What does my actual QA process look like?
When I wrote the original article, most of it covered the “why” of AI detection in B2B content, how the problems of the B2B EdTech buyer are unique and severe, and how I rebuilt our entire QA process after a prospect referenced GPTZero on a sales call.
This post will cover the “how.”
Many of you messaged after that original post and asked what the workflow actually looked like. Honestly, the workflow looks like a pretty dull checklist. It takes us three people to ensure content quality. What the workflow looks like is a routine we’ve developed over six weeks. Two tools are now non-negotiable to support the workflow.
Here is what the workflow looks like. However, there’s a catch. This is designed specifically for a certain scenario: B2B EdTech buyers who understand professional AI literacy and three writers with varying degrees of experience with AI tools.
The problem I was actually trying to solve
My three writers fall somewhere on a spectrum. There’s a writer who uses AI frequently and tends to submit work that requires little-to-no revisions prior to submission. Then there’s a writer who never uses AI, which creates a bottleneck on our editorial calendar. Finally, there’s a writer who uses AI in thoughtful ways and generally submits work that meets or exceeds my standards with little to no revisions.
Given these different working styles of AI, creating a single policy for all three writers is impossible. And while an honor system may be fine for some organizations, it’s not going to work for ours. What I really need from my writers is transparency. I need to know exactly what’s being submitted to me, not just what the final product looks like. For example, there could be a very diligent writer who uses AI throughout the writing process but edits extensively before submitting. They could still submit a perfectly clean-looking draft to me that would raise red flags among a provost.
To put it bluntly, trust in your writers is not an acceptable substitute for developing a process.
How the workflow works now
Each writer runs their drafted content through Walter Writes before sending it to me. I chose Walter Writes primarily due to the integrated AI detector. Using Walter Writes allows my writers to see their likelihood of AI-generated content from four detectors: GPTZero, Turnitin, Originality.AI, and Copyleaks without having to leave the tool. Reducing friction for my writers makes them much more likely to follow each step consistently.
Walter Writes has three rewrite levels: simple, standard, and enhanced. My default level for the team is standard. I only use enhanced if a section of the written content returns as flagged by one of the other tools after the writer passes the first round. I’m not trying to create unrecognizable content. I’m simply trying to ensure that once published, it appears as though it was generated entirely by humans, because in large part it already is.
Once my writers send their completed drafts to me, I then run each draft through Proofademic’s sentence-level detector as a second verification check. This was the single biggest change made to my editorial process by adding Proofademic to our toolkit. When I run a draft through Proofademic’s detector, it shows me which sentences within a paragraph were identified as possibly having been generated via AI algorithms and provides a rationale behind why each sentence was flagged. When a particular section of content continues to trigger even after my writers have passed their version through Walter Writes, I can identify the exact sentences that triggered and provide feedback based on those findings. For example: “You’ve got three sentences that read like they were generated using patterns. Take note of how they were constructed.”
GPTZero gave me a number. Proofademic provided me with something I can turn into an actionable directive for my writers.
What six weeks of running this has taught me
As it turns out, the writer who used AI so heavily has altered her processes since running her work through the accountability layer created by this new workflow. She’s done so without needing direct intervention on my behalf. She knows she’s being held accountable and is adjusting her habits accordingly, even though I never spoke directly with her about changing her behaviors.
On the flip side, the writer who hardly ever uses AI hasn’t had to adjust her workflows whatsoever. Because her work has always cleared both tools, this new workflow has given me valuable data regarding how effective this new workflow is.
Finally, the writer who uses AI thoughtfully loves the workflow. In fact, she uses Walter Writes as another way to validate her own instincts about whether or not she needs to revise her drafts before submitting them.
What I tell my writers
The point of all of this isn’t to clear whatever artificial barrier you set up in terms of passing some sort of test. The end result should be publishing content that you believe would be trustworthy enough for a provost to consume prior to meeting you during a sales call. The tools help us demonstrate that we’re holding ourselves to that standard.
In terms of EdTech specifically, your buyers are also the ones who determine their institutions’ AI policies. So they’re your target audience. The workflow exists because they do.