The Candy Machine Trap

Why leaders use AI for the work that matters least, and the allocation move that changes the odds.

A governance professional in one of our cohorts recently told me what she had asked AI to do that morning. ‘Please generate me a summary of these board meetings.’ The request had been essentially the same every week for the past quarter. The board packs came in; she fed them into AI; she got a readable summary in three minutes; she sent it to those people who needed it.

The request was fine. The output was probably fine. What caught my attention when we started talking was that most of her AI week consisted of variations on the same ask. Summarise this document. Polish this email. Rewrite this paragraph to sound less formal. When we audited her last ten AI interactions together, all ten fell into one quadrant of a 2×2 matrix, which we will get to in a minute. The same quadrant.

She had been struggling with strategic work that had not touched AI at all: a governance-reform proposal her board was about to push back on, and a diagnosis of why her organisation’s risk committee kept talking past each other. She had the most powerful thinking tool most leaders will ever have access to, and she was using it like a vending machine. Press the button. Receive a snack. Repeat.

This is what I call the Candy Machine Trap.

Two distinctions, stacked

Two older frameworks sit underneath this Candy Machine trap and you need both of them.

Cal Newport (2016, Deep Work). Work divides into deep and shallow. Deep work is sustained, uninterrupted focus that pushes cognitive limits and creates lasting value. Shallow work is dip-in, dip-out, fragmented, and low in cognitive demand. Necessary, but different. A simple test: would twenty-five minutes of uninterrupted attention materially improve this task? If yes, it is deep-eligible; If not, it is shallow.

Elizabeth Grace Saunders (2013, The 3 Secrets to Effective Time Investment). Activities divide into investment and maintenance. Investment activities compound forward; the benefit reproduces. Maintenance activities keep things running and are then done. An hour of investment is worth many hours of maintenance because investment pays back beyond the moment.

Saunders’s full framework goes beyond this two-cell distinction. She introduces additional categories, including neutral time (necessary but neither building nor draining) and optimisation time (sharpening the systems that produce the investment hours); the fuller version is worth reading in her own work. For the AI-allocation question, the investment-versus-maintenance split carries the analytical load. Simple test: does this hour pay back in hours I do not have to spend next month? If yes, investment. If not, maintenance.

Elizabeth Grace Saunders’s time-allocation framework: investment activities compound; maintenance activities give diminishing returns.

The AI Allocation Matrix

Stack Newport’s deep/shallow on one axis and Saunders’s investment/maintenance on the other. This creates four quadrants. Every AI working session falls into one of them. I call this the AI Allocation Matrix.

The AI Allocation Matrix, is a 2×2 that plots any AI working session by cognitive demand (deep/shallow) and time-allocation choice (investment/maintenance). Four named quadrants: Compounding leverage (top-left, target), Apple polishing (top-right, high-effort + no-compound), Quiet leverage (bottom-left, underrated), and The Candy Machine (bottom-right, the named trap — routine email, document polish, meeting summary, the default destination).

Four quadrants. The strategic leverage of AI sits in the top-left, where AI’s reframings, alternative viewpoints and adversarial scrutiny can elevate the quality of an investment-deep hour. The Candy Machine sits in the bottom-right, where the work is small, the rewards are fast, and AI is reliably good at what is being asked of it.

Most leaders’ AI hours land in the bottom-right. Almost none land in the top-left.

The Matrix is a diagnostic, not a prescription. It does not tell you that the bottom-right cell is bad. The Candy Machine is fine. Frictionless tools that do maintenance work fast are, in themselves, not the problem.

The problem is the displacement: the candy-machine hour quietly crowding out the deep-investment hour, with neither hour tracked, so the displacement is never seen.

What pulls the hours into the bottom-right cell

The Candy Machine Trap is the bottom-right Maintenance × Shallow quadrant of the AI Allocation Matrix where most senior leaders’ AI hours cluster that is routine email, document polish, meeting summaries - the default destination.

Four forces, each rational in the moment, pull AI hours into the bottom-right cell.

Frictionlessness. The marginal cost of an AI query is zero. Any task for which AI might help immediately becomes a candidate. The task with the lowest resistance wins, and your inbox has much lower resistance than your strategy paper. This is the physics of attention. Attention flows to the cheapest available use.

Immediate reward. The Candy Machine gives you output in seconds. Investment work gives you output in hours or days, sometimes never quite. The reward gradient points the wrong way. Every time you feel a small hit of competence after a sixty-second AI interaction, that reward trains your attention towards the next sixty-second interaction, not towards the three-hour strategic piece.

Availability. What you used AI for last week is what comes to mind when you open AI this week. The habit reinforces itself. The longer you are in the bottom-right cell, the harder it becomes to imagine AI anywhere else. The governance professional had not considered using AI for the governance-reform proposal. Not because she had considered it and rejected it, but because her pattern of AI use had become what AI is for in her mental model. A Pavlovian habit; the sugar-hit working as intended.

Gresham’s law of work. Gresham’s law of money, bad money drives out good, has a cousin in time: easy work drives out hard. Pre-AI, that pull was modest because routine work still cost effort. AI removes the effort. Easy reaches for easy; hard stays where it was; the trap deepens structurally.

Four forces, all rational, all compounding. The trap closes quietly, and what closes it is not weakness. It is the same attention economics that governs the rest of your week, sharpened by a frictionless tool.

The corrective: three moves

The corrective is allocation: see where the hours go, refuse the easy-first impulse, and place AI deliberately where the leverage is. No amount of ‘I should use AI better’ will get you there; the move is mechanical.

See where the hours go. Most leaders cannot answer the question ‘Of the AI hours you spent this week, what fraction went to deep-investment work?’ The first move is to make the trap visible. List your last ten AI interactions, specifically: the summary of the Monday board pack, not ‘meeting summaries’; the rewrite of the email to Acme, not ‘drafting emails’. Plot each on the Matrix. Most honest audits show eight or nine of the ten in the bottom-right cell. That is the data.

Refuse the easy-first impulse. When AI is reached for a maintenance task, ask first: ‘Is there a deep-investment task I am avoiding by doing this?’ Often there is. The maintenance task can wait an hour. The discipline is to interrupt the candy-machine reach long enough to notice what it is competing with, not to suppress it.

Apply AI where the leverage is. Pick three tasks in the coming week that belong in the top-left cell. Investment-deep work where AI in the loop would change the quality of your thinking rather than replace it. Not three topics. Three specific named tasks. A strategic diagnosis you have been putting off. A decision memo you have been avoiding writing because the argument is hard. A developmental conversation you are preparing for where you want to stress-test your approach.

Three moves, each capable of changing the shape of your AI week when done honestly.

The discipline: three habits

The three moves above outline what to do this week. The three habits below show how the practice embeds.

A weekly allocation review. Five minutes on a Friday. Review all the week’s AI prompts. Assign each to a quadrant on the Matrix. Note which received the most hours, and which got the least. The picture builds over months. Quarter by quarter, your distribution shifts, or it doesn’t, and you have data on the gap between intent and behaviour.

A deep-investment booking. At least one calendar block per week where AI is brought to a strategic question, not a routine one. The block is held against the candy-machine pull. If the block is recurring (Wednesday morning, every week), the discipline is in protecting it. Strategic work always has a reason to be deferred; the calendar holds the line.

A friction-asymmetry log. A few lines, captured in the moment when the easy task is reached, because it is easy, and a deep-investment task is visibly being avoided. The log surfaces the trap mechanic in your own behaviour, so it can be interrupted next time.

The discipline is to make AI hours appear where they generate compound returns, against the gravitational pull that puts them where they do not.

What this move does not solve

The top-left cell is harder. It takes longer and produces rougher first outputs that need more work. The work itself often sits uncomfortably, because the tasks that belong there are the ones you have been avoiding, and that avoidance does not disappear just because you have re-allocated an hour.

The top-left hour will feel harder, longer, and rougher than the candy-machine version. The reason to spend the hour is that it compounds. An hour spent using AI to stress-test the strategic diagnosis you have been avoiding produces thinking that serves you for months. An hour spent on the board-pack summary produces a summary. Both hours are real work. Only one of them pays back beyond the moment.

The reason to make the allocation move is that the Candy Machine is where the hours go by default, and that default leaves the compound-interest hours on the table.

Where this sits in the series

This is the opener of an arc on Mindset Reset for AI in leadership: allocate, aspire, invest, sustain. Four moves that describe how leaders shift their AI use from where it defaults to where it earns its keep.

The piece you have just read is the allocate step: placing the AI hour in the Matrix cell where it compounds.

The piece that follows is the aspire step (virenlall.com/gaussian-challenge-ai). Once an hour has been placed in the investment-deep cell, the next discipline is to ask whether the task itself is the right one, or whether a different framing makes the question disappear.

Then comes the invest step (virenlall.com/ai-investment-loop): the practice within the investment-deep hour. A small loop you run each time, turning the hour into deeper thinking and a captured artefact.

And the sustain step: the posture that keeps you in these disciplines when the candy-machine alternative would feel faster, easier, and more visibly productive in the moment.

Later in the series, a piece names the long-run cost of staying in the candy-machine cell (virenlall.com/skim-tax): the bill that compounds across years and levels of an organisation. Today’s allocation move is the upstream fix.

Where this has been tested

The Allocation Audit has been tested with chairs and governors in further education, deep-tech founders at the Royal Academy of Engineering, and manufacturing leaders in the UK government’s Made Smarter programme; it forms a design backbone for our COO, CDO and CPO open programmes.

Closing

Go back to the governance professional. When we looked at her ten interactions, she did not dispute the diagnosis. The pattern became obvious to her once the AI Allocation Matrix was on the board. What she found harder was choosing three tasks for the coming week. Each was one she had been avoiding, for reasons that had nothing to do with AI. AI had given her a way to look productive while the actual thinking waited.

That is the trap: a quiet accretion of hours spent in the cell where the work is smallest and the rewards are fastest, while the work that would have mattered most sits, unprompted, unallocated, indefinitely deferred.

The move is small. Audit your ten, pick three, commit aloud, and do one of them before the week is out. That is how you start.

“AI gravitates to the work that matters least, away from the work that matters most.”

About the author

Viren Lall FRSA is Managing Director of ChangeSchool LDN, an executive education design and delivery company that has worked with over 6,900 leaders and entrepreneurs across 39 countries. His current work focuses on how senior leaders use AI to improve judgement, allocation and organisational learning.

Viren Lall, Managing Director, ChangeSchool LDN (2026). virenlall.com/candy-machine-trap