There's a common assumption underlying most conversations about AI adoption: that the goal is always to move as fast as the tools allow, and that any deliberate slowing down represents a failure to fully capture the efficiency AI offers. Some of the most sophisticated AI users we've observed are doing something that contradicts this assumption directly. In specific, chosen moments, they're deliberately inserting friction, a pause before sending, a manual re-read, a moment of reflection, precisely because the speed AI makes available can bypass judgment that's still genuinely worth exercising.
This isn't caution born from unfamiliarity or hesitancy about the technology. It's a deliberate, selective practice from people who are otherwise highly fluent and fast with AI tools, applied specifically to the moments where speed carries a real risk that outweighs the time saved.
------------- Context -------------
AI's speed is generally an advantage, and defaulting to that speed across most tasks is entirely reasonable. But speed and judgment exist in some tension for a specific category of high-stakes moments, precisely because the ease of moving quickly can bypass a check that would have caught a problem if it had been given a moment of deliberate attention.
The clearest examples tend to involve moments where an AI-generated output is about to be sent, published, or acted on in a context where a mistake would be costly or hard to walk back. In these specific moments, the speed that makes AI valuable for most tasks becomes a liability if it isn't deliberately interrupted, because the natural momentum of a fast workflow can carry a flawed output straight through to action without the kind of pause that used to happen more naturally when producing that output required more manual effort and therefore more time to reflect along the way.
The sophisticated AI users who've recognized this aren't slowing down everything. They're identifying specific categories of high-stakes moments and building a deliberate pause into exactly those moments, while continuing to move at full AI-assisted speed everywhere else.
------------- Where Deliberate Friction Gets Applied -------------
A communications director at a mid-sized company described building a specific personal rule after a near-miss: an AI-drafted external communication that, in her rush to get it out quickly, nearly went out with a factual error she would almost certainly have caught with even a brief pause before sending. The near-miss prompted her to build an explicit rule specifically for external, client-facing, or public communications: a mandatory sixty-second pause and manual re-read before sending, regardless of how confident the AI-generated draft looked and regardless of how much time pressure existed in the moment.
This rule applied to a fairly narrow category of her total AI-assisted output. Internal notes, first-draft thinking, low-stakes routine communications continued at full AI-assisted speed without the added friction. But for the specific category of external, hard-to-retract communications, she deliberately built in a moment of friction that cost her almost nothing in aggregate time, since it applied to a relatively small share of her total output, but that had already caught at least two additional errors in the months since she'd implemented it, errors that would likely have gone out uncaught in her previous, fully fast-moving workflow.
The distinction she'd identified was specifically between speed applied to reversible, low-stakes actions, where a mistake costs little to correct, and speed applied to irreversible or high-stakes actions, where the same mistake, once it goes out, is significantly more costly and sometimes impossible to fully undo.
------------- The Skill of Knowing When to Apply the Brakes -------------
This is a genuinely different skill from simply being cautious or working slowly by default. It requires actively identifying the specific moments where deliberate friction is worth its cost, rather than either moving fast indiscriminately or slowing down indiscriminately. Both extremes miss the actual leverage point, which is a deliberate, selective application of friction targeted specifically at the moments where the cost of an uncaught mistake genuinely outweighs the time the pause costs.
Developing this skill requires an honest inventory of which categories of AI-assisted output in your own work carry real stakes if something goes wrong, distinct from the much larger category of output where speed carries little genuine risk, and building deliberate friction specifically into the former while continuing to move at full speed through the latter.
------------- Practical Moves -------------
First, identify the specific categories of AI-assisted output in your own work that are genuinely high-stakes or difficult to reverse if a mistake gets through: external communications, client deliverables, anything published publicly, anything with legal or financial consequences.
Second, build an explicit, mandatory pause into exactly those categories: a brief re-read, a specific check for factual accuracy, a moment of deliberate reflection before the action is taken. Keep this pause short and specific rather than turning it into a lengthy review process that undermines the speed benefit for the rest of your work.
Third, continue moving at full AI-assisted speed for the much larger category of low-stakes, easily reversible, or internal-only output, where the cost of an occasional mistake is low and doesn't justify the friction of a deliberate pause.
Fourth, periodically review any near-misses or actual mistakes that have occurred in your AI-assisted work, and use them specifically to refine which categories deserve deliberate friction going forward, rather than assuming your current categorization is complete or correct.
Fifth, resist the instinct to apply this friction indiscriminately across everything, which would sacrifice much of the genuine efficiency AI provides. The skill is selectivity, not caution as a general default.
------------- Reflection -------------
The assumption that faster is always better in AI-assisted work misses an important distinction between the vast majority of tasks where speed carries little genuine risk and a smaller, specific category of high-stakes moments where the same speed can bypass judgment that's genuinely worth exercising. The most sophisticated AI users aren't slower across the board. They're selectively, deliberately slower in exactly the moments that call for it.
Building this selectivity deliberately, rather than either moving fast indiscriminately or building caution into everything, is itself a skill worth developing, and it's one of the clearer differences between AI users who are simply fast and AI users who are fast and reliable.
Which categories of your AI-assisted work genuinely carry high stakes if a mistake gets through undetected, and do you currently have any deliberate friction built into exactly those moments, or does everything move at the same speed by default?