Marketing Just Got Its First Engineering Outage. Nobody Wrote the Postmortem.

A Growth Engineer’s Tuesday looks like this.

At 8am, a Claude-powered Competitor Research Agent scrapes ad libraries across the category, summarizes the winning angles, and drops a PDF into a shared folder. By mid-morning, a Design Agent has produced 50 variations of the week’s hero ad, copy and layout tweaked across the matrix. By afternoon, a Campaign Management Agent has launched the variants on Meta, reviewed yesterday’s performance, and paused the immediate losers. The whole stack runs on Claude, Meta, GitHub, and what’s left of Zapier.

This is a real workflow described by Jonathan Pogue, a Growth Engineer writing for other Growth Engineers. The piece is celebratory. It treats the autonomous stack as best practice. It is best practice, in the sense that it works. Read the full essay and one word never appears: governance. Not a passing reference. Not a footnote. Not a “we should think about this someday.”

That absence is the story.

The role exists. The practice doesn’t.

Engineering teams spent roughly 15 years building Site Reliability Engineering. The practice has a literature, a job ladder, a vocabulary, and a set of conventions that survive when people leave. When a deploy goes wrong, an engineer at a serious company reaches for a runbook, opens an incident, pages on-call, and writes a postmortem. None of that is invented in the moment. It is muscle memory backed by infrastructure.

Marketing now runs a comparable stack. Autonomous decisions, machine speed, customer-facing blast radius. What it does not have is the muscle memory.

Engineering practice	Marketing-autopilot equivalent (status)
Postmortem	Campaign-failure review (doesn’t exist)
Change window	Brand-launch freeze (doesn’t exist)
Eval pipeline	Brand-safety eval (doesn’t exist)
On-call rotation	Brand-incident on-call (doesn’t exist)
Observability	Did the agent pause the right thing? (no convention)

Five practices. Five gaps. Each one is the kind of thing an engineering organization would never ship a system without. Each one is missing from the marketing org running the same system in production right now.

What an outage looks like in this stack

Imagine the Campaign Management Agent at 2am. Meta’s ads API returns malformed performance data for an hour. The agent reads the error stream as collapsed CTR, applies its pause heuristic, and shuts off 30% of active campaigns across the account. By the time anyone notices, the morning spend baseline is already gone.

In an engineering org, this is a P1. Someone is paged. A status page updates. A timeline gets reconstructed. A postmortem ships within five business days, with action items and an owner. The same incident in a marketing org is a Slack message that begins, “did anyone change anything yesterday?” There is no on-call. There is no severity rubric. There is no template for the writeup, because the writeup is not expected.

This is not a hypothetical. As we explored in Marketing Agent Governance: Klaviyo Composer, platforms are now shipping autonomous campaign builders with explicit “Agent Guidance” controls. The vendors see what’s coming. The buyers, mostly, do not.

The Colgate signal

The first publicly visible failure mode has already arrived. Earlier this month, Colgate posted what readers immediately identified as AI-generated slop on its social channels. The image was off, the copy was generic, the brand voice was absent. It became a meme inside a day.

A single bad post is not an outage. It is a signal. The same generative pipeline that produced that post is sitting inside thousands of marketing orgs, pointed at the public, running with no eval gate between the model and the audience. The Search Engine Land team has started calling the broader pattern the “Bland Tax”: AI systems quietly penalize generic, repetitive content, sometimes filtering it from results entirely, sometimes absorbing it into a summary with no attribution. Brand identity dissolves at machine speed.

The Bland Tax is the slow version of the failure. The Colgate post is the fast version. Both have the same root cause: a generative pipeline running without a quality gate that understands brand.

This is a marketing problem dressed as a technical one. It is also the inverse of what we covered in AEO Is Already Commoditized: the visibility game is being won by orgs that govern what AI trusts about them. The same orgs are now learning to govern what their own AI says on their behalf.

Growth Engineers are the heroes

A clarification, because this part matters.

Growth Engineers are not the problem. They are the people who saw the autopilot opportunity first, learned the tools fastest, and shipped working systems while everyone else was still in workshops. Pogue’s essay is a generous walkthrough of a craft that did not exist three years ago. Most marketing orgs are healthier because someone like him is in the building.

The problem is the practice gap around them. A Growth Engineer running three agents in production needs the same scaffolding a Site Reliability Engineer needs: a postmortem template, a change window calendar, an eval suite, a pager schedule, and a dashboard that actually answers the question “did the agent do the right thing.” None of that ships with the model. None of it ships with the ad platform. None of it ships with GitHub.

It has to be built. By the marketing org. On purpose. Before the incident, not after.

Borrow the playbook before you need it

The convergence we wrote about in AI Governance Is Leaving the Engineering Silo is now operational, not theoretical. Marketing is the next function to inherit engineering’s hard-won discipline. The good news is that the playbook is already written. Engineering teams have spent 15 years documenting how to run autonomous systems without burning down the brand. The artifacts are public. The patterns transfer.

A marketing org that wants to skip the painful version of the curve can do five concrete things this quarter:

Write a one-page postmortem template, and run it the next time a campaign underperforms because of an agent decision. Not the next time a human errs. The next time the agent does.
Define a change window. Pick the days when the autonomous system is allowed to push net-new variants to production, and the days when it isn’t. Holidays, earnings, sensitive news cycles. Make the freeze explicit.
Stand up a brand-safety eval. Three to five tests the generated asset has to pass before it ships. Tone, claim accuracy, visual identity. Run it as a gate, not as a review.
Put someone on-call for brand incidents. Not the CMO. A named person, a defined window, a rotation. The whole point is that the response does not depend on who happens to be online.
Instrument the agent. One dashboard that answers a single question: did the agent pause the right thing yesterday, and how do I know?

None of this is exotic. All of it is overdue. Every week the Growth Engineer’s stack runs without it is a week the org is one Meta API hiccup away from a Slack message that starts “did anyone change anything yesterday.”

The first marketing-autopilot outage already happened. Somewhere, on a brand you’ve heard of. Nobody wrote the postmortem because nobody knew they were supposed to.

Write the next one.

This analysis synthesizes Jonathan Pogue’s Day in the Life of a Growth Engineer (April 2026), Search Engine Land’s The Hidden ‘Bland Tax’ That Could Erase Your Brand from AI Search (April 2026), and reporting on Adobe CX Enterprise (April 2026).

Victorino Group helps marketing teams build the SRE-grade practices their autopilot stacks already need. Let’s talk.