The approval gate AI marketing needs

For about a decade the constraint on content marketing was writing it. That constraint is gone. A founder with a decent prompt can produce more email copy in an afternoon than their agency shipped in a quarter.

Which means the bottleneck moved. It's no longer can we make this? It's should this go out? — and almost nobody has built a step for answering that, because for ten years the answer was implied by how expensive the asset was to make.

This is the step Tilly is designed around: everything it writes is shown to you before it goes anywhere. Not a log of what it sent. A queue of what it's about to.

Table of Contents

• Why volume changes the risk
• The six-point gate
• What the gate caught on our own site
• How long review actually takes
• What not to gate

Why volume changes the risk

When a post took a week, it got read carefully — by the writer, an editor, and usually the founder. Three sets of eyes, because the asset was expensive enough to deserve them.

At twenty assets a week, that process doesn't scale and quietly gets dropped rather than redesigned. What replaces it is a spot check: the founder reads the first two, they're fine, and the rest ship unread.

The failure mode isn't that AI writes badly. Modern drafts are competent. The failure mode is that AI writes confidently about things it doesn't know — your pricing, your shipping policy, what your product actually does — and competent, confident, wrong is much harder to catch at a glance than obviously bad.

The six-point gate

This is the list we run before anything publishes. It takes a couple of minutes per asset and it is deliberately short, because a checklist nobody completes protects nothing.

1. Is every factual claim about us actually true? Prices, timelines, integrations, guarantees, headcount, what the product does. This is the one that catches the most, and it's the only one where being wrong is genuinely costly.

2. Is every number sourced? Any statistic in an AI draft is guilty until proven innocent. Either it links to something real or it comes out. "Studies show" is not a source.

3. Does every link go where it says? Click them. All of them. Internal links to pages that were renamed, external links to pages that moved, booking links to the wrong calendar.

4. Does it contradict something else we've published? Drafts don't know what your other pages say. Two posts confidently stating different refund windows is worse than either one being wrong.

5. Is the media current and ours? Screenshots go stale faster than copy. A product shot of a UI you shipped two redesigns ago undermines everything around it.

6. Would I say this out loud to a customer? The catch-all. If a sentence would make you wince on a call, it doesn't ship.

What the gate caught on our own site

We ran this against withbranva.com in September 2026, on content that was already live. Six things came back:

• Two paragraphs of Spanish inside an English post. A translation step had run at some point and nobody checked the output. It had been live long enough that it was indexed.
• The same stale screenshot in seven posts — an image of our own site as it looked before a redesign, still being presented as current.
• A claim that our AI routed support tickets. It didn't. Nobody wrote that line maliciously; it was a plausible-sounding capability that got drafted, read as reasonable, and published. This is exactly the failure described above: confident, competent, false.
• A booking link pointing at the wrong calendar — repeated across every page and post on the site.
• Twelve buttons that did nothing when clicked.
• An article that was one sentence long, in the sitemap, offered to search engines as a real page.

Four of those six are mechanical and could be caught by tooling. Two of them — the Spanish and the false claim — needed a human who knew what was true.

How long review actually takes

The objection to a gate is always speed: if a founder has to approve everything, the founder becomes the bottleneck and the whole point of automation is lost.

In practice review runs at roughly one to three minutes per asset once the gate is a habit, because you're not editing, you're deciding. Edit-versus-approve is the distinction that makes it survivable. If you find yourself rewriting drafts rather than accepting or rejecting them, the problem is upstream — the brief was wrong, or the tool doesn't know enough about your brand — and rewriting each one is treating the symptom.

A week of output at twenty assets is under an hour of review. That is a real cost and it is much smaller than the cost of one false claim about your own product sitting on your site for a quarter.

What not to gate

Gating everything is how gates get abandoned. Things that don't need a human:

• Sends of an asset that was already approved.
• Scheduling and timing changes.
• Audience or segment selection, as long as the segment definition itself was reviewed once.
• Reporting.

The gate is for new claims and new creative going out under your name. Everything else is mechanics.

FAQ

Doesn't this defeat the purpose of automating marketing?

It defeats the purpose of unattended marketing, which was never a sensible goal for a brand with a reputation. Automation is worth having for the drafting, the scheduling, the publishing and the measurement — four of the five steps. Keeping the fifth is what makes the other four safe to run at volume.

What does this look like in practice with a tool?

In Tilly it's a queue: the work is drafted, you see it before it goes anywhere, and on approval it publishes into Klaviyo, Omnisend, Shopify Email or Meta — the tools you already use, rather than somewhere you'd have to check separately. The free tier shows you the queue without sending anything, which is the cheapest way to find out whether you'd actually approve what it writes.

We're a two-person team. Who reviews?

Whoever knows what's true about the business — usually the founder. That knowledge is the scarce input in the whole loop, and the gate exists precisely to spend it in the one place it can't be substituted.

This is step three of the loop we run: audit, build, publish, measure, again.