Don’t Blindly Accept AI Deliverables: The Three-Column Review Method

The smoothest scenario when assigning tasks to AI is when it gets it right on the first try. The most common failure mode occurs when, to save effort, you simply click "Accept"—and the errors slip directly into your final deliverable.

Here are two real-life examples. Last month, during a compliance check, AI pasted an audit screenshot for me with the note "All Passed" at the end. My finger hovered over the mouse as I wondered: Were there actually 23 PASSes or 20 in that screenshot? Was there a WARN hidden in one of the prompts? If I hadn’t looked closely, that screenshot would have ended up in my weekly report. Here’s another: I asked AI to generate a cover image. The script reported "HTTP 200," but when I squinted at the screen, the loaded page was a 502 error. Even a 200 status code can be deceptive.

These two pitfalls share a common trait: AI packages up something that "looks like a result" and hands it to you. The 30 seconds you save on judgment now will cost you multiples of that time later to fix.

My solution is simple: divide every AI delivery into three columns, sorting before acting.

What the Three Columns Look Like

When you receive a deliverable, don’t rush to paste it. Instead, mentally categorize it into three tiers:

**Full Rewrite**: The direction is wrong, or it conflicts with your original request. For instance, you asked for a "one-pager for the boss," and it gave you a 3,000-word essay. In this case, don’t edit it. Send it back for a redo, clearly explaining "where it deviated."

**Partial Edit**: The general direction is correct, but there are local issues—incorrect terminology, unsourced data in a section, or formatting drift. This is the main event: provide clear, complete feedback and let AI fix it itself. For example, "In the second paragraph, replace 'significantly improved' with specific numbers; if no data is available, write 'N/A'."

**Mental Fill-in**: Gaps that AI didn’t write but you must verify. In engineering deliverables, these are called missing edge cases—error branches, boundary values, empty data. In writing, it’s things like "the third part it promised is completely missing." This column is most often overlooked and also the most dangerous.

Two Rules to Speed Up Sorting

1. **Content over format**. First judge "is it correct?" then look at "is it well-written?" Formatting issues can always be cleaned up in one pass at the end; content errors become more expensive the later they are discovered.

2. **Provide only 1–2 points of feedback at a time**. Plainly put: AI’s capacity to absorb feedback is lower than you think. If you bundle five suggestions, it often only fixes the first one and ignores the rest. Feed feedback in batches, verifying the results of each batch before sending the next.

When Not to Use This Method

Don’t idolize this method either. There are two scenarios where I skip the sorting process: first, during the exploration phase, where the goal is divergence and being picky kills the discussion; second, when you intend to fully outsource the task and follow its structure anyway, in which case sorting just adds unnecessary drama. Also, for small, one-off tasks (like translating a single sentence), the cost of sorting exceeds the time saved.

Five-Step Quick Check

1. Pause for 10 seconds upon receiving the deliverable; do not paste it immediately.

2. Run through the three columns: Rewrite / Partial Edit / Mental Fill-in.

3. Prioritize content over format; focus on "correctness" first.

4. Provide only 1–2 points of feedback at a time, including the original request. Don’t say "you know what I mean."

5. Only deliver after it passes your own check. AI saying "done" doesn’t count; you must verify it with your own eyes.

Three Common Failure Points

**Trusting its "Completed" status**. Just because AI says it has finished all corrections doesn’t mean it corrected them correctly. Last time, it "fixed" a SQL query, but upon saving, I saw it had deleted the entire WHERE clause. Its reported completion rate and actual completion rate are two different things.

**Hoarding feedback for a final dump**. Saving up 10 points until the end overwhelms AI, and you’ll forget which point referred to what. Feedback should be like feeding someone: small bites, served hot.

**Forcing a "Partial Edit" after two failed attempts**. If the same deviation persists after two rounds of editing, it means we are not on the same wavelength. At this point, escalate to a "Full Rewrite," restating the background, requirements, and constraints. This is much faster than continuing to patch.

The Three-Column Method isn’t mysterious: Rewrite, Partial Edit, Mental Fill-in—just these three boxes. Combined with the "Relay Notes" (switching sessions only when tasks get too long) and "Acceptance Statement Method" (itemized verification before closing) I wrote about recently, the AI delivery pipeline—accepting tasks, executing them, and verifying them—is now closed-loop. Before handing off work to AI today, install these three boxes in your mind. The time you save will be enough for an extra cup of coffee.