In partnership with
The override that cost an afternoon
I spent three hours rewriting an AI-generated outline last month. The structure felt off to me-too linear, missing the tension I had in mind.
I pulled it apart, rebuilt it from scratch, and sent it to a colleague for a read. She said: the original was better.
I had overridden the tool not because the output was wrong, but because it felt wrong-and those are not the same thing.
My instinct was pattern-matching against my preferences, not against the actual quality of the work. I had wasted the afternoon on an edit that made the piece worse.
The calibration problem is not about trusting AI blindly or distrusting it reflexively. It is about knowing which signal to act on when the output makes you uncomfortable.
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
Two kinds of discomfort
When an AI output makes you want to change it, that discomfort comes from one of two places. The first is substantive discomfort-something is factually wrong, logically incomplete, or missing crucial context that only you have.
This is worth acting on. The second is stylistic discomfort-the output does not sound like you, does not match your usual structure, or feels unfamiliar in a way that your brain reads as incorrect. This is frequently not worth acting on.
The problem is that both feel identical from the inside. Your unease does not come labelled.
The probe question
There is one question that separates them: can I articulate what is specifically wrong? Not generally wrong-specifically. If you can name the error (the timeline is inverted, this claim needs a qualifier, the tone misreads the audience), override. If you cannot-if the objection is closer to 'I wouldn't have written it this way'-pause before you rewrite.
This is the calibration model. Three positions: approve when the output is correct; override when you can name the error; annotate when the discomfort is real but unresolved-flag it, return to it with fresh eyes before committing to a rewrite.
What changed for Ines
Ines, a content strategist at a mid-sized agency, was spending roughly four hours a week rewriting AI drafts that her team had already reviewed. She introduced one rule: before rewriting any section, write a single sentence describing what is wrong with it.
Within two weeks, she found she could not write the sentence for roughly 40% of her planned rewrites. She approved those sections instead.
Her revision time dropped by half, and her editor-who reviewed the final pieces-noticed no quality difference.
The rule did not make the AI better. It made her judgement more precise.
What research on human-AI disagreement shows
A 2022 study by Logg, Minson, and Moore at Harvard Business School, published in Organizational Behavior and Human Decision Processes, found that people systematically overrode accurate algorithmic advice when it conflicted with their own prior judgements-even when the algorithm had a demonstrably better track record.
The effect was strongest when people felt they had relevant expertise. Crucially, the study found that asking participants to explain their override before making it reduced erroneous overrides by 29%, without reducing appropriate ones.
The mechanism was the same as Ines's rule: forced articulation separates real objections from preference.
Your support queue gets a head start every morning.
Viktor reads overnight tickets, tags them by product area, summarizes the patterns, and posts a brief in #support. The agent picks up the queue already triaged. The PM sees recurring requests rolled up by Friday.
One sentence before every rewrite
You do not need a new workflow. You need one sentence.
Before rewriting any AI output, write down what specifically is wrong with it. Not what you would change. What is wrong. If the sentence comes easily and names something real, rewrite. If it stalls-if the best you can produce is 'it doesn't feel right'-approve the output and move on.
This is not about deferring to AI. It is about being honest with yourself about whether your edit is improving the work or just making it more familiar.
I have written more about where this line sits across different kinds of work at
Reply and tell us: what do you usually override? We are curious whether the pattern holds.