How are B2B marketers handling AEO right now? Anita Moorthy at Rocksalt is running a 2-minute pulse check on who owns it, what's working and whether it's paying off. Answers are confidential, and everyone who takes part gets the results »
In January, I wrote an internal strategy memo outlining a second brand — an AI-run agency. A marketing factory.
One heading reads “The Operating Model: No Human Judgment.” And underneath it I wrote “The business is designed to run without an owner.” and “We treat B2B Demand Gen as a commodity code-base that can be deployed, not a creative project that needs to be invented.”
Fixed price, zero touch, no sales calls, no kickoffs, no strategy meetings. We don’t consult, we just execute SOPs.
I wanted to make a production line.
And I believed every word when I wrote it.
But seven months later, I’m now running the agency in the opposite direction. Not because the models underperformed. They did everything I outlined in the memo.
The problem was what happened when they did…
What everyone gets wrong about AI
The argument that AI writing is bad is weakening by the day. Models are getting better every few weeks, and by 2028 the argument will be dead in the water.
But quality was never the problem. We went hard on AI copy across our clients. And it worked for about six months. Output went up, costs went down, and the quality was…fine.
Not good. Not bad. Just fine.
It wasn’t until I sat down to read a quarter’s worth of output side-by-side that I saw what the real problem was — every client sounded identical.
Same rhythm, same structure, same patterns. Same tidy closing line that lands on nothing.
A financial services client and a healthcare client where you could swap the copy and it wouldn’t matter. Strip the logos and branding away and I wouldn’t be able to tell you which was which. When viewed in a vacuum, all of it would pass review. But when they’re placed side-by-side, it was the same one voice wearing eight different logos.
That’s not a quality failure, it’s a convergence failure. And it’s a different problem with a different cause.
A better model makes it worse, not better
The logical assumption is that this problem will inevitably get solved. If it’s not by a better model from Anthropic or OpenAI, then by a vertically trained marketing model. It’s only a matter of time until we have LLMs trained on ‘good’ B2B copy instead of blindly trusting what gtmking69 shared on Reddit.
But that won’t fix anything. If anything, it only accelerates the problem.
Genuinely good marketing doesn’t come from a paint-by-numbers playbook. It comes from thinking critically about a specific situation. A situation that has never existed in that exact form before. But a model trained on marketing can only paint by numbers. It’s trained on playbooks, templates, formulas, etc. It’s very good at producing the median of everything that has already worked.
Which is fine, if you’re the only one using that model.
But if everyone is running their briefs through the same model, we all get the same output. And that doesn’t change with better models. Improving the model doesn’t give you differentiation, it just gets you a higher quality average that everyone converges on at the same time.
The ceiling isn’t the model’s capability. It’s that we’re all drawing from the same well.
“Humans converge too” isn’t the counter you think it is
I’ve heard the argument that human marketers also converge. That B2B copy all sounded the same long before ChatGPT was released four years ago. Every SaaS homepage had the same hero, the same three-column benefits, the same gradient, etc. We didn’t need LLMs to make everything look identical.
It’s true. But it misses the point.
Who they’re describing is mediocre marketers. These are the people who see what works for someone else, then copy it. These people make up most of the industry, and always have. But the best marketers create and innovate. They make new things, not endlessly recycle what’s come before. These people drive the industry forward, carrying everyone else on their shoulders.
This is the thing no model can replicate. An LLM can only do an impression of originality by smashing together a bunch of existing ideas, then presenting them back to you confidently.
So yes, humans also converge. But the good ones are the exception, and the exception is what stops the whole industry getting stuck in a loop.
The second thing a better model doesn't fix
The memo I wrote wasn’t theoretical. We actually built the tier I outlined. Same strategy, same thinking, AI does the execution, and a human occasionally reviews it. A lower price, with lower touch, with a bigger market. But clients killed the idea with a single question:
“What happens when something goes wrong?”
And something always goes wrong. Media spend, for example, is a long sequence of judgement calls made with incomplete information. Some calls work out, some will be wrong. That’s not a risk you engineer out, it’s the nature of the work.
What I hadn’t considered is what people are actually buying when they buy this service. If I’m hiring someone and they outsource the work to an AI agent, I didn’t hire them. I hired a wrapper around a model. And you can’t hold a model responsible, you can’t fire it, and you definitely can’t expect a useful answer if you ask why it made a decision.
“What happens when something goes wrong?” isn’t really a question about process. It’s actually about ownership. And there’s no version of the model where the answer to that is the model.
Welcome to the Analogy Hotel
I recently asked Claude to book a hotel. I gave it a specific property and a specific price, and let it go to work.
That was a mistake.
It booked a completely different hotel for five hundred dollars more. Non-refundable. When I asked why it did this, it apologised and told me the booking couldn’t be reversed. It was fast, confident, and completely wrong. And I only caught it because I checked it manually.
What happens when that same agent is running a live campaign, and the two-week report is the first time anyone looks?
What I actually run now
This isn’t an anti-AI essay. We still use LLMs at 42, and that won’t change. But the line I’ve landed on is that AI does the work, but a person needs to decide what the work is for.
Every time I’ve tried to invert that, it breaks. But when I stick to this rule, it works great.
We use AI for internal knowledge retrieval, making every transcript, campaign, and past decision searchable. We save hours by generating production variants and custom graphics, with a designer part of the process deciding what good looks like, and whether it’s checked the box.
We also built a tool that detects anomalies on live campaigns. It was originally planned to be a product, but now it lives in Slack and flags things to the paid media team. Which means no-one has to log into the ad platform on a Sunday.
In all three cases, the pattern is the same. The judgement sits upstream with the domain expert, and the labour is downstream, handled by the AI.
That’s why, while other agencies are dropping copywriters in favour of AI-generated copy, we’ve recently hired a copywriter. In fact, we’re hiring more people than we were a year ago, not fewer.
The memo was right, for the wrong reasons
Better models are coming, and they’ll fix a lot. Cleaner copy, fewer hotel-shaped mistakes, better medians. But none of that will touch the two problems. The convergence will still happen because everyone will be running it against the same brief. And there’s still no-one to blame when things go wrong (and they will).
The memo designed both differentiation and accountability out of the business. On purpose. No human judgment, no creative project, nothing that needs to be invented. Everything that made the service a service, treated as inefficiency.
The proudest line in my memo was that the business was designed to run without an owner.
It got that right. That was the problem.