There is no shortage of writing about what AI agents will do to marketing work. There is very little about what has already been handed over, by name, by people who had to live with the result.
So this article has a deliberately narrow entry requirement. Every person quoted below describes something they built and ran — a pipeline, a product, an org chart of agents, a workflow they measured before and after. No predictions. The one exception is the last section, and it is there because somebody has to say the sceptical thing and mean it.
Read in order, they describe a progression: hand over the assembly, then hand over the diagnosis, then hand over the routine admin — and then stop, hard, at a line that turns out to be the same line for everyone who has one.
The whole content pipeline, in a terminal
Zak Ali works on SEO at Finder US and runs weekly internal training on this. His build is the most complete end-to-end handover anyone has described to me on the show:
Basically what I did is I created a publisher, essentially, which does the keyword research using the Ahrefs MCP, builds a customer profile, does the brief, the content outline, does the competitor research, does the deep research to be well-cited, can do the internal linking — basically the end-to-end process — and then eventually push into WordPress itself. So you don’t even have to leave your terminal to do your entire job.
Zak Ali, Finder US
The important part is what he says immediately after, because it is where most people’s thinking stops and his does not:
And once that agent’s built, well now you have 20 sessions going at the same time. What’s to stop you from building a hundred? You don’t want to get into the scaled-content-abuse realm, but that part of the job is mostly commoditized at this point.
Zak Ali, Finder US
Two claims are stacked there and both are load-bearing. The first is that the constraint on content production is no longer capacity. The second is that he chose not to take the capacity, and the reason is a quality boundary rather than a technical one.
What he redirected the freed time into is the actual thesis: “now that means you can spend more time creating awesome YouTube videos and guest blogging, creating more lifestyle pieces where there’s no clear ROI, but the community building is there and people are building trust with you.” The automation did not replace the work. It moved the work to the part a model cannot do.
He also runs a correction mechanism most people skip, and it is the single most transferable idea in this article:
Have a judgment ledger as part of your second brain, where you see it makes judgments, and having to close the feedback loop when those judgments don’t pan out the way they thought they would has been such a massive growth loop for making the AI perform better and better as time goes on.
Zak Ali, Finder US
A judgment ledger is a log of the calls the system made, revisited when they turn out wrong. It is the difference between an agent that produces output and an agent that gets better. Most stacks have the first and not the second.
Closing the loop that dashboards leave open
Somya Goyal built Stellarcast around a complaint about the entire analytics category — that it stops one step short of doing anything:
So it is basically monitor, diagnose, execute, and then prove, so it continuously moves in, it completes the flywheel. Most of the tools stop at the dashboard. The whole point of Stellarcast is closing the loop and keeping the things in log and moving the things continuously rather than just one-time audit or one-time fix.
Somya Goyal, Stellarcast
That four-verb sequence is a good test to run against anything sold to you as an SEO agent this year. Most tools do monitor and diagnose. Very few do execute. Almost none do prove, which is the one that determines whether the previous three were worth paying for.
Note where she puts the human, though. Not outside the loop — inside it, at the gate:
It will give you drafts of the fixes and once you approve with the details given it can also ship it, fix it, and will definitely move the needle for you.
Somya Goyal, Stellarcast
Drafts the fix, ships on approval. That is the same architecture Zak described from the other direction — full automation of the assembly, a human on the judgment. Two people who built very different things landed on the same division of labour.
The unglamorous win: four hours down to thirty minutes
James Ernst of Core Order works with small businesses, where the wins are less exciting and considerably easier to prove:
Where I’ve seen the biggest win is more on leveraging and understanding the data that’s there. So whether that is finding a task that takes a company a good four hours a week per ten employees and finding a way to bring that down to thirty minutes, by leveraging that time and allowing them to focus on what’s actually gonna move the needle for the company.
James Ernst, Core Order
That is the sentence to steal if you are trying to get an AI project funded. It is scoped to a single named task, sized per ten employees so it scales legibly, and it has a before and an after. Nobody has to believe anything about the future of work to approve it.
It is also, structurally, the same move as Zak’s: the point is not the hours saved, it is what the hours get redirected into.
The line nobody crosses
Erica D’Arcangelo of Love Content Development is not an AI sceptic. She is explicit that the tooling is remarkable — she called an internal-linking tool I was describing “phenomenal” and said outright, “I absolutely love technology. It saves so much time.” Her position is about a boundary, not a verdict.
She writes literary fiction, young adult, and children’s books alongside the client work. And that is where the handover stops:
That is something that I, as an artist, would feel very compromised if I used AI for something that I was putting out as my own work.
Erica D’Arcangelo, Love Content Development
The word is compromised, and it is doing something the usual objections do not. This is not a claim that the output would be worse. It is a claim about authorship — that the thing published under your name has to be yours, and that a good enough result obtained the other way is still a loss.
Her general principle is the same one Somya and Zak arrived at by different routes: “you may be able to get data and research and get help with different, how the article sounds, but you have to be responsible for what you’re saying in there.” Responsibility does not delegate. She makes the same argument at length on her own site in Why AI Content Can Never Replicate the Magic of Human Creativity.
The counterweight: none of this is settled
Chris Panteli of Linkifi supplies the sobering note, and it is not about agents. It is about whether the underlying models are ready to be trusted with anything at all:
They have a disclaimer. It says these models can hallucinate. And Google was always about delivering safe results on its SERPs, wasn’t it? It never had a disclaimer saying like these links may be unsafe on our first page.
Chris Panteli, Linkifi
That comparison is hard to shake once you have heard it. We adopted a class of tool that ships with a written admission of unreliability, at a speed we would never have accepted from a search engine.
His second point is about where the curve goes, and it is the reason to build correction mechanisms now rather than later:
At some point it’s going to flatten. And then I read somewhere about how we’re having this dilution of information because if everything is being produced by AI, then what is the next generation of training data going to be based on? It’s like its own waffle. That’s pretty scary. Like we do need new content.
Chris Panteli, Linkifi
He is candid that he does not know: “it’s so new and experimental.” That is a more honest position than most of the confident ones, and it points at the same conclusion as everyone above — the scarce input is original human material, which is precisely what Zak redirected his freed hours into and precisely what Erica refuses to hand over.
What five people who shipped things actually agree on
- Assembly is commoditised; judgment is not. Every automated pipeline here keeps a human at the approval gate, and none of them describe that as a temporary limitation.
- The metric is redirected hours, not saved hours. Zak into community and video, James into the work that moves the company. Saved hours that go nowhere are not a win, they are a headcount argument.
- Prove is the verb everyone skips. Somya’s monitor-diagnose-execute-prove is a purchasing checklist. Run it against your stack.
- Build the correction loop before you scale. Zak’s judgment ledger is what makes the difference between output and improvement, and Chris’s training-data point is why it will matter more, not less.
- Know where your line is, in advance. Erica’s is authorship. Zak’s is scaled content abuse. Having one is not sentimentality; it is the only thing that stops capacity from making the decision for you.
If you want the more prosaic version of the same argument — what to hand over first, where the cheap wins are — that is in 7 Quick Wins Hiding in Your Last SEO Crawl. Most of the tasks worth automating are hiding in an export you already have.
Every quote in this article is verbatim from Unscripted podcast transcripts and attributed to the guest who said it. Light cleanup has been applied to filler words only.