own mistake.
That one stung a little. It also made me trust the system more.
Step 3: The Output Review Loop
Every agent that produces outward-facing content marks it as "Pending Brand Review" before passing it to the next agent. The Branding Manager reviews it. Only after that does it reach me.
I'm the last link in the chain, not the first. That's deliberate.
What I delegate to the agents: generating drafts, catching errors, checking consistency, flagging violations, proposing corrections.
What I don't delegate: publishing, sending, approving anything that goes outside the system. The final call is always mine.
This matters because the risk with AI automation isn't that the system does too little, it's that it does too much, too fast, before anyone has checked it. The review loop slows that down. It means nothing goes out without having been checked at least twice: once by the Branding Manager, once by me.
Step 4: Learning Feedback Built In
Every violation the system catches gets documented in `agent-learnings.md`. Every new forbidden phrase gets added to the brand constitution. Every stat correction updates the approved list.
The system doesn't just run, it gets better.
By week two, the error rate had dropped from roughly 13% of outputs containing some kind of brand deviation to under 5%. Most of those remaining errors were subtle things: a slightly off-tone sentence in a German message, a stat with the right number but wrong formatting. Catchable, correctable, logged.
The brand constitution is now a live document, not a static one. It grows every time something goes wrong.
What AI Still Can't Do
I want to be honest about the limits, because this is where most AI automation writing gets slippery.
Contextual intelligence. The agents don't know I had a difficult client week. They don't know a competitor just made a public move that changes the framing of our outreach. They write from what I've taught them, not from what I'm thinking today. When I haven't logged a recent case study or updated the context documents, that gap shows up in the output. A German prospect email that should have referenced a specific logistics sector win instead referenced a generic "enterprise growth" angle that landed flat. Same technical quality. Wrong strategic context.
Personal anecdotes. I have 18 years of client stories, the Amazon Netherlands expansion, the AirHelp crisis campaign, the Gaastra relaunch. Those are not in any document. The agents can't reference them because they don't know them. When they try to fill that gap without guidance, the output sounds like a case study from someone else. Competent, but not mine.
Nuanced message timing. Knowing when it's too early to mention a specific case study, or when a particular angle is too aggressive for a German audience this week, that kind of judgment doesn't come from a brand constitution. It comes from experience. There's no document that captures "a major German competitor just had a public data breach, so our data security angle needs to lead earlier than usual in the conversation."
This is why I use interview formats for some content. I talk out loud about a case or a strategic decision, the agent writes it down, and the result sounds like me, because it literally came from me. The agent is the scribe. I'm the source.
The AI will average the internet unless you give it something better to average.
Before You Deploy AI Marketing Content
Most companies skipping these steps aren't doing it out of laziness. They're doing it because nobody told them the failure modes were this specific.
Any company deploying AI marketing content should do three things before running it at volume. First, build a brand constitution, not a brand book, but a two-page operational reference with exact titles, approved stats, forbidden phrases, and tone notes per market. Second, assign a review function, whether that's an agent, an editor, or a checklist run before anything goes external. Third, build a feedback loop, every deviation caught should be documented and fed back into the brand constitution so the same mistake doesn't happen twice.
The AI will average the internet unless you give it something better to average. The mechanical half of that check is public: you can run your own copy through the same gate. That's not a flaw in the technology. It's just how it works.
Where This Leaves Me
Brand governance in an autonomous AI system is not a one-time setup. It's a recurring loop: run, catch, document, tighten, repeat.
The system I have now sounds like me. Not because the AI is clever enough to infer my voice from the internet, it's not, but because I've spent months explicitly teaching it what I say, what I don't say, what I'm called, what numbers I stand behind, and what tone I use with which audience.
That's the real work behind "AI that sounds like you." Not a prompt. Not a persona setting. A system with guardrails, a reviewer agent, a feedback loop, and someone, me, who stays in the chain.
The agents handle the volume. I handle the voice.
Want to know what this looks like for your company? Book a 45-minute Marketing Scan. I'll give you an honest read on where your current setup is drifting, and what to fix first.
Sources & References
- Salesforce, State of Marketing Report, 8th Edition (2024), AI adoption rates and brand governance gap in marketing organizations. https://www.salesforce.com/resources/research-reports/state-of-marketing/
- Gartner, Hype Cycle for Digital Marketing (2024), AI-generated content quality and governance as an emerging risk category for brand integrity. https://www.gartner.com/en/marketing/research
- Own system data (2026): AI agent error rate dropped from 13% (week 1) to under 5% (week 2) after brand constitution and Branding Manager agent deployed across the autonomous marketing pipeline (13 agents at the time; the team has since grown to 18).