A pull-request culture for knowledge. The discipline AI-native marketing teams are converging on.

A marketing strategist at a Series B SaaS company opens her quarterly planning doc on a Monday in February. Three months earlier, her team had wired up an AI workflow to generate competitive briefs every two weeks — one-pagers titled How Competitor X Wins. The briefs landed in a shared Drive folder. Nobody had reviewed any of them in detail. They looked authoritative. They became the input to the strategist's roadmap. By Tuesday afternoon, a quarter's worth of dev spend has been allocated against a feature Competitor X publicly deprecated last August. Two of the three claims about Competitor X in the brief had been drawn from an archived 2023 marketing page that the AI had treated as current.

This piece is about how AI-native marketing teams stop that failure from being normal. The short answer comes from a phrase three different speakers landed on, independently, at Speero Circus 2026 last week in London: a pull-request culture for knowledge. Treat AI-generated artifacts the way engineers treat code. Never merged without a labeled reviewer. Never enters the team's knowledge base without a pass/fail signal. Never operates as truth until someone has signed off.

Three independent voices, one shape

Marcel Toben, Head of Engineering and Experimentation at Zalando, named the shape sharpest in his published abstract for the conference talk. "AI coding agents let teams ship 10–50× faster — but confidence hasn't scaled with speed. As changes pile up, review breaks down, metrics get noisy, and decisions become harder to trust." He calls this the velocity–confidence gap. His argument is that experimentation has to become the control plane for change — the discipline that turns AI-driven speed into outcomes a team can actually act on. Without that, faster shipping is faster drift.

Leah Tharin, a B2B SaaS growth executive at Fyxer, put it more directly in her own conference abstract. "The real AI risk in experimentation isn't hallucination, it's unreviewed context becoming organizational truth." She made the case for what she calls a pull-request culture for knowledge: let AI build the artifact, but treat approval like engineers treat code merges. Especially when the people consuming the output — the leaders making strategic bets on it — aren't technical and can't independently audit the inputs.

The third independent voice was Caitlin Sullivan, who ran one of the Day 1 workshops. Her published workshop description was the most concrete. Build multi-agent systems where each agent's job comes with built-in verification: it checks its own output against source data, it cross-references findings across multiple data sources, it flags contradictions. Citations and evidence requirements are baked into every output. The artifacts the system ships aren't slides or dashboards — they are markdown files that produce consistent, repeatable results from each run.

Three different speakers, three different vocabularies, one shape. Toben names the gap. Tharin names the failure mode. Sullivan ships the architecture. They wrote these framings independently — three separate submissions to one conference, written before any of them took the stage. Put them together and you get the same picture: AI generates knowledge faster than humans can absorb it, so the discipline that has to scale isn't generation — it is review. Teams that are still calling AI productivity tools without an explicit review gate are running a quiet risk. Their faster outputs are accumulating into a knowledge base nobody has actually approved.

What the gate looks like inside an AI-native marketing team

This is where the engineering metaphor stops being a metaphor and starts being a workflow. In an engineering org, every code change runs through a pull request: a labeled reviewer, an explicit diff, a pass-or-fail signal, a comment thread that gets resolved before merge. Code that does not go through that gate does not ship to production. The discipline is structural, not aspirational.

The translation to marketing knowledge is direct. Every AI-generated brief, summary, research doc, customer-persona update, or campaign retro is a knowledge artifact. Today most marketing teams treat those artifacts as "outputs" — things that land in Drive and get used. They are actually proposed merges. They become the team's working knowledge by default unless someone reviews and approves them. The default is the problem.

The flow

Here is the shape we ship inside an AI-native marketing team. Five steps. Embedded in the team's actual workflow, not bolted on as an extra process.

The classic flow. An AI tool produces an artifact — a brief, a summary, a draft, a research doc. It lands in Drive or Slack. The person who needed it skims it, uses what looks useful, and moves on. No reviewer is named. No pass/fail signal is recorded. The artifact enters the team's working knowledge by default.

Test-drive first. Before rolling the gate out across the whole team, run a two-week dry pass. Assign reviewers under step 2 below, but let artifacts ship as usual. The reviewer logs what they would have flagged in step 3 — pass, fail, or back-edit — without actually blocking anything. After two weeks, the log tells you how many artifacts would have failed review, what kinds of failures dominate, and who the bottleneck reviewer is. Most teams find one or two artifact types are 80% of the value. Start there and extend.

The AI-native version we propose.

1. Every AI-generated artifact ships with a header. An artifact is anything that will enter the team's knowledge base — a brief, a summary, a research doc, a customer-persona update, a campaign retro. Scratch work and exploratory drafts don't count. Each artifact carries four header fields: who triggered the build, what sources the build used, what the agent was asked to do, what signal the build is supposed to produce. No header, no review possible.
2. A labeled reviewer is assigned at trigger time, not after. When the build runs, it names the human who has to sign off. That human's calendar gets a slot. Review is a step in the workflow, not a backlog.
3. The reviewer's job is a pass/fail signal plus a back-edit. Pass means the artifact is approved as-is. Fail means the reviewer either rejects it or back-edits it with corrections. The back-edits matter — see step 5.
4. Approved artifacts get a "merged" marker before they enter the knowledge base. The reviewer applies the marker at sign-off. Anything without the marker is still a draft. Nobody is allowed to cite a draft in a strategy doc, a board deck, or a customer-facing message.
5. The reviewer's back-edits flow back into the agent's instructions. This is the compounding loop. When a reviewer corrects an artifact, the correction doesn't just fix that artifact — it edits the underlying instruction file the agent reads on the next run. The same kind of mistake doesn't recur. Over six months, the agent's behavior is shaped by the brand's accumulated review history, one back-edit at a time. The reviewer is teaching the system what good looks like for this team.

Where it breaks. Reviewer fatigue. If one person owns review for every artifact, the queue grows faster than they can clear it. Either approvals get rubber-stamped, or the system slows down until the team works around it. The fix is rotation by domain — retention artifacts get reviewed by the retention owner, paid-media artifacts by the paid-media owner — plus a cap on how many artifacts any one agent ships per week. The flow does not survive on willpower. It survives on capacity discipline.

What this changes for the buyer

The teams that ship AI-generated knowledge without an explicit review gate are not deciding to skip the review. They are deciding by default. Every artifact that lands in the Drive folder without a pass/fail signal is a vote that says this is approved. Over six months, a thousand unreviewed artifacts becomes a thousand-page knowledge base nobody has actually read. The team is operating on that base every day. The strategist's quarterly plan came from somewhere.

The deeper cost is trust. Once a few of those plans turn out to have been built on bad inputs — a competitor brief drawn from an archived 2023 page, a customer summary that quietly invented a complaint, a retention diagnosis built on the wrong cohort — trust in the AI artifacts collapses internally. Operators stop using them. The team reverts to manual research, which costs the speed advantage that justified the AI install in the first place. Without the gate, the speed itself is what dies.

Three things change when the gate is in place. First, the senior operator stops absorbing context loss every time a junior strategist uses an AI artifact in a planning doc — the artifact has either passed review or it has not, and that signal travels with it. Second, the agent improves week over week, because the reviewer's back-edits are training data for the next run instead of corrections that live in someone's head. Third, the engagement becomes auditable. A reviewer with a pass/fail history is a paper trail. Without it, an AI-native engagement is a black box for both the agency and the brand.

The teams converging on a pull-request culture for knowledge — the ones at Speero last week, the experimentation leaders at Zalando and Fyxer naming the problem out loud, the AI-native marketing rooms quietly running their own version of the gate — are not slower than the teams without it. They are moving the same speed. They are refusing to merge knowledge into the same base they will operate on tomorrow until a labeled human has signed off.

The discipline AI-native marketing teams are starting to share is not about generating less. It is about merging less.