Why agencies hide it — and why hiding is the tell
Ask ten agencies whether AI touches their client deliverables and you'll get ten versions of "we use it for research." That's usually not true. The economics are too obvious: content retainers are priced on human hours, and if the client knew a draft took eleven minutes to generate, the invoice would get a harder look. So the industry settled on a quiet arrangement — use the tools, bill the old way, hope nobody asks.
We think that arrangement is a liability, for two reasons. First, it's fragile. Clients are using the same tools now. They can smell an unedited AI draft, and the day they catch one in their deliverables folder, the relationship is done — not because AI was involved, but because the agency lied about it.
Second, hiding the tool usually means nobody built controls around it. If your agency won't tell you AI is in the pipeline, it probably also can't tell you what stops a hallucinated statistic from reaching your blog. The secrecy and the sloppiness tend to travel together.
So we say it first. AMC is two senior operators and an AI operating system we built ourselves. The system does the production. Humans do the judgment. And the whole thing runs on hard gates — nothing publishes without passing them, and no gate can be waived by the automation itself.
The pipeline, end to end
Here is the actual production line behind the content we ship for clients, on a Tuesday/Thursday publishing cadence. Nothing below is aspirational — this is the system as it runs today.
Keyword-tier brief
Every post starts from a keyword strategy the client approved before any writing began, organized into three tiers by intent and difficulty. The brief locks the target query, the search intent, the length (our posts run 1,100–2,400 words), the internal links, and the schema types the post must carry.
Generation
The system drafts against the brief under constraints: the client's voice profile, a banned-phrase list, required sources, and a rule that any number must trace to a document a human can open. A draft is raw material at this stage — it has never once been publishable as-is.
Human editor pass
A named editor reads every word. Claims get checked against the source, filler gets cut, anything the machine asserted confidently but can't support gets deleted or rewritten. This is the most expensive step in the pipeline and the one we will never remove.
E-E-A-T check
Does the post carry a real named author with a real bio? Does it contain something only this business could say — first-hand detail, real numbers, an actual opinion? If the answer is no, it goes back. A post that could sit on any competitor's site interchangeably is a post that shouldn't exist.
Schema injection
Every post ships with validated JSON-LD — Article plus FAQPage, Product, or TechArticle depending on the content. This isn't decoration: structured data is how search engines and AI answer engines parse who wrote what, about what, and why it's credible.
Publish
Only when every gate reads green. The scheduler holds the Tuesday/Thursday slot; a failed gate means the slot goes empty rather than filled with something that didn't pass. An empty slot costs us nothing. A junk post costs the client.
The guardrails, specifically
"Human review" is what every agency claims. Here's what ours actually consists of — four gates, each of which can kill a post on its own:
- Brand-voice gates. Each client has a documented voice profile — register, sentence length, what they'd never say. Drafts are checked against it before an editor ever sees them, so the human pass is spent on substance, not tone repair.
- Banned-phrase lint. A hard blocklist of AI-isms and agency filler — "delve," "in today's fast-paced digital landscape," "it's important to note" — plus each client's own no-go terms. A lint failure blocks publish automatically. It is not a warning; it is a wall.
- Quality scoring. Every draft gets scored against the brief: intent match, source coverage, internal linking, originality. Below threshold, it goes back to editing. There is no path where a low-scoring draft publishes because the calendar said so.
- Human sign-off. The final gate is a person clicking approve, with their name attached. The automation prepares, scores, and flags. It never approves. That asymmetry is the entire design.
The gates exist because we've seen what the machine produces without them. Left alone, a generation model will invent statistics, cite studies that don't exist, and write with the bland confidence of something that has never had to face a client. The pipeline isn't built on trust in the model. It's built on the assumption that the model will get things wrong, every day, and the system's job is to catch it before you do.
What Google's June 2026 spam update actually punished
Google's June 2026 spam update went after what its policies call scaled content abuse: publishing large volumes of pages whose purpose is to manipulate rankings rather than help anyone. Read the policy closely and you'll notice what it doesn't say — it doesn't say "AI content." The policy is medium-agnostic. A thousand junk pages written by underpaid freelancers die in the same update as a thousand junk pages written by a script.
What the update reliably caught was content with the fingerprints of unreviewed volume: no named author, no first-hand anything, no number that traces anywhere, pages interchangeable with every other result for the query. That profile is easy to produce at scale and increasingly easy for Google to detect at scale — which is exactly why unreviewed AI junk has a short shelf life. It's not that AI wrote it. It's that nobody read it.
Our position on this predates the update, and we've acted on it even when it hurt a dashboard. For one client — a Canadian consumer lender we won't name because the niche invites judgment — the honest move was deleting content, not adding it. We noindexed 30+ off-strategy pages and watched clicks drop 23% on purpose, while impressions for the queries that actually matter rose 34% (55,882 to 74,921) in the first month (Google Search Console, Month 1 of engagement, 2026). A content system has to be able to make a number go down when down is correct. A content mill can't.
And when the system points the other direction — real expertise, real volume, properly reviewed — it holds up:
That byline detail is worth sitting with. The content for a 39-year-old industrial furnace manufacturer in Spain didn't move because we generated it faster. It moved because the manufacturer finally had a real engineer's name and real technical depth on pages that used to carry a fake author. The AI made the volume possible. The E-E-A-T made it work.
The honest limits: what AI is still bad at
If this post were pure sales copy, it would end at the receipts. It's not, so here's the other column — the things we've learned the machine cannot do, no matter how good the prompt:
- First-hand experience. The model wasn't on the client call, didn't stand in the shop, has never smelled a furnace. Every piece of genuine experience in our content is put there by a human, because there is nowhere else it can come from.
- Numbers. Left unsupervised, a model will produce a statistic for anything, with total confidence and no source. Our rule — every number traces to a document — exists because the alternative is publishing fiction with a percent sign on it.
- Knowing what not to publish. The lender story above required deciding that 30+ pages should die. No generation system suggests deleting content; the incentive gradient of the tool always points toward more. Strategy is subtraction, and subtraction is human.
- Taking a position. Models are trained to be agreeable, and agreeable content ranks like wallpaper. The opinions in our clients' content — and in this post — are ours, because "on the other hand" is not a point of view.
That's the honest shape of it: AI compresses production from days to minutes, and it moves the human work up the stack — from typing to judging. Any agency telling you AI replaced the judgment is either wrong or about to be.
Questions we get about this
Do you tell clients when AI was involved in their content?
Yes, before they sign. It's in this post, it's in our FAQ, and it comes up on the first call. The pipeline is part of what clients are buying — a small team that ships on a Tuesday/Thursday cadence because the production layer is automated and the judgment layer is human.
Will Google penalize AI-assisted content?
Google's spam policies target scaled content abuse — mass-produced pages made to manipulate rankings — regardless of whether a human or a machine produced them. What gets punished is unreviewed volume with no author, no expertise, and no reason to exist. Reviewed, attributed, genuinely useful content is fine under the policy however it was drafted.
Who actually reviews the drafts?
A named human editor — Nico on delivery, Andres on strategy pieces — reads every draft before it can publish. The automation can prepare a post, score it, and flag it. It cannot approve it. Human sign-off is a hard gate, not a policy suggestion.
Andres Marin
Founder & Strategy, AM Consulting Marketing. Builds the systems described in this post and takes the hard client conversations personally — including the ones where the answer is "you shouldn't pay us for that."