Companies are celebrating the work AI can finish in hours, while ignoring the human effort needed to check, repair, manage and live with the results.
Curatorโs Note: An examination of the hidden cost of AI-driven work, the unfair rules behind modern productivity targets, and a different way to build companies where human ability is valued rather than consumed.
Introduction
I read the discussion among engineering managers and found a problem much bigger than employee burnout. We have changed how work gets done, but kept the old rules for judging people. Machines produce more, managers demand more, and employees carry the difference. Something in this arrangement needs to change.
We Are Counting the Machineโs Speed but Ignoring the Human Cost
The original discussion began with a question about why engineers were becoming tired even though AI was helping them finish work faster. One manager described how expectations had changed. Another said some people were hoping to lose their jobs just to get time to recover.
That sentence says more about a workplace than any productivity report. When an employee begins to see unemployment as a chance to rest, the company has a serious problem.
The mistake is not that management wants better results. Every business has that right. The mistake is expecting more from people without finding out what the extra work is doing to them. A faster machine does not give a human being a new nervous system, more hours in a day or unlimited attention.
There is a simple question that managers should ask before setting a new target. Where exactly did the saved time go? If an engineer completed a task in four hours instead of eight, did the company use the remaining four hours to improve the product, fix old problems or help the team learn?
Or did another task arrive before the person could even close the first one? This is where the meaning of productivity gets lost. Saving time has value only when that time is used wisely.
Filling every free hour with another demand does not necessarily create more value. It may simply increase the amount of unfinished work, decisions and pressure carried by the same employee.
The first change should happen in the way companies approve productivity targets. No manager should be allowed to increase a teamโs expected output merely because an AI tool has produced a faster result in a demonstration.
A demonstration proves that a tool can perform a task under certain conditions. It does not prove that the whole team can repeat the result safely, every day, across different projects.
Before changing targets, run a measured trial with real work. Record the time saved, the extra checking required, the errors found and the effect on the teamโs other responsibilities. Only then should the company decide whether the workload can change. This one rule would prevent many false promises from reaching employees.
A Task Is Not Finished When AI Stops Generating Code
One of the most important comments in the discussion came from an engineer who challenged the claim that a feature could be built in two afternoons. The engineer explained that such estimates often come from watching someone create a quick demonstration.
They do not include review, integration, repairs or the person who may have to deal with a failure after release. This is a basic accounting problem. The company counts the visible part of the work and leaves the rest outside the calculation. Then it asks employees why the actual delivery is taking longer.
A task that has been generated but not checked is not a finished task. It is unfinished work waiting for someone to accept responsibility.
Consider a feature that AI produces in three hours. A developer spends another two hours reading the code, a reviewer spends three hours checking it, and a tester finds a problem that takes four hours to repair. The feature then needs another round of testing and approval.
Calling this a three-hour task is not an innocent mistake. It gives management a false picture of the work and makes the people doing the checking appear slow. Companies should measure the complete journey from request to accepted release.
Every time a change is returned for repair, that time should be recorded against the original task. Review hours should not be hidden under general team activity. They belong to the cost of the work.
There is a useful lesson here about honest measurement. A shopkeeper cannot claim to have given a customer the correct amount while measuring only half the goods. In the same way, a company cannot claim that a task took three hours when it has left ten more hours outside the calculation.
The practical answer is to create a delivery record for every important piece of work. It should show the original request, generation time, human review time, testing time, repair time and final acceptance. After several weeks, managers will have real information about their teamโs delivery capacity. They can use that information to plan future work instead of repeating estimates that came from the fastest person in the room.
We Are Creating More Code Than We Can Understand
A manager in the discussion raised the problem of knowledge debt. AI was building features without full code review and making changes to existing tests without proper attention.
This deserves more concern than the number of lines being generated. A software system is not just a collection of files. It carries years of decisions, customer needs, old failures and rules that may not be obvious to someone seeing the code for the first time.
A change can look correct in one place and create trouble somewhere else. The danger is that the person reviewing the work may not even know what question to ask. The code may pass its tests while the company slowly loses its understanding of how the system behaves.
The solution should begin with a rule about knowledge. Every important part of a system must have a named person who understands its purpose, its risks and the reason behind major design choices.
This person does not have to write every line of code. The responsibility is to make sure that important changes do not enter the system without someone understanding their effect. When AI changes an existing test, the reviewer should ask what behaviour the test was meant to protect.
If the test now passes because the requirement has been weakened, the company has not improved its software. It has simply removed the warning light. Test changes should therefore receive their own review, especially when they affect security, payments, access or customer data.
There should also be a limit on how much new code a team accepts before understanding what it already has. For example, a company could require a review of a featureโs design, tests and ownership before allowing another major feature to depend on it. If the team cannot explain how the current feature works, it should not keep adding more work on top of it.
This may appear slower in the first few weeks. But the purpose is to prevent a much larger problem later, when a system has grown beyond the teamโs ability to understand and repair it. A company should never confuse the amount of software it owns with the amount of software it can safely maintain.
The Real Shortage Is Human Attention
One engineer said that managing two AI-assisted projects had become unbearable within three months. Another described the pressure of running several projects at the same time, checking the output and making decisions throughout the day.
These comments point to a problem that is easy to miss. AI can work on several tasks at once, but the person supervising those tasks still has one mind. Every unfinished branch, open question and pending review asks for attention. A person may not be typing much code, yet the mind is constantly moving between different problems.
That movement has a cost. It also makes mistakes more likely because people lose the full picture of the task they are trying to complete.
Companies should stop treating the number of active projects as a sign of ambition. A team with seven unfinished projects may be doing less useful work than a team with two completed projects. The first needs to remember seven sets of decisions, risks and deadlines.
The second can give its attention to finishing the work properly. Set a limit on active projects for every engineer. If a new urgent request arrives, the manager must decide which existing task will be paused. Do not add the new task quietly and expect the employee to carry the extra burden. This is a management decision, not a personal test of how much pressure someone can tolerate.
There is another useful measure that companies should introduce: the number of unfinished tasks waiting for human attention. Count the AI-generated branches that need review, the decisions awaiting approval and the projects waiting for integration.
If this number keeps rising, the team is creating work faster than it can finish it. The company should slow the arrival of new tasks until the pending work comes down. This is better than asking employees to review faster because the actual problem may be too much work entering the system. A person cannot solve a traffic jam by driving every car faster. Sometimes the road simply cannot carry that many vehicles at once.
We Are Losing the Knowledge That Makes Experienced People Valuable
Another comment described how experienced engineers were no longer being seen as essential because management believed that someone using AI could do the same work. This is a serious mistake.
Four years inside a company can give a person knowledge that does not appear in the source code. They may know why a customer was given a special arrangement, why an old service cannot be changed easily, which supplier has unreliable data or what happened during a failure several years ago.
Much of this knowledge is gained through experience, conversations and difficult decisions. It is not always written down. A new employee with an AI tool may produce code quickly, but that does not mean the person understands the business in the same way.
Companies should change how they reward experienced engineers. Their value should not depend only on the amount of code they personally write. They should also be recognised for preventing mistakes, teaching others, explaining difficult systems and helping teams make better choices.
Give senior engineers a formal role in reviewing major changes and guiding less experienced workers. Pay them for this work rather than treating it as an extra duty added to their existing workload. Also create time for them to record important knowledge.
If an expert spends every day meeting deadlines, there will never be enough time to pass that knowledge to the next person. The company then becomes dependent on individuals while pretending that AI has made everyone replaceable.
Junior employees need protection too. If AI writes every difficult function and answers every technical question, a young engineer may complete many tasks without learning how to solve the underlying problems. The company may see quick results today and discover a skill gap two years later.
A good training system should give junior employees a mix of AI-assisted work and tasks that require them to understand the problem themselves. They should explain why a solution works, test different approaches and learn to find errors without depending on a model.
Senior staff should review their reasoning, not just their final code. The goal is not to keep people away from AI. It is to make sure that using AI does not prevent them from becoming skilled engineers.
The Person Who Approves the Work Must Have the Power to Reject It
There is a quiet unfairness in many AI workflows. An engineer may be asked to approve code that has been produced at a speed the engineer cannot properly review. The manager wants the work released, the deadline is already announced and the employee is expected to sign off.
If something fails later, that same employee may be asked why the problem was not found. This arrangement gives responsibility to one person while keeping the decision-making power somewhere else. It is not fair, and it is not a safe way to run a technical business.
Responsibility means little if the person carrying it cannot ask for more time, reject a change or demand better evidence.
Every company should have a clear approval rule. The person responsible for a release must be allowed to stop it when important checks are missing. A manager may challenge the decision, but cannot simply force an approval without recording the risk and accepting responsibility for that choice.
For high-risk systems, approval should depend on evidence such as test results, security checks, review records and a clear plan to reverse the change if it fails. The same rule should apply whether the code was written by a junior employee, a senior engineer or an AI agent. The method of creation does not remove the duty to check the result.
This also changes how mistakes should be handled. When an AI-generated feature fails, the company should not immediately look for one employee to blame. It should examine who requested the task, what instructions were given, which checks were performed, who approved the change and whether the company gave the team enough time.
If a person ignored a known risk, that is one kind of failure. If the process made proper review impossible, that is a different failure. Mixing the two prevents companies from learning. A fair review must follow the chain of decisions, not simply find the nearest person whose name appears on the release.
Cheap Creation Can Become Expensive Ownership
AI has reduced the effort needed to create many things. A feature that once required a team can sometimes be built by one person with an AI tool. But the cost of creating something is not the same as the cost of owning it. Every feature brings future work.
Someone must update it, answer questions about it, check it when other parts of the system change and repair it when customers find a problem. A company can create hundreds of features in a year and quietly build a maintenance burden that its employees will have to carry for years. The first team may receive praise for fast delivery.
A later team may spend its time looking after code nobody remembers approving.
Before a new feature is accepted, the company should record who will maintain it and what work may be needed later. It should also decide what would make the feature worth keeping.
If customers do not use it, if it does not solve the problem or if its maintenance cost becomes too high, the company should be willing to remove it. Software removal should be treated as normal engineering work, not as an admission of failure. The same attention given to adding features should be given to retiring them.
This prevents the product from becoming a collection of old promises that nobody has the time or courage to remove.
A useful financial rule can make this change real. Every new feature should carry an estimate of its future support and maintenance cost. The estimate does not need to be perfect. It needs to make the future cost visible before the company approves the work.
If a team has room to build ten features but can maintain only six, management must choose carefully. The remaining four are not free just because AI can create them quickly. They will still need attention, and that attention will come from people who already have other work to do.
More Features Do Not Mean More Value for Customers
One comment in the discussion pointed to unclear customer needs, review delays and unreliable tests. These are old problems, but AI can make them larger because it allows teams to produce more work before asking whether the work was needed.
A company may build a feature in a day, but customers may never use it. The developers have spent time on it, testers have checked it and support staff may have to explain it. The company has created activity without creating much value.
The ability to build quickly can make people less careful about deciding what deserves to be built in the first place.
A simple change in the approval process can help. Before a feature enters development, the person requesting it should explain the customer problem in plain language.
The team should identify how customers experience that problem today and what would improve if the feature worked. There should be a small test before a large build whenever possible. After release, the company should check whether customers actually used the feature and whether the original problem became smaller. If the expected result does not happen, the team should not keep adding more work just to defend the original decision. It should change the idea or stop it.
This approach also protects engineers from being used as a service desk for every new management idea. A request should not become a project simply because someone senior has suggested it. The person asking for the work must take responsibility for explaining its purpose and accepting the result. If the request has no clear customer or business reason, it should wait until that reason is known.
This creates a useful boundary between having an idea and asking other people to spend their time on it. It also gives managers a better way to say no without turning every disagreement into a personal argument.
The Benefits of AI Should Not Belong to the Company Alone
One person in the discussion said that the workload had increased while pay remained the same. This raises a question that companies cannot avoid. If employees learn new tools, change their working habits and help the business produce more, why should all the gain go in one direction? The company may earn more, serve more customers or reduce the cost of delivery. The employees may receive more tasks, more responsibility and less time to recover.
That is not a fair exchange. It also creates a reason for skilled people to leave as soon as they find a workplace that values their contribution.
Companies should decide in advance how productivity gains will be shared. The benefit does not have to take only the form of a salary increase. It may include shorter working hours, paid learning time, better leave, a bonus linked to verified gains or more freedom to choose how work is done. But there should be a clear connection between the improvement employees help create and the benefit they receive.
A company should not tell people that AI is making their work easier while using the same claim to increase their targets and reduce the number of people doing the work. If the tool creates a real gain, the people who helped produce that gain should have a fair share in it.
There is also a need for honest discussion about job changes. If a company plans to reduce roles because AI can handle some tasks, it should not hide the decision behind vague claims about efficiency. People deserve to know what is changing, what skills will be needed and what support will be available. Where possible, employees should be trained for new roles before outside hiring begins.
A person who has spent years learning the companyโs systems may be able to move into testing, system design, customer problem-solving or AI process review. The company should examine these options before treating the person as an old cost that can be removed.
A Workplace Needs Rules for the Use of AI, Not Just Access to AI
One manager in the discussion suggested that companies should stop making developers prompt and manage several AI agents by hand. Instead, they should build systems that handle the routine parts of the process. This is a useful idea because the current approach often adds another layer of work. An engineer asks an agent to make a change, checks the result, sends it back, opens another agent and then tries to remember what happened in the first branch.
The tool may be doing the coding, but the person is doing the coordination. Companies should not assume that this work is free simply because it happens through a chat window.
The better approach is to build a shared process around the tools. Every agent should receive the correct project information, follow the same coding rules and work within clear limits. Routine tests should run automatically. Changes should be grouped and presented in a way that helps a reviewer understand what was done and why.
The system should show which files changed, which tests passed, which tests failed and what still needs human attention. An agent should not be allowed to silently change important tests or approve its own work. If the task is outside its permission, it should stop and ask for a human decision. This reduces the number of small decisions that employees must make throughout the day.
The company should also decide which work should not be given to an agent. A task involving sensitive customer information, a major security change or an unclear business rule may need a person from the beginning. Not every task becomes better when AI is added.
The team should test where the tool helps and where it creates extra work. This is not resistance to technology. It is the same care used when choosing any tool that can affect a business. A hammer is useful, but nobody uses it to repair every object in a house. The value of a tool depends on knowing when it belongs in the job.
Rest Should Not Depend on How Close Someone Is to Breaking
The original post mentioned people who wanted to be laid off because they needed time away from work. That is a painful thing to read. A person should not have to lose a job to get a chance to recover.
Companies often wait until someone resigns, becomes unwell or stops performing before they take workload concerns seriously. By then, the damage may have been building for months. A better system would notice the signs earlier. Managers should look at repeated late work, growing review queues, missed leave, constant task switching and employees who never seem to finish their assigned work.
These are not proof of burnout by themselves, but they are reasons to have a real conversation.
A company should create a way for employees to report overload without putting their reputation at risk. If someone says that the workload cannot be handled safely, the manager should review the actual tasks and remove or delay some of them.
The answer should not be another motivational speech. Teams also need protected periods without meetings and a real right to disconnect after working hours. If an urgent incident requires someone to work late, the company should arrange recovery time afterwards. This is not special treatment. It is a way to keep people able to do careful work over many years instead of using them until they cannot continue.
There is a deeper point here about the limits of human beings. A person can carry a difficult responsibility, but that does not mean every responsibility should be placed on that person.
Good management must consider what someone is already carrying before adding more. The question is not simply whether an employee can complete the next task. The question is what will be neglected, delayed or damaged if the task is added. That is how a company should judge capacity. It should not wait for someone to fail before accepting that the workload was too heavy.
The Rules of Management Must Change Along With the Tools
The research gives us a reason to avoid simple claims about AI. A 2025 study involving 4,867 developers across Microsoft, Accenture and a large company found an average increase of 26.08% in completed tasks among developers using AI coding assistants.
That is a real result, but it does not mean every developer or every kind of work will see the same gain. A separate METR experiment involving experienced open-source developers found that the tasks took 19% longer when AI tools were allowed in that particular setting. Later METR updates also warned that changes in tools, participants and measurement made newer estimates difficult to interpret.
These findings should not be forced into one simple story. They show why companies need to measure their own work rather than copy a number from another setting.
The 2025 DORA research found that AI use was widespread and that many developers felt more productive.
It also found a gap between the value people saw in AI and the level of trust they placed in its output. That matters because a company may see more work being released while still facing questions about whether the work is correct and safe.
A 2025 workplace study also found that developers enjoyed using coding assistants, but their trust in the generated code did not rise in the same way. Enjoying a tool and trusting every result from it are two different things. Management must leave room for both facts. Employees can welcome AI and still need more time to check its work.
The new management rule should be based on evidence from the companyโs own teams. Each department should run a small trial before changing its targets. The trial should compare work with and without AI on similar tasks. It should record time to completion, review effort, defects, customer results and employee workload.
The company should also check whether the tool is helping junior and senior workers in different ways. At the end, managers should decide what to keep, what to change and what not to use. This is better than announcing that every team must become faster because another company has reported a productivity gain. The right to use a tool should not become a duty to produce an untested amount of extra work.
We Need to Stop Confusing Human Worth With the Amount of Work Produced
There is a belief behind much of this pressure: if a person can do more, that person should always do more. This belief has shaped workplaces for years, but AI is making its effects harder to ignore.
When a machine can produce work at a lower cost, the temptation is to keep raising the target until the human part becomes the only limit. But a personโs worth cannot be measured by the number of tasks completed in a week. Some of the most valuable work is hard to count. It may be a warning that prevents a failure, a conversation that stops a poor decision or a lesson that saves another employee months of confusion.
Companies should make room for this work in performance reviews. An engineer who prevents a serious problem should not be judged as less productive because the problem never happened. A senior employee who teaches five younger workers may have written less code that month, but the company may have gained five people who can work with greater skill.
A person who removes an unnecessary feature may save the company years of maintenance. These contributions should be recorded and valued. The company should ask what became safer, clearer or easier because of the employeeโs work, not only what new item appeared on a project board.
I also believe employees need a voice in deciding what productivity means in their own teams. The people doing the work often know where time is wasted, which checks are useful and which reports exist only to satisfy management.
They should be able to challenge measures that reward the wrong behaviour. If a metric encourages people to produce more code while ignoring defects, the metric should be changed.
If a deadline causes people to skip important reviews, the deadline should be questioned. A number is useful only when it helps the company see the truth. When it hides the truth, it becomes another problem that employees must work around.
The Future of Work Depends on What We Choose to Reward
The biggest change does not require another AI tool. It requires a different idea of success. A company should reward work that solves real problems, remains reliable, can be maintained and does not depend on exhausting the people who built it. It should reward the person who says a release is not ready when the evidence supports that decision.
It should reward the manager who removes unnecessary work, even if the team appears less busy for a while. It should reward the engineer who teaches others and the employee who finds a simpler way to solve a customer problem. These choices may not create impressive numbers every week, but they build a company that can keep doing good work.
There is a practical way to begin. Each team should review its work every month and ask what it created, what it had to repair, what it stopped doing and what knowledge it gained. Managers should compare planned hours with actual hours, including review and maintenance.
Employees should be asked whether they have enough time to do careful work and whether the tools are reducing or adding to their daily burden. If the numbers show that work is arriving faster than it can be checked, the company should reduce the amount entering the system.
If the team is spending too much time managing agents, the process should be redesigned. If features are not being used, the company should stop making more of them.
A company does not become weak when it refuses to overload its employees. It becomes weak when it depends on people who are too tired to notice mistakes, too afraid to question deadlines and too busy to pass on what they know.
AI may change the cost of writing code, but it cannot remove the need for good judgement, clear responsibility and trust. The company that understands this will have a better chance of keeping its skilled people and serving its customers well.
The company that ignores it may enjoy a short period of higher output while slowly losing the knowledge that made its work valuable.
I want to see a workplace where a person can use AI without feeling that every saved minute will be taken away. Where an experienced engineer is valued for what years of work have taught them.
Where a junior employee can learn instead of being judged only by speed. Where a manager can say no to an unrealistic deadline without being treated as an obstacle. And where a person can finish the working day without feeling that the real work has only just begun.
The point of creating better machines should not be to make human beings feel less useful. It should be to remove the work that does not need human effort and give people more room for the work that does.
If AI can write the first version, let people spend more time asking whether it should be written at all. If AI can test routine cases, let engineers study the difficult ones. If AI can reduce repeated work, let the saved time go towards learning, better decisions and a life outside the office.
A person is not a machine with a slower processor. A person has a family, a body, a mind, a need to learn and a limit to what can be carried at one time. Any company that forgets this will eventually pay for it, even if the cost does not appear in the next quarterly report.
The real test of AI will not be how many tasks a machine can complete. It will be whether the people working with it have gained a better way to live and work. That is the result worth measuring. That is the progress worth paying for.


Leave a Reply