Daniel Foss manages a nine-person engineering support team. One Friday afternoon he asked an AI tool to draft a note to a customer whose integration had failed during a product launch, costing them a weekend of downtime. The draft came back fast, fluent, and completely wrong in tone: cheerful, full of phrases like "we appreciate your patience" and "rest assured," reading like a form letter from a company that did not grasp what had happened. Daniel spent the next 40 minutes tweaking it sentence by sentence, softening this, sharpening that. When he finally read his patched-together version aloud, it still sounded like a hostage note written by a committee. So he deleted the whole thing and wrote three honest paragraphs from scratch in eight minutes. The painful part was not that the AI failed. It was that he had wasted 40 minutes refusing to admit it had failed. This lesson is about never wasting those 40 minutes again.

What This Lesson Covers

You generate output from an AI and then second-guess yourself. Is it actually bad, or are you just being picky? Is it worth redoing, or should you keep editing? Many managers waste real time tweaking mediocre output instead of overriding it, because they do not trust their own judgment about when AI has failed. Spending 45 minutes editing something you should have deleted is a false economy. Sometimes the fastest path is "delete this and write it myself."

This lesson gives you a fast, repeatable way to decide. You will learn a three-way decision framework (accept, edit, or override), the red flags that signal poor output, a time-based rule for when editing has become slower than rewriting, and a short checklist of judgment questions you can run in seconds. The throughline is confidence: you have context, judgment, authenticity, and emotional intelligence that the AI does not, and your job is to recognize when the output does not match those, then act on it without hesitation.

Why This Matters for Managers

The whole value of AI is speed. The moment you start editing line by line for 40 minutes, that advantage is gone; you would have been faster writing it yourself. But the opposite mistake is just as costly: sending output you know is mediocre because you are in a hurry. Mediocre work quietly erodes your credibility, and a bad decision built on bad output takes far longer to recover from than a clean rewrite would have taken.

The skill, then, is calibration: a clear, consistent threshold for when to keep AI output, when to fix it, and when to throw it away. Managers who lack that threshold are inconsistent. They edit one mediocre draft for half an hour and override another instantly, with no logic connecting the two. A framework makes the decision fast and repeatable.

The Override Decision Framework

For each piece of AI output, sort it quickly into one of three buckets.

ACCEPT, use as-is. The output meets your requirements, you have verified the parts that matter, it says what you meant, it sounds like you (or is appropriately professional), and it is genuinely good enough to send or present. Accept does not mean perfect. It means good enough that no remaining issue would actually stop you from using it.

EDIT, keep the structure and fix details. The general approach is sound but the specifics need work: a couple of factual corrections, minor phrasing, a tone that is mostly right but needs a small adjustment. The output has good bones and just needs customization. Edit is the right call when the skeleton is correct and only the surface needs attention.

OVERRIDE, delete and redo from scratch. Something fundamental is broken. The AI missed the point of what you asked. It produced the wrong type of output (you asked for an email and got a structured analysis). The tone is completely misaligned. It is missing critical context and reads generic. The logic is broken: conclusions do not follow from the data. It is emotionally wrong: technically correct but blind to the human reality. Or it is a sensitive moment, feedback, a difficult conversation, an apology, where people need to hear you, not an AI-polished version. Override is not a sign the AI is useless; it is a normal outcome for the subset of tasks where your authentic voice or judgment is the actual content.

The Red Flags That Signal Override

Learn to spot poor output quickly. The warning signs cluster into five groups.

• Structural problems: vague or generic language ("the company should focus on growth"), missing details that would be obvious to you, oversimplified treatment of a complex issue.
• Authenticity problems: it does not sound like you or your team's culture, it is over-polished with corporate jargon you would never use, it lacks your point of view, or it papers over real uncertainty with false confidence.
• Logic problems: conclusions that do not follow from the premises, a recommendation that ignores the obvious risk, an overconfident tone about genuinely uncertain things, or assumptions you never stated being treated as fact.
• Emotional and interpersonal problems: it reads cold when it should be warm, ignores the human impact, lacks empathy, or sounds like a memo when it should sound like a conversation.
• Context problems: wrong names or company facts, missing your specific constraints, generic advice that does not fit your situation, or no account of what your team can actually do.

One or two minor flags usually means edit. A fundamental flag, broken logic, completely wrong tone, missing the point, almost always means override.

The 10-Minute Rule: When Editing Costs More Than Rewriting

Here is the rule that would have saved Daniel his Friday afternoon: if fixing the output will take more than 10 minutes, override. A rough calibration:

• Fixing a few details: 2 to 3 minutes. Edit.
• Rewriting substantial parts: 5 to 10 minutes. This is the edit threshold; you are near the line.
• Rewriting most of it: more than 10 minutes. Override.

The simplest tell of all: if you catch yourself thinking "I would be faster writing this myself," you almost certainly would be. That thought is the signal to stop editing and start over. The 10-minute rule turns a fuzzy feeling into a clean decision.

Worked Example: A Two-Question Override Triage

Daniel turned his Friday lesson into a triage he now runs on every AI draft in under a minute. It has two gates and a timer.

Gate 1, does it accomplish what I asked? Not "is it good," just "does it do the job at all." If the AI misread the task or produced the wrong type of output, stop here and override. No amount of editing fixes a draft that answers the wrong question.

Gate 2, is anything fundamentally wrong? Is the logic broken, are core facts wrong, is the tone completely off for the moment? If yes, override. If no, continue.

The timer: estimate honestly how long the fixes will take. Under 10 minutes, edit. Over 10 minutes, override. Then two final sanity checks: would I be comfortable putting my name on this, and would I be proud to share it? If either is no, override or keep editing.

Watch it run on three real drafts from Daniel's week, with the numbers attached:

• The launch-failure apology. Gate 1: it produced an email, so it did the job. Gate 2: the tone was fundamentally wrong, cheerful where it needed to be accountable, and it missed the human reality of a lost weekend. Fundamental flaw, so override at Gate 2. Outcome: deleted, rewritten in 8 minutes. (The previous version had cost 40 minutes of editing before he gave up: a 32-minute lesson.)
• A status update to his director. Gate 1: passed. Gate 2: nothing fundamental, the structure and message were right. Timer: two numbers were stale and one sentence was clunky, about 3 minutes of fixes. Outcome: edit, done in 3 minutes. Overriding this would have wasted the AI's genuine head start.
• A draft performance note for a struggling team member. Gate 1: passed. Gate 2: technically fine, but it read like a template and said nothing only Daniel could say; for feedback this sensitive, the authentic voice is the content. Override at Gate 2, even though "fixing" looked like only a few minutes. Some override calls are about what the moment requires, not the edit time.

The numbers tell the story. Across these three, Daniel spent 11 minutes total (8 + 3 + a fresh write) and produced work he would put his name on. The old habit, editing all three toward "good enough," would have cost him well over an hour and produced two pieces that sounded like nobody at all. Triage first, then act, is faster than tinkering every time.

When the Bones Are Good: An Edit in Practice

Override gets the attention, but the more common correct answer is edit, and getting that call right matters just as much. Daniel needed talking points for a quarterly review with his leadership team: five points, each with a claim, supporting data, and a strategic rationale. The AI returned exactly that structure, five well-formed points at the right level of detail, on the right topics.

His review found three soft spots rather than fundamental ones. One point overstated a performance improvement by quoting the best-case number as if it applied across the board, when the gain was much smaller for standard cases. Another was framed tactically ("adding more integrations") when the strategic story was about building a platform other teams could extend. A third leaned on a number so small it would read as noise to the audience, when a different and stronger metric was available in the same data.

None of that is a broken draft. The structure was doing real work, and rebuilding it from scratch would have thrown away 20 minutes of genuine value to fix problems that took five minutes to correct. He qualified the performance claim with an honest range, reframed the integration point around the platform story, and swapped the weak metric for the stronger one. Five minutes, done. That is the shape of a good edit call: strong structure plus weak details equals fix the details and keep the rest.

When AI Answers a Different Question

The hardest override to spot is the draft that is fluent, accurate, and about the wrong thing. Daniel once asked for a short paragraph explaining to his team why the group was pausing new feature work to pay down accumulated technical problems. He asked for three specific things: what the problem was costing them, why they were addressing it now rather than later, and how it benefited the team specifically.

What came back was a clean, correct explanation of what technical debt is, complete with the standard analogy to financial debt, followed by generic benefits: faster shipping later, less time debugging. Nothing in it was false. All of it was useless. His team already knew what technical debt was. What they needed to understand was why growth work was being deprioritized right now, and the honest answer was not "we will ship faster someday." It was that the team was worn down, the foundation was unstable, and continuing to build on it was making them slower rather than faster.

Salvaging that draft would have meant changing the structure, the approach, and the core message, which is another way of saying writing a new one. He overrode it and wrote the real version himself: direct about the cost, direct about the frustration, direct about what the team actually got out of it. The tell was not a factual error or a tonal slip. It was that the AI had explained a concept when the moment called for addressing what people were feeling. When the draft answers a question you did not ask, editing cannot rescue it.

The Checklist You Can Run in Seconds

Formalized, the triage is five questions in order, each with a clear branch. Run them on anything consequential.

• Does it accomplish what I asked? Not whether it is good, only whether it does the job. If no, override. If yes, continue.
• Is anything fundamentally wrong? Broken logic, wrong core facts, completely misaligned tone. If yes, override. If no, continue.
• Can I fix the problems in under 10 minutes? If yes, edit. If no, override.
• Would I be comfortable putting my name on this? If yes, use it or edit as needed. If no, override, because a draft that does not represent you is not finished, it is borrowed.
• Would I be proud to share it? Not perfect, but good enough to feel fine about. If yes, send it. If no, keep editing or start over.

Anti-Patterns to Avoid

Perfectionism, editing too much. You edit every sentence, even the good ones, and a 5-minute task becomes 45. The fix: ask "would this actually stop me from using it?" If no, leave it alone. Good enough beats perfect for most work, do not edit for style when the substance is right, and a 10-minute edit cap forces the issue.

Accepting bad output to save time. You know it is mediocre but you send it anyway under time pressure. The fix: ask "would I be proud of this?" and "would this damage my credibility if people knew AI wrote it?" Bad output takes longer to recover from than the time to redo it.

No clear threshold. You are inconsistent, editing one mediocre draft for 30 minutes and overriding the next on instinct. The fix: commit to the 10-minute rule and the framework so your decisions become consistent.

Not trusting your own judgment. The AI says something and you question yourself instead of trusting your instinct that it is wrong. This one usually comes from wanting to believe the machine is objective, plus a little self-doubt. The fix: remember that you hold context, judgment, and authenticity the AI does not. If it feels off, it usually is.

The Deeper Question: Should I Have Used AI Here at All?

The override decision is tactical: is this output good enough? Underneath it sits a more strategic question that will matter more as you grow: should I have reached for AI in the first place? Not every task benefits from it. Some moments of leadership require your unfiltered, unassisted presence: a difficult conversation with a struggling team member, a genuine apology, a moment when your team needs to hear you think out loud rather than read polished prose. Start building your personal AI-use framework with three questions.

• Is this a task where AI adds value, or am I using it out of habit? If you can write it yourself in five minutes and it needs your authentic voice, just write it.
• What am I trading for the efficiency? Sometimes the answer is "nothing meaningful." Sometimes it is "authenticity" or "the thinking itself." When working through complexity is how you develop judgment, outsourcing it makes you worse at your job.
• Would I be comfortable explaining my AI use to the person affected? If you drafted someone's feedback with AI, would you tell them? If that makes you uncomfortable, the discomfort is information pointing at a boundary worth respecting.

This framework deepens considerably at Level 3, where ethical judgment becomes the focus rather than a footnote. For now, simply start noticing. When do you reach for AI, and when should you not? The override decision is your first tool. This broader question is your next one.

Practice and Reflection

Calibration is built by doing, not by reading. Over the next two weeks, work through these.

• Build your calibration. Collect three pieces of AI output next week. For each, decide accept, edit, or override, and write down which you chose and why. Then look at the pattern. Did you choose well? Are you systematically over-editing or under-editing?
• Test your threshold. Find a draft you feel tempted to edit. Estimate honestly how long the fixes would take. If the answer is more than 10 minutes, practice saying no and override instead. Notice how much time you actually save.
• Trust the instinct. The next time something feels off about a draft, pause instead of editing. Override it. Afterward, ask whether your instinct was right. This is how you stop second-guessing yourself.
• Identify your patterns. Which kinds of output do you tend to override, and which do you over-edit? Emails, analysis, strategy documents? Knowing your own tendencies makes the next decision faster.
• Run a speed test. Take something you would normally iterate on three or four times and instead make one fast call: override or edit. Track the time saved against the quality of the result.

Then spend two minutes on this. Think back over your past week and find one task, decision, or message where this framework would have changed what you did. What would you have done differently, and what would the outcome have been? Write it down. That link between the concept and your own work is where the learning actually sticks.

Related Lessons

Overriding is one move in a larger practice of human oversight, and three lessons sit directly alongside it.

• Verification Workflows comes first in sequence: it is how you check output systematically before you ever get to the accept, edit, or override call. Good verification is what surfaces the red flags in the first place.
• Feedback Loops and Iteration covers how to refine a prompt and re-run it, which is often the better move when a draft is close but not right. Knowing when to iterate rather than override is the natural complement to the 10-minute rule.
• Documenting AI Assisted Work is where you record what you overrode and why. Those notes become your calibration data over time, and they matter when someone asks how a piece of work was produced.
• Ethical Judgment in Practice takes up the deeper question raised above, moving from "is this output good enough" to "should I have used AI here at all," with the depth that question deserves.

Key Takeaways

• Sometimes AI gets it wrong, and recognizing that is a skill. Spotting failed output is a capability to build, not a personal shortcoming.
• Sort every draft into accept, edit, or override. Some flaws are fatal (override), some are cosmetic (edit), and some output is simply ready (accept). Knowing the difference is the whole game.
• Use the 10-minute rule. If the fixes will take more than 10 minutes, delete and rewrite. "I would be faster writing this myself" is your signal to override.
• Run a fast triage, do not tinker. Does it do the job? Is anything fundamentally wrong? Can I fix it in under 10 minutes? Would I put my name on it? Would I be proud of it? Five questions beat 40 minutes of editing.
• Good structure plus weak details means edit, not override. When the skeleton is right and only facts or framing are soft, fixing them in five minutes beats throwing away real value.
• Watch for the fluent draft that answers the wrong question. Accurate, well-written output can still miss the point entirely, and no amount of editing rescues it.
• Trust your judgment. You hold context, authenticity, and emotional intelligence the AI lacks. If output feels off, override it rather than second-guessing yourself into accepting it.
• Some moments demand your unassisted voice. Feedback, apologies, and hard conversations are where people need to hear you, not a polished draft. Override on principle, regardless of edit time.
• Ask whether to use AI at all. Beyond "is this good enough" sits "should I have used AI here?" When the thinking or the authenticity is the point, do it yourself.