Where AI freed up hours, most of them go into adding: another feature, another workflow, another agent. Very few go into reviewing decisions that were already made. Four researched mechanisms explain it: people think of adding before removing, the person who made a decision invests the most in it when it fails, builders overvalue what they built, and everyone reasons better about someone else's problem than about their own. Three of the four come from attachment to what was built or decided, and someone from outside carries less of it. The first affects everyone, so an outside reviewer needs the removal question written into the brief. This piece covers the research, what a team can do on its own, and when a review from outside is the better tool.
Where AI saved time, the time went into more tasks
Many knowledge workers now have AI in their daily tools. What did that change in the places that decide whether a company makes money?
The best data so far says: less than it feels like. Anders Humlum and Emilie Vestergaard linked large AI-adoption surveys in Denmark to administrative payroll records. They found precise null effects on earnings and recorded hours, for workers and for workplaces, ruling out effects larger than 2% two years after ChatGPT launched. Workers did report productivity gains. The useful detail is where they went. Employers absorbed AI by reorganising tasks, and workers picked up new ones: generating content, overseeing AI output, integrating AI into their work.
That study measures pay and hours, not profit, so it does not settle the question. My own view, after 10+ years inside tech companies, is that technology was rarely what stood between a company and its margin. What decided it was what got built, what got stopped, and how sound the base underneath was. All of that could be done before AI and can be done after it. Where AI does free up hours, the open question is what they are spent on.
People think of adding before they think of removing
Gabrielle Adams, Benjamin Converse, Andrew Hales and Leidy Klotz published eight experiments in Nature in 2021 on how people improve things. In one, a Lego structure could be made stable by removing a single block. Most participants added blocks at the corners instead. In another, people were asked to improve a mini-golf hole, and those reminded that they could take things away were much more likely to do it. Klotz's summary, in the University of Virginia's account of the study: "Even with financial incentive, we still don't think to take away."
Two findings from that paper map onto how companies plan. People overlooked the better subtractive option more often when nothing cued them to consider it, when they had one pass at the problem, and when they were under higher cognitive load. And a plain cue brought subtraction back.
A quarterly planning meeting looks a lot like the no-cue, one-pass, high-load condition. The agenda asks what to build next, and the template has a column for it. It has no column for what to stop.
The person who made a failing decision is the most likely to keep funding it
Barry Staw ran the classic study in 1976. 240 business-school students played an executive allocating research money inside a company, and some were told an earlier allocation had gone badly. The participants who had personally made that earlier choice committed the most new money to it.
Michael Norton, Daniel Mochon and Dan Ariely found the companion effect for things people build. In one experiment, participants who assembled a plain IKEA storage box themselves bid 63% more for it than people who bid on the same box already assembled, and they expected other people to value it the same way.
Put those two findings in one room. The people who know the most about a feature, a process or a vendor choice are the ones who chose it and built it, and they are the ones for whom a clean review is hardest. None of this requires bad faith.
Four mechanisms, all pushing toward adding
| Mechanism | What the research found | What it looks like at work |
|---|---|---|
| 1The additive defaultAdams et al., 2021 | People search for things to add and overlook removing, more so under load and without a cue. | The roadmap has a "next" column and no "stop" column. |
| 2Escalation of commitmentStaw, 1976 | The person responsible for a failing decision invests the most in it. | The sponsor of a project asks for one more quarter. |
| 3Overvaluing what you builtNorton et al., 2012 | Builders value their own work well above what others would pay. | The team that built a module rates it as core. |
| 4The self-other gapGrossmann and Kross, 2014 | People reason more wisely about another person's problem than their own. | A manager gives a peer the advice they do not take themselves. |
Why a team cannot easily see this from inside
Igor Grossmann and Ethan Kross tested the fourth mechanism directly. Across three studies with 693 participants, people reasoned more wisely about another person's problem than about the same problem framed as their own: they were more likely to recognise the limits of what they knew, to consider other perspectives and to look for compromise. The authors named it Solomon's paradox, after the king known for his judgement in other people's disputes. Their studies used personal relationship dilemmas, so the step to business decisions is by analogy.
Dan Lovallo and Daniel Kahneman described the planning version in Harvard Business Review in 2003. Executives forecast from the inside view, built from the details of this project, this team and this plan, and neglect the outside view, which asks how projects of this kind usually turned out.
The feeling of progress does not flag any of it. Adding produces something visible every week. A review produces, at best, a shorter list. In the METR trial I covered in Why work that feels productive often moves nothing, experienced developers using AI took 19% longer and believed they had been 20% faster.
What a team can do on its own
Each mechanism has a cheap counter-move.
Put removal on the agenda, in words
In the Adams experiments a cue was enough to bring subtraction back. Add a fixed line to the planning template: what do we stop, remove or merge this quarter?
Set the review date when the decision is made
Write down which result, by which date, would mean the decision was wrong. That threshold is easier to set before anyone is attached to the outcome.
Give the review to someone who did not make the decision
Staw's effect attaches to personal responsibility. A peer from another team carries less of it. A peer still shares the budget and the leadership.
Describe the problem in the third person
In the Grossmann and Kross studies, self-distancing removed the gap. "What should this company do about this module?" is a cheap version of that.
When someone from outside is the better tool
In the studies, a cue or some distance reduced the bias. Inside a company the attachment stays, because everyone inside still built something, decided something, or is measured by something continuing. Someone from outside did not build it and did not decide it, and if they are paid for the job rather than for its continuation, their fee does not depend on which answer they give. The additive default acts on outsiders too, so the brief asks in words what to stop.
That advantage has limits. An outsider starts without context, so the review is only as good as the access the company gives: the data, the people who built the thing, the history of why it was decided. An outsider can also propose removing something whose purpose they did not understand. A careful review names the reason each item exists before it proposes stopping it.
My approach starts from what the company already has: finding where the existing people, tools and budget can give more. As talent on demand I join for a specific job and do the work people inside have no time for: review, research, documentation, ideas. Then I hand it back in writing. The aim is for the team to get time for that work themselves. In my experience it is the work that improves the deliverables, the culture, and the professional experience of the people doing it.
For investors I offer the same kind of review before an investment, as operational due diligence: an execution-readiness read, not investment advice.
Where the freed hours should go
The work that makes an organisation run better was always short of hours, and it is the work AI has now made room for:
- Planning - deciding what to build before building it.
- Changing direction - stopping what no longer pays.
- Strengthening the base - data, architecture, the processes other work depends on.
- Testing - checking that what was shipped does what it was built for.
Pick one decision older than a year that still costs money every month. Multiply its monthly cost by the months since anyone last checked it. Give it a review date this quarter and a reviewer who was not in the room when it was made.
Sources
- Adams, G. S., Converse, B. A., Hales, A. H. and Klotz, L. E., "People systematically overlook subtractive changes", Nature, 592(7853), 258-261, 2021. Source of the eight experiments, the Lego and mini-golf tasks, and the three conditions under which subtraction is overlooked more. The Klotz quote is from the University of Virginia's account of the study. A 2024 study in Scientific Reports found the effect varies with age, culture and task, so read it as a strong default rather than a universal law.
- Staw, B. M., "Knee-deep in the big muddy: A study of escalating commitment to a chosen course of action", Organizational Behavior and Human Performance, 16(1), 27-44, 1976. Source of the 240-student simulation and the finding on personal responsibility. A laboratory role-play with students.
- Norton, M. I., Mochon, D. and Ariely, D., "The IKEA effect: When labor leads to love", Journal of Consumer Psychology, 22(3), 453-460, 2012. Source of the 63% figure (bids for an assembled versus a pre-assembled IKEA storage box) and of the finding that the effect disappears when the build is not completed. A consumer-product experiment, applied here to internal builds by analogy.
- Grossmann, I. and Kross, E., "Exploring Solomon's Paradox", Psychological Science, 25(8), 1571-1580, 2014. Source of the 693 participants across three studies and the self-distancing result. The studies used personal relationship dilemmas, not business decisions.
- Lovallo, D. and Kahneman, D., "Delusions of Success: How Optimism Undermines Executives' Decisions", Harvard Business Review, July 2003. Source of the inside view and outside view distinction.
- Humlum, A. and Vestergaard, E., "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI", NBER Working Paper 33777, May 2025, revised March 2026 (first circulated as "Large Language Models, Small Labor Market Effects"). Source of the null effects on earnings and hours and of the task-reorganisation finding. A working paper on Denmark, measuring pay and hours rather than company profit.
- METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity", July 2025. Randomised controlled trial with 16 developers. Small sample, read as a signal about perception.