The problem with ad-hoc review
You know you should check AI output before it goes out. But under deadline pressure, “I’ll review it” becomes a quick scan. A quick scan misses the fabricated statistic buried in paragraph three. A written verification workflow doesn’t.
Most communicators who use AI have some version of a mental note to review the output before publishing. The problem is that mental notes are not reliable under time pressure, fatigue, or the cognitive bias that makes polished-looking text feel trustworthy. If your verification process exists only in your head, it will compress under pressure until it’s gone.
A written verification workflow solves this. It takes the decision-making out of the moment — you don’t have to decide what to check each time, because you’ve already decided. You just run the checklist.
Key Insight
A good verification workflow is fast because it’s specific. It doesn’t say “check everything.” It says exactly which categories of content require checking, and exactly how to check them. Specificity is what makes it usable under deadline pressure.
Not all content carries the same verification risk, and treating it all the same wastes time. A tiered model lets you apply the right level of scrutiny to the right content — quickly.
Tier 1: Always verify. Statistics, figures, percentages, and data points. Named quotes attributed to real people. Citations, studies, or reports referenced by name. Regulatory or legal details. These require source confirmation every time, regardless of deadline or how plausible the output looks.
Tier 2: Verify before it goes external. Factual claims about your organization, clients, or partners that didn’t come from your own brief. Descriptions of events, partnerships, or product features that AI may have inferred rather than been told. For internal drafts, these may be lower priority; for anything that reaches a journalist, client, or the public, they need a check.
Tier 3: Light review only. Structural elements, transitions, formatting, tone adjustments, and copy variations where you provided all the underlying facts. These are lower-risk because the factual inputs came from you — AI was just doing something with them. A read-through for voice and accuracy is sufficient.
Before you close out of AI and move to verification, try this: ask AI to flag the claims in its own output that should be independently verified. The prompt is straightforward: “Review the draft above and list every specific fact, statistic, quote, or citation that I should verify before publishing. Include the exact claim and note why it might be inaccurate.”
AI won’t catch everything, and it may flag things unnecessarily, but it gives you a starting list. Think of it as a first pass, not a final answer. Your own knowledge of the subject is still the primary check.
Build your personal three-tier verification checklist. By the end of this session you’ll have a written document you can use on any AI-assisted content going forward.
List the content types you produce most often — press releases, internal memos, social posts, exec briefings, whatever applies to your role. Aim for five to eight types.
For each content type, decide which verification tier applies: always verify (Tier 1), verify before external use (Tier 2), or light review only (Tier 3). Write it down next to each type.
Draft the specific checks for each tier. Tier 1 should list the exact categories you always verify (stats, quotes, citations, dates). Tier 2 should list what triggers a deeper check (going external, named clients or partners, regulatory content). Tier 3 describes the light review standard (read for tone and accuracy, confirm factual inputs were yours).
Take a piece of AI-generated content you have access to — something you’ve drafted recently or a sample you create now. Run it through your checklist. How long does it take? What did you find?
Save your verification checklist. This is Module 3’s output and a core piece of your AI Confidence Kit. You’ll expand it when you build the full QC checklist in Module 7.
