What Mechanical Engineers Know About Friction, Failure And AI
Lessons from lubricated systems for designing and operating AI-enabled work without losing expertise, feedback and control.
A direct tool is something you consciously pick up, use, and set down. You wield a hammer with a clear intention : wall, nail, thwack, job done. Lubricants are different. They sit within the system itself, reducing resistance and changing how the system behaves.
Using LLMs often feels more like applying a lubricant than using a tool. They smooth the transitions between tasks. They make it easier to move from question to research, idea to draft, thought to words, goal to plan. They do not simply perform tasks. They make transitions easier - almost frictionless.
One advantage of thinking about LLMs as cognitive lubrication is that we inherit centuries of engineering intuition about lubricated systems. Mechanical engineers understand what friction does, what happens when it is removed, and how smooth-running systems can still fail.
What can we learn from that experience that could be applied to AI-enabled work?
Lubrication Can Hide Wear
Engineering lesson
One of the paradoxes of lubrication is that it can make a machine appear healthier than it really is. By reducing friction, heat and noise, lubrication allows components to keep operating smoothly even as wear accumulates beneath the surface. A well-lubricated machine may show few obvious signs of deterioration until the damage is already significant.
AI lesson
AI can create a similar effect.
Reports are written more quickly. Code is generated with less effort. Analysis arrives faster. Work continues to flow, which makes it easy to assume that capability is being preserved as well.
Yet that may not always be the case. People may be exercising less judgement, less reasoning and less expertise than before. Productivity remains high, while the capabilities that underpin it gradually change.
Practice: Inspect the lubricant
Engineers do not assume that smooth operation means everything is fine. They inspect the lubricant because it often reveals problems long before the machine itself does. Flecks of metal, clumping, strange smells.
Organisations should adopt a similar habit. Rather than focusing only on outputs, they should pay attention to how work is being done, where reliance on AI is increasing, and whether expertise is still being exercised. Output is easier to observe than capability.
A useful starting point is to look beyond the final output and examine the process itself. Review transcripts of human-LLM interactions, which means storing them in the first place. Look at review comments, how long reviews take, and whether reviewers are making substantive changes or simply approving work. Track how often work needs to be re-done and where AI-generated outputs require significant correction. These are often the first signs that capability is changing beneath the surface.
Lubrication Changes Failure Modes
Engineering lesson
Lubrication does not eliminate failure. It changes where failure occurs and how it appears.
A poorly lubricated machine may fail through heat, wear or seizure. A well-lubricated machine may instead suffer from contamination, lubricant degradation or hidden fatigue. Some problems become less likely but others emerge in their place.
AI lesson
AI appears likely to follow the same pattern.
Certain errors should become less common. Basic drafting mistakes may decline. Routine coding errors may become easier to avoid. At the same time, new risks begin to appear: overconfidence in plausible answers, shallow understanding, unchallenged assumptions and growing dependence on systems that few people fully understand.
The system is still vulnerable, just in different places.
Practice: Study the new failures
Engineers pay attention to the failure modes that lubrication introduces, not just the ones it prevents.
Organisations should do the same with AI. Productivity gains and lower error rates are useful measures, but they do not tell the whole story. Leaders should also look for emerging weaknesses. Which decisions are becoming harder to challenge? Where is confidence increasing faster than understanding? Which mistakes are appearing more often than before?
Those questions can be more revealing than efficiency metrics alone.
One way to approach this is to track the mistakes that are becoming more common, not just the ones that are disappearing. Record near misses as well as outright failures, and reward people for surfacing them. Review cases where AI-generated work was accepted without sufficient challenge. Look for patterns in corrections, escalations and post-implementation fixes. Pay particular attention to situations where outputs were technically correct but still led to poor decisions because context, assumptions or uncertainty were overlooked.
Friction Can Be A Valuable Diagnostic
Engineering lesson
Engineers do not always treat friction as an enemy. Resistance, vibration, heat and noise often provide valuable information about the condition of a system. A change in friction can be the first indication that something inside the machine needs attention.
What looks inefficient may also be informative.
AI lesson
Many forms of workplace friction serve a similar purpose.
The effort required to explain an idea can expose weak understanding. A difficult design review can uncover hidden assumptions. A disagreement can reveal risks that were previously invisible.
These activities consume time and effort, but they also generate information about the quality of thinking taking place within the organisation.
Practice: Preserve diagnostic signals
The aim should not be to eliminate every source of friction.
Some forms of friction help organisations learn, challenge assumptions and adapt. As AI removes effort from many activities, leaders should identify which forms of friction provide useful feedback and make sure they are not lost in the process.
Frictionless work sounds attractive. The difficulty is that some forms of friction are also sources of information.
This requires some discipline. Before streamlining a review, approval or challenge process, ask what information it currently provides. Track how often AI-generated outputs are questioned, how much discussion takes place during reviews, and whether alternative options are still being explored. If design reviews, peer reviews or decision meetings are becoming shorter, determine whether that reflects better understanding or simply fewer questions being asked. Some forms of friction act as sensors. Remove them and you lose visibility into how the system is functioning.
Lubricant In The Wrong Place Can Be Catastrophic
Engineering lesson
Lubrication is valuable in bearings and gears because it reduces friction. Place lubricant on a brake disc or tyre, however, and the consequences are very different.
Some components depend on friction to do their job. Remove it and performance does not improve. The system becomes harder to steer and control.
AI lesson
Not every activity benefits from increased automation or reduced effort.
Some tasks depend on direct human engagement. Critical decisions, expert judgement, challenge processes and accountability mechanisms derive much of their value from the fact that people must think carefully, argue explicitly and take ownership of outcomes.
These are the organisational equivalent of brakes. They are among the mechanisms that keep the system stable.
Practice: Deliberately preserve friction where control depends upon it
Engineers understand that friction is essential in certain parts of a machine.
Organisations should apply the same discipline when introducing AI. The issue is not simply where AI can be used, but where human reasoning, challenge and accountability need to remain central. Some forms of friction are not inefficiencies. They are part of how the system maintains control.
Removing them may increase speed, but it can also make the organisation harder to steer.
A practical test is to identify decisions that require clear human ownership and then check whether those controls remain intact. Record who approved significant decisions, what evidence they considered and whether alternatives were discussed. Review cases where AI recommendations were accepted without challenge. If accountability becomes difficult to trace, or if nobody can explain how a decision was reached, the organisation may already be losing some of its ability to steer itself.
Lubrication Increases Capacity, Not Power
Engineering lesson
A lubricant does not make an engine more powerful. It does not create additional energy, increase the size of the motor or fundamentally change what the machine is capable of doing.
What it does is reduce losses. Less energy is wasted overcoming friction. More of the available power is converted into useful work. As a result, the machine can often run faster, carry greater loads and operate for longer periods.
Lubrication increases capacity by making better use of the power already available.
AI lesson
AI often has a similar effect on knowledge work.
People can write more documents, analyse more information, produce more code and complete more tasks in a given period of time. Capacity increases, sometimes dramatically.
That should not automatically be mistaken for greater capability. Producing more output does not necessarily mean people understand more, exercise better judgement or develop deeper expertise.
An organisation may become capable of doing more work without becoming any better at deciding which work matters or whether the results can be trusted.
Practice: Measure capability separately from capacity
Engineers distinguish between power, efficiency and throughput because they describe different characteristics of a system.
Organisations should take the same approach when evaluating AI. Productivity metrics are useful, but they only tell part of the story. Leaders should also ask whether people still understand the work they are producing, whether decision quality is improving, and whether expertise is growing or declining.
Otherwise, increased activity can easily be mistaken for increased capability. An organisation may be producing more than ever while understanding less than before.
The easiest mistake is to measure output alone. Alongside productivity metrics, assess whether people can still perform critical tasks without AI assistance. Ask whether they can explain and defend the reasoning behind AI-generated work. Track skill development, independent problem-solving and decision quality over time. More work completed is a useful indicator. It is not, by itself, evidence that organisational capability is improving.
Conclusion
Lubrication makes machines better. It reduces wasted energy, increases capacity and extends operating life. No engineer would look at a well-lubricated machine and conclude that lubrication was a mistake.
Engineers understand that lubrication can conceal wear, introduce new failure modes and alter the signals that operators rely on to understand what is happening inside the machine. As a result, they adapt their maintenance practices, monitoring techniques and operating procedures. They do not simply celebrate the improvement and move on.
AI-enabled work deserves the same treatment.
The goal is not to preserve every human skill exactly as it exists today. Nor is it to maximise automation wherever possible. The goal is to improve the performance of the overall human-AI system, without losing skill, identity, and dignity.
Doing that well requires understanding how the different parts of the system interact. Human judgement, expertise and reasoning remain important, not because they are ends in themselves, but because they contribute to the adaptability, resilience and controllability of the wider system. If those capabilities begin to weaken, the effects may not be obvious immediately. Like wear in a well-lubricated machine, they may only become visible when the system is placed under stress.
My suggestion is: to gain the most from AI-enablement, treat it as engineers treat lubrication: as a powerful improvement that changes how the system behaves and therefore alters what we need to monitor, maintain and understand.
[ Below is a diagram trying to capture some of the signals and practices that are derived from this thinking ]
