Inside Our AI Marketing Lab

What follows is the most under-explained part of how Creative Marketing actually works. Most agencies do not publish their R&D process because the process is either unflattering, embarrassingly thin, or genuinely confidential. Ours is none of those things, so we publish it. The lab is the engine behind the products, the essays, and the strategy work — and seeing it is the most useful thing a journalist, a senior prospective hire, or a CMO evaluating a long retainer can do.

1. The room

The lab is a dedicated room at our Rochdale headquarters with the Monday Brief written on one wall in marker pen, and the seven active prototypes pinned up on the other. There is a long table, eight chairs, two screens, no glass partitions and no whiteboard that wipes clean — both deliberate. The Monday Brief in marker pen means we have to literally rewrite it every Monday morning, which forces the team to argue about whether it should change. A whiteboard that wipes clean encourages cleaning it; a wall in marker pen encourages argument.

About a third of the people involved in the lab work remotely. The rituals run on video for them. The room is where the in-person work happens; the lab itself is more accurately a process and a cadence than a place. Visitors usually expect the lab to be a gadget room. It is not. It is a small, deliberately uncluttered space designed to make conversation hard to dodge.

2. The three gates

Every AI-first idea — feature, framework, methodology, agency service, product hypothesis — has to clear three gates before it gets called a feature, ships to a client, or appears on the website. The gates are deliberately uncomfortable. Most ideas die at one of them, which is the design intent.

Gate 1 — Research-grade evidence. Before an idea gets prototyped, someone has to write a one-page research brief that establishes the problem is real, the approach is defensible, and the failure modes are named. The brief is reviewed in a Wednesday Tear-Down. About six in ten ideas die here. Most fail because the problem turns out to be already-solved, the proposed approach has a public counter-example, or the failure modes turn out to be larger than the upside.

Gate 2 — Six-week prototype. Ideas that survive the research gate get a strict six-week prototype window. Not "a project". A working thing built end-to-end against a real client problem with a real user. About three in ten of the ideas that reach this gate die here. Most fail because the prototype reveals a structural problem the research missed — usually a brittleness under real-world data noise that did not show up under test data.

Gate 3 — Three-client live test. Prototypes that survive get a three-client live test. Three different brands, three different sectors, eight to twelve weeks each, with the team involved being explicitly told they have permission to kill the idea. About half of what reaches this gate becomes a shipped feature. The other half is killed because the live performance, across three brands, is not differentiated enough to justify shipping.

One shipped feature per twenty ideas is the rough rate the gates produce. We think this is the right rate. A higher hit rate would mean the gates are too lenient and we would be shipping things that should have been killed. A lower hit rate would mean we are not generating enough ideas at the top of the funnel.

3. The four rituals

Monday Brief (45 minutes). Every Monday at 9.30 the lab convenes around the wall and rewrites the brief in marker pen. What are we building this week, what did we kill last week, what is the most important question we are trying to answer. Anyone in the company can attend; they cannot speak unless invited by the lab lead. The point is for the rest of the agency to see what the lab is doing, not to debate it in real time.

Wednesday Tear-Down (90 minutes). Two ideas in the research-gate phase get torn down. The author of the research brief presents for fifteen minutes, then the rest of the lab attacks the brief for thirty. The remaining time is decision — kill, revise and re-tear-down next week, or advance to the prototype gate. The tone is deliberately uncomfortable. The easiest thing in the world is to be polite about a colleague's idea; the most expensive thing is to ship it.

Friday Ship-Or-Kill (60 minutes). Prototypes that have been running for six weeks and live tests that have been running for the agreed window come to Friday for a binary call. Ship to clients, kill it, or extend with a named reason and a hard new deadline. The rule is no third-time extensions. If a prototype cannot be decided after two extensions it is killed by default, regardless of how attached the team is to it.

Quarterly Public Post-Mortem. Once per quarter we publish a long-form piece on something we got wrong. A killed feature with the autopsy. A thesis we held in a previous essay that has not aged well. A hire that did not work and our share of the reason. A client engagement that ended with us at fault. The point is to make it socially safe inside the company to surface failure, and to make the agency externally trustworthy by being on record about its own mistakes. The first 2026 Public Post-Mortem publishes in May.

4. The rules nobody breaks

The lab runs on five rules that are not written on the wall but are enforced ruthlessly. Naming them publicly is mostly a way to hold ourselves to them.

1. Nothing ships to clients without surviving all three gates. Including things the senior team is excited about. Especially those.
2. Marketing language is not allowed in the lab. If you cannot describe what a feature does in operator-language, without using "AI-powered", "intelligent", or "next-generation", the feature is not ready.
3. The killer of an idea names the next attempt. If you kill someone else's idea in a Wednesday Tear-Down you owe the room a sketch of what you would try instead. This stops critique from being free.
4. The author of the brief gets the last word, but not the decision. Final ship-or-kill calls are made by the lab lead, not the author. The author can disagree on record; the record is referenced six months later when the decision proves right or wrong.
5. Failed ideas get a one-page autopsy filed. Within two weeks of a kill. The autopsy is read at the next Monday Brief. The point is to compound the learning across the team rather than re-attempt the same idea in slightly different clothing six months later.

5. Why we publish this

Every agency website has a "how we work" page. Most of them are decorative. The reason we publish the actual mechanics — the gates, the rituals, the rules, the failure rate — is that the audience for an agency website in 2026 has read enough of those decorative pages to be immune to them. The CMO evaluating a multi-year retainer in their thirties is not looking for "we are strategic, we are creative, we are data-driven". They are looking for evidence that the engine producing the work is real.

Publishing the engine carries some competitive risk. A competitor could read this post and copy the rituals. We think they will not, because the rituals are uncomfortable to run and most agencies will not adopt uncomfortable rituals voluntarily. And if they do, the industry gets better, which we are fine with. The lab's edge is not the rituals. The lab's edge is the willingness to enforce them when it would be easier not to.