The Polished Deck Problem: When AI Does the Thinking, What Happens to Human Judgement at Work (By Chavvi Bhingare)
It is 11:40 pm in a consultant’s apartment, and the client presentation is due at nine. Meera is two years into her career. Earlier, she would have spent the night wrestling with spreadsheets and half-formed slides. Tonight, she pastes the client’s data into an AI tool, describes the brief, and receives a sharp 18- slide narrative with a market-entry recommendation, a risk matrix and an executive summary. She tidies the formatting and sleeps by 1:30.
The next morning, the client’s regional head stops at slide 7. “Why 14 per cent, and why not the western cluster first?” Meera knows the number is in the deck. She cannot say where it came from, what it assumes, or what would change it. Her manager steps in, a little too quickly. The meeting recovers. The deadline was met, the deck was admired, and something quietly went missing.
To be clear, this is an illustrative scenario, not a documented case from any real company. But if you manage people in a consulting firm, an IT services major, a bank or a pharma sales organization, it will probably feel familiar. And it points to a question that the usual debate (“AI improves productivity” versus “AI takes jobs”) skips entirely: As AI takes over more of the thinking, what happens to the human capabilities organizations still need most: judgement, expertise and accountability?
The honest answer from research is neither alarmist nor comforting. AI can genuinely expand what people do well. But the benefit depends on how it is used, and the risks are quietest exactly where organizations are looking least: at understanding, not output.
Three ideas that explain Meera's morning
1. Offloading feels efficient, and we are not always good at judging when it is
Cognitive psychologists call it cognitive offloading: using an external aid, whether a notepad, a calendar reminder or now an AI assistant, to reduce the mental demands of a task. Risko and Gilbert (2016) reviewed the evidence and described offloading as a normal, often smart strategy. People offload when they believe it is more efficient than relying on their own minds. The catch is that those beliefs are not always accurate, so people can offload more, or less, than is good for them.
That last point matters for the workplace. Remembering a client’s phone number is a harmless thing to offload. Reasoning through why a market should be entered in a particular sequence is not. When the same tool does both with equal ease, the line between them blurs, and the person doing the offloading may not notice it has been crossed.
2. Trust in a tool can quietly replace scrutiny of it
Human factors researchers have studied this for decades in aviation and industrial control rooms. Parasuraman and Riley (1997) described how people can misuse automation (over-rely on it), disuse it (ignore it when it would help), or abuse it (when designers automate without thinking about the human). Lee and See (2004) added a crucial framing: the goal is not maximum trust but appropriate trust, matched to how reliable the system is in each situation.
Generative AI makes calibration harder than older automation did, because its failures look exactly like its successes. A wrong answer arrives in the same confident, fluent tone as a right one.
3. Skills are built by doing the hard part, and AI can quietly remove it
In 1983, the psychologist Lisanne Bainbridge wrote a short paper that has been cited ever since. Her “ironies of automation” argument: when you automate most of a task, the human is left with the difficult exceptions but has lost the routine practice that would have made them capable of handling those exceptions (Bainbridge, 1983). Automation removes the easy parts and can make the hard parts harder.
Where the real risk sits efficiency is visible, understanding is not
Put these three ideas together and a pattern appears. Organizations measure artefacts: the deck, the report, the draft, the turnaround time. AI improves exactly those. What it can erode (the employee’s grasp of why, their ability to challenge an output, their growth in expertise) does not show up on any dashboard, and shows up very late, usually when someone senior asks “why 14 per cent?”
This is not an argument against AI. It is an argument against measuring only what AI makes easy
A second scene: the first-time manager
Kartik has just been promoted to lead a team of six at a mid-sized financial services firm. Annual appraisals are due. Writing honest feedback is uncomfortable, so he asks an AI tool to draft comments for each team member. The language is balanced and professional. He edits lightly and submits.
Nothing is technically wrong. But the part of the task that develops a manager (recalling specific incidents, weighing them, deciding what he believes about each person’s performance, and preparing to defend that view in a calibration meeting) was never done. Kartik has produced feedback. He has not formed a judgement. And when a team member asks, “Can you give me an example?”, he is in Meera’s position.
In both cases, nobody was careless. The tool simply made it possible to finish without thinking.
Meera, again: the same night, a different way
Return to Meera’s 11:40 pm and imagine a small change in how she and her organization work.
She still uses AI to draft the deck. But before accepting the key recommendation, she does three things that take perhaps forty minutes in total:
- She forms her own view first. Before reading the AI’s recommendation, she writes two lines on what she would recommend and why. Now there is something to compare against, rather than a polished answer to simply accept.
- She interrogates the number. She asks the tool to show the assumptions behind the 14 per cent, then checks them against the client’s actual data. One assumption does not hold for the western cluster. She changes the recommendation.
- She prepares to be challenged. She writes the three questions a skeptical regional head would ask, and the answers.
The next morning, when the question comes at slide 7, she has an answer, and a better recommendation. She used the same tool, took slightly longer, and ended up with both a stronger deck and stronger expertise. In the language of the BCG study, she treated the AI as a collaborator whose work needed checking, not an authority whose work needed formatting.
What this means for managers, HR and L&D leaders
Individual discipline helps, but Meera should not carry this alone. Research on appropriate trust points to design and environment as well as personal habit. Some practical directions:
- Ask for reasoning, not just results. In reviews, make “walk me through how you got here” a normal question, not a challenge. If a team member can only describe the output, that is information.
- Protect some practice on purpose. Following Bainbridge (1983), if a skill matters, people need regular chances to exercise it. Junior employees should still build first drafts, analyses or client summaries unaided on selected tasks, the way pilots still practice manual flying.
- Teach calibration, not just prompting. The BCG findings and Lee et al. (2025) suggest that knowing how to use the tool is different from knowing when not to trust it. L&D programs that cover only prompting miss the harder skill: recognizing where the jagged frontier might be and verifying accordingly
- Design AI use like the guarded tutor. The Bastani et al. (2025) result suggests the same technology can help or harm depending on its setup. Where learning is the goal, prefer workflows that prompt hints, explanations and self-testing over those that hand over answers.
- Value the dissenting view. An AI-generated answer often reflects a kind of average of what it has seen. Teams that retain people willing to say “this doesn’t fit our client” keep a source of insight that no tool provides.
- Keep accountability human. Whoever signs off on a recommendation should be able to defend it. If they cannot, it is not ready, however good it looks.
The question to take back to your team
The most useful question for an organization is not “Are we using AI enough?” It is closer to this: if the tool were unavailable tomorrow, or confidently wrong, would our people still know enough to notice?
AI will keep getting better, and the jagged frontier will keep moving. That makes human judgement more, not less, important to cultivate someone must know where the edge is. Organizations that treat AI to skip thinking will move fast and understand little. Those that treat it to think further will keep the capability they need most.
References
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005- 1098(83)90046-8
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), Article e2422633122. https://doi.org/10.1073/pnas.2422633122
- Dell’Acqua, F., McFowland, E., III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science. https://doi.org/10.1287/orsc.2025.21838
- Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 1121). ACM. https://doi.org/10.1145/3706598.3713778 Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39(2), 230–253. https://doi.org/10.1518/001872097778543886
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002