Someone Still has to Sign Off
There is a common sort of meeting happening in all sorts of companies all over the world at the moment. Someone presents something like a strategy, a report, an analysis, or a proposal. When asked where it came from, they say: “I ran it through the AI and it came back with this.” Everyone in the room nods, the document gets approved, the topic moves on and finally the meeting ends.
But, does anyone ask the one question do you stand behind it?
From what I am seeing, often times not, and this presents an accountability problem (not a technology problem). To me, this is one of the defining human failure modes of the AI age.
Microsoft’s 2026 Work Trend Index 1, based on surveys of 20,000 knowledge workers across ten countries, found something that sounds reassuring on first read: 86% of AI users say they treat AI output as a starting point, not a final answer and that they stay responsible for the thinking. That is almost everyone. What a relief!
Then you look at the second number. Only 19% of those workers are classified as “Frontier Professionals” meaning the people who are actually doing this with any real skill and intentionality. The other 81% believe they are evaluating AI output, but the truth is many are just relaying it.
This is exactly what the accountability gap is, and it’s widening as AI execution accelerates 2.
The speed trap
When AI can produce a research memo in four minutes and a ten-page analysis in twelve, the pressure to move fast is enormous. The bottleneck is now the evaluation of work, not the production of it. Evaluation is slow and uncomfortable because it requires you to understand the thing you are reviewing, not just approve the format of it. You must catch errors in paragraphs, assumptions buried in methodology (often wildly incorrect), missing context that changes the conclusion. The evaluation step requires you to explain why, not just what.
Many people are skipping this step from time pressure, habit, and a confused sense of what “using AI” actually means. Using AI to produce a document is not the same as using AI to help you think about a problem. The first removes the work. The second augments the judgment.
The distinction is so important, because accountability for the output will sit squarely with the individual. When your name is on the memo, it means your judgment is on the memo. The tool that generated it is not professionally, ethically, or legally responsible for its contents. You are.
Three questions before you sign off on AI-generated work
Before you sign off on AI-generated work take a few minutes to ask these questions. It is so simple but right now, it seems most people skip them entirely.
First, the explanation test. Can you explain the reasoning, not just the conclusion? For example, not “the analysis suggests we should enter the market,” but why the model reached that conclusion, what it weighted, what it left out, and whether you agree with those trade-offs. If you can describe the conclusion but not the logic, you have not evaluated the work properly, you’ve only read the summary.
Second, the person test. If the person most affected by this decision could see how it was produced (the AI prompt, the raw output, the amount of human review it received) would you be comfortable? This question is clarifying because it forces you out of the abstraction of “AI-generated” and into the specific reality of the output’s consequences. Someone – an actual human – is on the receiving end, be it a hiring decision, a client recommendation, or a communication to the board. Don’t they deserve actual judgment applied to their situation, not a pattern-matched approximation of it?
Third, the correction test. What did you catch? What did you change? What did you push back on? If the answer is “nothing, it was very good,” then either the AI was producing genuinely superior work that even a careful human reviewer could not improve (not likely) or you were not reviewing it carefully enough.
An argument about what you owe the work before you sign it is worth making plainly: using AI is fine. Signing off without understanding what you are signing is not ok.
The skill atrophy problem
Microsoft’s research found that Frontier Professionals (the most effective AI users) are more likely than others to intentionally do some work without AI to keep their skills sharp. 43% of them report this as a deliberate practice. Only 30% of other AI users do the same 3.
This shows that the capacity to evaluate AI output depends on the capacity to do the underlying work. A strategist who has stopped doing strategy without AI loses the reference point for what good strategy looks like. A writer who has stopped writing without AI loses the ear for what a weak sentence sounds like. A data analyst who only reviews AI models loses the instinct for what suspicious numbers look like. You get my drift.
Judgment atrophies if you don’t use it – it’s not a fixed asset. The people who maintain good judgement are the ones who treat AI as an augmentation tool rather than a replacement engine, and who understand that the moment you can no longer evaluate the output, you have outsourced your accountability 4.
The anxiety running through the workforce right now (40% of knowledge workers fearing job loss to AI 5) is pointed at the wrong target. The roles most vulnerable are those built around the reliable execution of well-defined tasks. The roles most durable are those built around the ownership of consequential, ambiguous outcomes. That ownership requires judgment, which requires consistent practice by doing some of the work yourself, even when you don’t really have to 6.
What to do this week
Pick one piece of AI-generated work that you signed off on this week (a document, an email, an analysis, a recommendation) and apply the three tests.
Can you explain the reasoning, not just the conclusion? Would you be comfortable if the person most affected saw how it was produced? What did you catch and change?
If you can answer all three with confidence, well done! If you can’t, well, you now know what the work actually is.
The Frontier Professionals in Microsoft’s research are better because they have maintained the judgment to use AI well. They pause before starting a task to decide what belongs to AI and what belongs to them. They keep skills sharp on purpose and they understand that the expanding availability of AI execution raises the premium on human evaluation and that the premium is paid in attention, not time.
On courage
Owning work that AI helped produce means being willing to be wrong in public. In other words, you, as the human evaluator who missed an error takes responsibility for missing it. The human sign-off that approves a flawed analysis owns that approval. The executive who presents AI-generated insight as their own judgment is accountable for that judgment’s quality.
It’s uncomfortable, sure but it’s also not really optional. The question that AI cannot answer for you is: do you stand behind this?
- Microsoft 2026 Work Trend Index (May 2026): Based on 20,000 AI-using knowledge workers across 10 countries, 86% say they treat AI output as “a starting point, not a final answer” and that they “stay responsible for the thinking.” But only 19% are classified as Frontier Professionals — the workers actually doing this with skill and intentionality. The other 81% believe they are evaluating AI output. Many are relaying it. (Source: microsoft.com/worklab) [↩]
- AI accountability crisis in organisations: 67% of organisations that experienced significant AI-related incidents had no formally designated accountability owner. Accountability distributed so broadly it belongs to no one. (Source: Corporate Compliance Insights / ISACA, 2026)
[↩] - Frontier Professionals vs. the rest (Microsoft WTI 2026): Top AI users are more likely to “intentionally do some work without AI to keep their skills sharp” (43% vs. 30%) and to “intentionally pause before starting work to decide what should be done by AI versus a human” (53% vs. 33%). Accountability as a daily practice, not just a stated value.[↩]
- Ethan Mollick, Wharton (OneUsefulThing.org): “The skills that are so often dismissed as ‘soft’ turned out to be the hard ones.” Management judgment — understanding the problem well enough to direct AI and evaluate its output — is the differentiator. Not technical fluency.[↩]
- Job displacement anxiety, 2026: 40% of knowledge workers fear losing their job to AI. The irony: the roles most resistant to displacement are those requiring ownership of ambiguous, consequential outcomes — the exact capacity many workers are quietly outsourcing to AI. (Source: Metaintro, 2026[↩]
- Yale Insights: AI disruption is “hitting before careers can start” — entry-level knowledge workers, who lack the experience base to evaluate AI output critically, face the steepest challenges. Judgment is a developed capacity, and many workers have not yet had the chance to build it.[↩]
