Which decisions should become slower because of AI?

The ones where speed was previously doing work nobody had costed. When a decision took three days, part of that delay was research, part was a second opinion, and part was the friction that let an error surface before it was acted on. A model can compress the research to seconds without replacing the other two. Four conditions mark a decision that should now take longer than it used to: the consequence is hard to reverse, the environment does not reward intuition, the only human check happens after the machine has already spoken, and the reviewer could not produce the work unaided. Where all four hold, adding delay is a control rather than a delay.

Speed was carrying more than anyone put in the business case#

The efficiency argument for AI in decision-making treats elapsed time as pure cost. Some of it is. A lawyer waiting four days for a document review is waiting, and nothing is being checked in the interval. But a good deal of ordinary organisational slowness was doing three jobs at once: gathering evidence, obtaining a second view, and giving somebody the chance to notice that the question was wrong.

Generative tools remove the first job almost completely and leave the other two untouched. The output arrives fluent, complete and formatted, which suppresses the impulse to seek a second view precisely when the material looks most finished. That is the mechanism behind automation bias, described in the human factors literature long before this technology existed. The same mechanism explains why confident-sounding output is a design property rather than a signal of correctness.

Dell'Acqua and colleagues supplied the sharpest illustration. Given the same GPT-4, 758 BCG consultants were dramatically better inside the model's competence and worse than consultants using no AI at all on a task placed just outside it. Nothing in the interface marked the boundary. The consultants who went wrong went wrong quickly and confidently, which is the combination a slower process is for.

The experiment that deliberately slowed people down#

Buçinca, Malaya and Gajos ran the direct test. Working from dual-process theory, they argued that people rarely engage analytically with each individual AI recommendation and instead develop general heuristics about whether and when to follow it. They designed three cognitive forcing interventions to compel more thoughtful engagement with the AI's explanation, and compared them against two simple explainable-AI approaches and a no-AI baseline, with 199 participants.

Cognitive forcing significantly reduced overreliance compared with the simple explainable-AI designs. The finding that matters for anyone implementing this is the trade-off they report in the same paper: participants gave the least favourable subjective ratings to the designs that reduced overreliance the most. The interventions also benefited participants higher in Need for Cognition more, so the effect is not uniform across a workforce.

A control people dislike is a control that gets removed at the first efficiency review. Anyone adding deliberate friction to an AI-assisted decision should expect the user satisfaction score to fall and should decide in advance that this is acceptable, because the alternative is discovering it in month three and reversing the design. The related point about explanations is covered in does explaining an AI decision help, where the short answer is that an explanation alone does not reliably reduce overreliance and can increase it.

One regulator has already mandated a slower decision#

Article 14 of Regulation (EU) 2024/1689 requires that high-risk systems be designed so a natural person can effectively oversee them, and it names automation bias in the text as something the overseer must be enabled to remain aware of. For one category it goes further and specifies a procedure. Article 14(5), for remote biometric identification systems under Annex III point 1(a), requires that the oversight measures ensure that

no action or decision is taken by the deployer on the basis of the identification resulting from the system unless that identification has been separately verified and confirmed by at least two natural persons with the necessary competence, training and authority.

That is a mandated slowdown with a stated reason. The regulator has decided that in a category where an error is severe and hard to reverse, one human check is insufficient, and has priced the delay in. The same paragraph then disapplies the two-person rule for law enforcement, migration, border control and asylum where Union or national law considers it disproportionate, which deserves reading alongside the rule rather than after it: the exemption falls on several of the uses with the gravest consequences for an individual. Article 26(2) puts the corresponding duty on the deploying organisation, that human oversight be assigned to natural persons "who have the necessary competence, training and authority, as well as the necessary support".

Most organisations will never be in scope. The design logic is available to all of them anyway: name the categories where a second, independent, competent human confirmation is required before action, and accept that those decisions will be slower than the technology allows.

Kahneman and Klein tell you which decisions to leave fast#

The counterweight to all of this is that most decisions should get faster, and slowing the wrong ones is its own failure. Kahneman and Klein's 2009 adversarial collaboration supplies the test. Judging whether an intuitive judgement can be trusted requires assessing two things: the predictability of the environment in which the judgement is made, and the individual's opportunity to learn that environment's regularities. Where both hold, recognitional expertise is trustworthy. Where either fails, confident intuition is not evidence of skill.

Run that test on a decision before deciding its speed. A high-volume, reversible, well-fed-back decision in a stable environment, made by someone who has made thousands of them, is a candidate for compression. A one-off decision in an environment that gives feedback years later, or none, is not, whatever the model produces in four seconds. The authors give criteria rather than a classification, so applying it to your own decisions is itself a judgement.

Four conditions that should add a deliberate delay#

• The consequence is hard to reverse. Dismissals, clinical decisions, credit refusals, publication, anything that reaches a customer or a court. Reversibility predicts how much verification a decision can justify better than anything else on this list, and it beats a risk score on the practical ground that everybody agrees on it.
• The environment does not reward intuition. Kahneman and Klein's two conditions fail: outcomes are delayed, noisy or never observed, so neither the human nor the model has learned the regularities that would make a fast answer trustworthy.
• The only human check comes after the machine has spoken. A reviewer reading a finished draft is anchored on it. If the human contribution is meant to be independent, some part of it has to happen before the output exists. This is the argument of human at the start.
• The reviewer could not produce the work unaided. Oversight by someone who cannot do the task is presence rather than scrutiny. See who supervises work they cannot do and synthetic seniority.

The four are cumulative rather than alternative. One of them is a reason to be careful. Three or four together describe a decision where the responsible design is to make the process take longer than it needs to, on purpose, and to say so in the policy rather than leave it to individual conscience.

What a deliberate delay actually looks like#

• Write the position before you read the output. Two lines on what you expect and why. It costs a minute and it converts a review into a comparison. The forcing-function literature is the evidence for this shape of intervention.
• Require the disagreement to be recorded. A review process that has never changed an outcome is not a review process. Sampling for override rate is the cheapest audit available. Human in the loop is not a safeguard sets out why presence is not oversight.
• Put a fixed interval between generation and action on the reversibility-critical categories. Overnight is usually enough. The value is in breaking the continuity between the fluent output and the irreversible act.
• Name a second competent human for the worst category. The EU has done this for one class of system; an organisation can do it for its own by listing three or four decision types and no more, because a list of thirty gets ignored.
• Track how long these decisions take, and defend the number. The pressure to compress them will come from a dashboard that treats all elapsed time as waste. Someone has to own the answer that this particular slowness was bought deliberately.

Where this sits in my own argument#

"The Decision You Never Made" (2025) put the position that the consequential choices about AI in most organisations were never made by anyone, but accumulated out of individual convenience, leaving nobody to hold to them and no moment when the position was set. Decision speed is the clearest instance. Nobody in a large organisation ever decided that approvals should now take an hour instead of a week; the tooling made it possible and the calendar filled the space. "The Architecture of Drift" gives the general form, that drift is the absence of a decision rather than the presence of a bad one.

The Entrepreneur UK column of 7 July 2026, "Why AI doesn't create bad decisions, it just exposes them faster", is the compressed version of the argument on this page: the technology is a magnifier of an existing decision process, and an organisation with a weak one now gets weak decisions sooner and in greater volume.

Attribution note. Cognitive forcing functions are Buçinca, Malaya and Gajos's term. The conditions for intuitive expertise are Kahneman and Klein's. Drift versus design and human at the start are Hirji's. Automation bias is established human factors vocabulary and is not his.

Slower is not being claimed as better#

It does not claim that slower decisions are better decisions. Buçinca and colleagues measured reduced overreliance in a controlled task with 199 participants, not improved organisational outcomes over time, and the intervention was disliked by the people it helped most. There is no field evidence that deliberately slowing a class of business decisions improves results, because nobody has run that experiment.

It does not claim the four conditions are validated. They are a synthesis of the reversibility logic in the EU's own risk tiering, Kahneman and Klein's two criteria, and this research's position on where oversight fails. Treat them as a structure for an argument, not as an instrument with psychometric properties.

It does not claim the EU biometric provision generalises. It is a specific rule for a specific Annex III category, cited here because a regulator has done the reasoning in public, not because two-person verification is proportionate anywhere else.

Key sources

• Buçinca, Z., Malaya, M. B. and Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1).
• Kahneman, D. and Klein, G. (2009). Conditions for intuitive expertise: a failure to disagree. American Psychologist, 64(6).
• European Union (2024). Regulation (EU) 2024/1689, Article 14: Human Oversight. Text read at the European Commission's AI Act Service Desk.
• European Union (2024). Regulation (EU) 2024/1689, Article 26: Obligations of deployers of high-risk AI systems.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier.
• Parasuraman, R. and Manzey, D. (2010). Complacency and Bias in Human Use of Automation.
• Bainbridge, L. (1983). Ironies of Automation.

Related SuperSkills research#

On the decision itself, human and AI decision-making, decision quality and when to override AI. On oversight, meaningful human oversight, why human in the loop is not a safeguard and the invisible work of oversight. On authority, allocating AI decision rights and who can override an AI system. On the drift argument, design versus drift.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Buçinca abstract and the text of Articles 14 and 26 were read at source for this page, Article 14 at the European Commission's own AI Act Service Desk. Resolved 6 September 2026. This page previously gave no application date, because the two published texts consulted did not agree. They still do not, and the reason is now clear: the Commission's own Article 113 page displays the unamended text under a disclaimer saying it has not been updated for the Digital Omnibus, while the Commission's Omnibus FAQ is written in the language of a proposal. Reading both together gives a position rather than a date. The high-risk rules follow the availability of standards, so they begin once the Commission confirms those are sufficiently available, with a backstop of no more than sixteen months later than originally envisaged for Annex III, reported as 2 December 2027, and twelve months for Annex I products, reported as 2 August 2028. Anyone told the provisions do not apply until December 2027 at the earliest has it backwards: that is the latest date, not the first.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.