Most agency owners are asking the wrong question about AI. It’s not “how do we use AI to produce faster?” It’s “how do we make sure what AI produces is actually any good?” Speed is easy now. Quality control is where the real leverage sits – and most agencies have no system for it.
This post covers why editorial judgment becomes more valuable as AI handles more production, how to build human verification into AI workflows without slowing them down, why defining your quality standard is the first step to improving AI output over time, and how the STANDOUT framework helps identify exactly where that editorial layer belongs.
The Shift From Producer to Editor
AI is exceptionally good at the start of a task. Give it a brief and it produces a first draft, a summary, an initial structure. None of that requires strategic judgment. It’s pattern recognition applied at speed – mechanical work that, until recently, took human time. When AI handles the mechanical start, the constraint shifts. The bottleneck is no longer production speed. It’s quality judgment.
Journalism understood this long before AI arrived. Editors earn more than writers not because editing is technically harder, but because it requires something writing doesn’t – a standard. The editor knows what good looks like, can identify the gap between what was produced and what was needed, and has the judgment to close it. Across the audits I’ve run, the agencies getting the most from AI aren’t the ones generating the most output. They’re the ones with the clearest eye for what good output actually looks like.
“The best AI SOPs don’t just chain prompts together. They have deliberate human checkpoints – because speed without judgment is just faster garbage.”
What AI Cannot Reliably Do in Your Workflow
There’s a version of this conversation where AI becomes capable enough that human review becomes optional. We are not there. For any agency using AI in client work, there are four things AI cannot reliably do: catch its own hallucinations, identify when something technically correct is strategically wrong for a specific client, apply contextual judgment that exists in your head rather than a knowledge base, and recognise when an output meets the brief but misses the point.
My background in psychology is useful here. One of the most consistent findings in research on overconfidence is that people – and by extension, the systems trained on people’s outputs – are better at producing answers than at assessing whether those answers are correct. AI is a very fast, very confident producer. That’s the asset. It’s also the risk if there’s no editorial layer actively managing it.
Building Human Verification Into AI SOPs
The fix isn’t to slow down your AI workflows. It’s to design human checkpoints deliberately, rather than hoping someone catches problems before delivery. The best AI SOPs don’t just chain prompts together. They have explicit verification gates – moments where a human must upload a document, answer a specific question, or approve the output before the next step runs.
These checkpoints harden the output by forcing the human to actually engage with the AI’s work rather than passively passing it along. In the STANDOUT framework this sits across two pillars – Operations (O) and Team (T). Operations because it’s about where the decision gates sit in the workflow. Team because human verification is only valuable if the people doing it actually know what good looks like. Both have to be addressed together, or neither sticks.
Defining Your Quality Standard Before You Train AI on It
Here’s what most people miss: building human verification into your SOPs is also the first step toward improving your AI outputs over time. When you define what you’re checking for – what separates a good output from a passable one, what client context must be reflected, what structural rules apply – you’re creating something more valuable than a checklist. You’re creating training data.
With enough documented examples of quality output you can begin to embed those standards into your AI prompts and knowledge bases. The agencies furthest ahead aren’t the ones with the most tools – they’re the ones who invested early in defining what good work looks like across their core deliverables. In STANDOUT terms this is a Development (D) question as much as an Operations one. It’s proprietary IP built on your specific work, your clients, your criteria.
The Bottom Line
Standstill agencies use AI to produce. STANDOUT agencies use AI to produce, and then edit deliberately. The irony worth naming: define your quality standard clearly enough to check it, and you have enough data to improve your AI outputs over time. The better your editorial eye, the better your AI performs. The human judgment that catches errors today becomes the training signal that reduces errors tomorrow.
Frequently Asked Questions
Why is quality control important in AI workflows?
AI produces outputs confidently – including confidently wrong ones. Without a human verification step, hallucinations, misaligned context, and strategically incorrect outputs go straight to the client. Quality control is what separates fast AI workflows from reliable ones.
How do you build human checkpoints into an AI workflow?
Design explicit verification gates at key points in the SOP – moments where a human must upload a document, confirm a judgment, or approve an output before the next step runs. These checkpoints don’t slow the process down; they harden the output and create an audit trail.
What does good AI output quality control look like in an agency?
It starts with defining what ‘good’ actually means for each deliverable – specific criteria, not vague standards. That definition becomes the basis for human review checkpoints in the workflow. Over time, with enough examples, it also becomes training data that improves AI outputs toward your standard.
What is the STANDOUT framework in the context of AI adoption?
The STANDOUT framework is an AI audit model that maps readiness and opportunity across 8 levers of agency performance: Sales, Team, Ambition, Numbers, Development, Operations, Uniqueness, and Technology. Quality control in AI workflows sits primarily in the Operations and Team pillars.
