Originally published in my newsletter on sustainable pace. Subscribe for free.
This essay has been more than a month in the making, and it took me around 30 hours to write. I hope you find it insightful and useful. I would also appreciate you sharing it with those who might benefit from it.
Enjoy!
Seoul, 2012
I often get asked if my book on sustainable pace in knowledge work (Kestävä työtahti, published in Finnish in August 2026) covers the use of AI and its implications. My answer is that AI is referred to in the book, and in some cases used as an example, but its main focus is to present a comprehensive research-backed theory of effectiveness for modern-day knowledge work. And if that theory turns obsolete as the use of AI becomes more commonplace, it was not a very good theory to begin with.
Nonetheless, the question of what sustainable pace in knowledge work looks like in the age of AI is very much relevant. According to Microsoft's 2025 Work Trend Index, a survey of 31,000 knowledge workers in 31 markets, 53% of leaders say that productivity must increase. At the same time, 80% of the workforce, leaders included, say they don't have enough time or energy to do their work. Unsurprisingly, Microsoft's answer to this "capacity gap" is AI. I can see the appeal. It's tough to squeeze more out of people who are already over their limits.
Whether or not the expected productivity improvements will be realised remains to be seen, but it’s not as straightforward as the big tech companies and AI labs would have you believe. One glaring issue is that measuring productivity in knowledge work is notoriously difficult, and it doesn’t take much to spin the data to tell whichever story best fits what you want to see. For most of us, a more practical question is how AI can help us do work that matters without wearing us out in the process.
Sustainable pace, in the way I talk and write about it, means doing meaningful work that creates value, at a pace you can keep up year after year without exhaustion and burnout. The theory presented in my book is based on unchanging facts about how our brains work, and the unique characteristics that make knowledge work different from industrial labour. To name a few: knowledge work is intangible, so you can't see at a glance how much of it is piling up. There are always more things that could be done than there is time to do them. The results are often non-linear, which means that a small change in the right place can matter more than months of effort somewhere else. And knowledge work is still the responsibility of people — not machines — with limited attention, working memory and energy, in a world where things rarely go according to plan.
AI doesn't change any of this. However capable the tools become, the people using them can still direct their conscious attention to only one thing at a time, and they still run out of energy and need to recover. That's why I'd argue that the theory of Sustainable Pace in Knowledge Work also holds up in the age of AI. The more interesting question is what the theory tells us about where and how to use AI. This essay sets out to explore it.
A core part of the theory is a framework of six principles. They are meant to be applied together, because they reinforce one another to make knowledge work effective, impactful and sustainable in the long run. Limiting work-in-progress, for example, can help you maintain a sense of control and avoid burnout, but it won’t make your work any more meaningful if you have no idea what creates value and impact. The principles are:
Focus on impact and value. Spend your limited time and energy on what creates value, make necessary work easier, and avoid waste.
Embrace uncertainty. Surprises are the only thing you can count on, so leave slack in your plans to deal with them.
Limit work-in-progress. Everything unfinished generates overhead, so the more things you have going on at once, the less time you have for the work itself.
Make work visible. You can't plan or prioritise work you can't see, and the more difficult it is to access the information you need, the more you end up wasting your time and energy.
Reduce context-switching. Every shift of attention from one task to another costs time and energy.
Manage your energy. The value of your work is rarely directly proportional to how much time you spend on it. This is especially true in decision-making, problem-solving and creative work. Therefore your physical and mental state has a direct impact on effectiveness.
The principles work in two directions. They give you a way to test a particular use of AI. For example, does it remove waste or automate it? Does it free your attention or fragment it? The principles also describe the conditions AI needs in order to pay off, such as slack in your schedule, visibility into tasks, and well-organised information. I believe that AI can help us knowledge workers get more meaningful work done, but it can also harm our effectiveness and wellbeing. Which of the two happens depends less on AI itself than on how and where we use it.
For the purposes of this essay, I've grouped principles 3 and 5 together as they're closely linked. The perspectives vary from individuals to teams to organisations. Leadership behaviour and management practices are also discussed.
Helsinki, 2024
Focus on impact and value
For me this is the most relevant of all the six principles when it comes to using AI. People have limited time and energy, so we should spend it on what matters most. Yet the organisational reality in many workplaces is that a large share of the workday is spent on things that don’t create value or impact: multitasking wastes time; disorganised information requires sorting and filtering before it can be used to make progress in actual tasks; unclear roles, responsibilities and lack of feedback lead to second-guessing and possibly doing the wrong thing in the first place. In order to preserve cognitive capacity for work that creates value, we should aim to minimise all forms of extraneous cognitive load (a term from cognitive load theory for mental effort caused by how the work is organised and presented, rather than by the contents of the actual tasks).
This is where AI can help. For example, it can become the interface for finding and sorting relevant information from a stack of sources and documents, each with its own formatting and structure. We then have more energy left for figuring out what to do with that information. A good example of this is a forensic architect interviewed by Dan Heath, who said that sometimes he gets sent thousands of files with the file names obfuscated on purpose so he can't tell what's in which file. In addition, sometimes textual information is converted into images so that text search can't be used. It used to be costly and tedious for him and his colleagues to try to find the information they needed. AI has changed that.
The same goes for optimisation and automation. There is always work that simply has to be done, but where more effort does not create more value. Think maintenance work, required reporting and administrative overhead. Using AI to simplify or automate these tasks frees up time for more value-adding activities.
For example, a significant amount of time could be saved in large organisations by adopting an AI tool for automated generation, and even approval within certain limits, of travel expense reports. The process where employees manually upload, categorise, and allocate each receipt is prone to errors. It then takes time for another person to review and check that everything is correct before the expenses can be approved. What if you could simply upload the receipts to a chat interface and explain in natural language what the expenses are about? An AI would extract relevant information from the receipts and generate the expense report for you. There are numerous administrative processes like this where AI has the potential to free up people's time and energy for more meaningful work.
One thing to be aware of is that as AI makes it easier to automate tasks, the likelihood of automating something that should be eliminated altogether also becomes greater. If a report is not read by anyone in the first place, automating its generation is also a waste of time and energy.
This is closely related to the fact that AI makes it nearly free to produce new work that nobody asked for. When a report, a slide deck or a summary takes seconds to generate, more of them get generated. It's not uncommon that AI-generated work looks polished but on closer inspection lacks substance. Problems arise when the job of evaluating AI outputs shifts away from the person generating them. Someone has to spend their time and attention doing it. These days you often see people vent their frustration on LinkedIn: "If I wanted the opinion of an AI, I would have asked the AI and not you!"
One useful rule of thumb: before you throw AI at a problem, ask whether you're dealing with a symptom or a root cause. The aforementioned forensic architect has no way of fixing the root cause of his problem. The files are made hard to search on purpose, by people who would rather he didn't find what's in them. For him, AI is the right tool. Most of the disorganised information, unclear goals and constant interruptions in an ordinary workplace are different. They're a by-product of the way the work is organised, and that should be addressed directly. The risk is that if, for example, AI is used as a sticking plaster to help employees think through unclear responsibilities with good-enough results, nobody is going to fix the underlying problem. Clarifying goals and responsibilities is a job for management, and AI should not become an excuse for doing it poorly.
The hardest question this principle raises is about the work that creates the value in the first place. Writing, analysis, design and problem-solving are the kinds of work AI keeps getting better at, which makes it tempting to hand them over. But the value of expert work often lies in the judgement that goes into it, and that judgement is built by doing the work and thinking through problems on your own. If AI does most of the thinking and you only check the results, the ability that makes your checking worth anything starts to fade. I wrote about this in more detail in an earlier newsletter article. So for each part of your core work, ask yourself: where does its value come from? How much of it should be handed to AI? And what way of working together with AI could help preserve or even build your own expertise, instead of letting AI slowly take its place?
When approached in a thoughtful manner, AI can help us create more space for work that creates real value and impact. It can free capacity and energy from activities that are either wasteful, could be made less burdensome, or even automated fully. Some of this can be put into practice at an individual level, and some requires bigger changes in the tools and processes of the workplace.
There is a catch, though. The space AI frees up is not going to stay empty. Getting more done faster will not exhaust the amount of work there is to be done. New demands, tasks, projects and ideas will fill that void. Oliver Burkeman calls this the efficiency trap: the faster you get the important work done, the busier you will feel about less important things. It is a trap that will eventually leave you exhausted and demoralised. The pursuit of efficiency becomes a never-ending quest. Whatever you do is never enough. There is always more you could have done.
The first step in escaping the trap is to recognise that it exists. The problem is that AI can make it more difficult to break free. One more draft, one more variant or one more idea to try out now costs almost nothing. When Aruna Ranganathan and Xingqi Maggie Ye from UC Berkeley spent eight months studying a 200-person US tech company, they found that AI didn’t reduce people’s workload. It intensified it. People worked at a faster pace and took on a broader range of tasks, often outside their previous roles.
If you want AI to make room for meaningful work, you can't just go and fill the space with all the things that have now become possible to do with AI. That is not sustainable, as the people in the UC Berkeley study started to find out. I'd say you have to become even more discerning and protective of your time. Otherwise it just fills up with more of everything.
Embrace uncertainty
This principle is about making sure you have capacity to react to surprises. AI can help you see some of them coming. An AI assistant could keep an eye on your projects and flag early signs of trouble, such as a dependency that is running late or a client whose messages have started to sound worried.
On the other hand, getting too reliant on AI-generated predictions instead of focusing on maintaining the capacity to react to the unexpected makes you more vulnerable to black swan events (also known as "unknown unknowns"). An improved predictive capability can make you overconfident, but it does not make you immune to surprises. And when the nasty surprise happens, it comes with a bigger negative impact if you're already stretched too thin.
A more robust approach to the uncertainty that is inherent in knowledge work is to focus on adaptiveness instead of trying to become better at predicting things. A system that has no slack is brittle. It has no room for dealing with the unexpected. The extra strain that the unexpected causes means that something else has to give.
AI can also add uncertainty of its own. Sometimes a simple task you expect to take a few minutes runs into problems, and you end up spending the next hour trying to sort it out. I remember being frustrated last spring at the utter inability of Copilot to transcribe text from images. Especially because I knew that ChatGPT had no problems with the same task, and they were supposed to be running the same language model. I was able to figure out after about half an hour of arguing with the AI that Microsoft's tooling that sat between me and the language model downscaled, without any indication of doing so, every image file to such a low resolution that all the text turned to gibberish. Instead of telling me this, Copilot just kept on hallucinating its outputs. Only after being pressed multiple times did it confess that the files were unclear, whereas at my end they were not. It still took a few rounds of discussion to figure out what exactly was causing the problem with the files.
If you plan your week on the assumption that AI makes everything faster, every hiccup like that is time away from something else. Not to mention what frustration does to your ability to think clearly. It's the same optimism that has always led to tight schedules and underestimating how long it takes to get stuff done. When you plan to do something that relies on AI, leave room for it to not work as expected.
Having slack may seem less productive than the full utilisation of resources, but the loss in productivity often pays for itself when surprises happen. If the surprise is a negative one, slack lets you deal with it without other work suffering. When it is a positive surprise, slack often means you can pursue an unexpected opportunity.
Therefore, assuming that AI frees up capacity, some of it should be kept in reserve for dealing with surprises. As I argued above, that won't happen by itself, because capacity that nobody protects has a tendency to get filled with other work. In practice this means deliberately limiting the amount of planned work to less than full capacity. For example, the Finnish author Sari Torkkola writes in her book on using lean principles to lead expert work (Lean asiantuntijatyön johtamisessa) that the ratio between incoming work and available capacity should not exceed 80%. Otherwise lead times get longer, quality goes down, the amount of work in progress increases (which causes its own problems, as we'll see later), people start to feel rushed and stressed, and there will be no time to learn new things or do continuous improvement. Applied more broadly to knowledge work, the rule suggests planning no more than four days of work a week and keeping the fifth in reserve for dealing with surprises.
Protecting slack is hard for an individual if the organisation keeps filling it, so this is as much a decision for managers as it is for you. Unfortunately it seems that hardly a week goes by without news of some large organisation laying people off with AI as the reason. Whether the cuts follow real productivity gains or only the expectation of them, those who remain end up with no more slack than before, and with less if the expected gains never materialise. AI may be the real reason or just a convenient explanation, but either way the message people take from such news is that the productivity gains from AI will cost them their jobs. That’s a surefire way to make people hostile towards AI adoption.
Lean manufacturing went through the same thing. In many organisations that adopted it from the 1990s onwards, the first improvement initiatives freed up capacity, and layoffs followed. People drew the obvious conclusion: using new tools, processes, and ways of thinking to improve how you work puts your own job at risk. It's one reason experienced lean practitioners insist that nobody should lose their job because of an improvement initiative. Once people link improvement to layoffs, continuous improvement grinds to a halt. What looks good in quarterly financial metrics makes the organisation more stagnant and vulnerable in the long term.
A better alternative would be to use the freed-up capacity to develop new capabilities. Research on organisational ambidexterity builds on a distinction James March made in 1991 between exploitation and exploration in organisational learning. Exploitation is about making the most of existing opportunities, executing everyday operations effectively and continuously improving in those areas. Exploration is about the skills, capabilities and structures that let an organisation pursue and develop entirely new opportunities, to be exploited in the future. In order to survive, to renew themselves, and to stay relevant in the long term, businesses need to be able to simultaneously pursue both incremental innovation (exploitation) and discontinuous innovation (exploration).
We are still in the early stages of AI adoption in most organisations. There's no playbook, and the future of human-AI collaboration is being created as we speak. This is exactly the kind of situation that calls for exploration and experimentation, but in most organisations people are too busy dealing with everyday work to spend their time innovating. In Microsoft's 2026 Work Trend Index, 45% of AI users said it feels safer to focus on their current goals than to redesign their work with AI. In other words, to exploit instead of explore.
It feels inherently uncomfortable to put a pause on what you're doing — especially if you're already behind schedule — and carve out space to try out something different. To experiment with new tools, methods, and ways of doing things with AI. The benefits of exploration come with a delay. When there are fires burning and people have no slack, the work that would help an organisation get more value out of AI in the future gets postponed, and then postponed again.
The risk is that one day the leadership team realises their business is being outcompeted, outpriced and out-innovated by new entrants who’ve built everything from the ground up with AI in mind. Or they end up putting their faith wholesale in what external vendors and consultants are promising. And as I've written before, outsiders can never understand the reality of work in your organisation as well as the people who are doing it day-to-day.
Nobody knows yet how large AI's benefits will turn out to be, and all the different ways in which it can create new opportunities for both individuals and organisations. That uncertainty is the strongest reason to put time and effort into exploration. Small, inexpensive experiments are how an organisation finds out what AI can mean in its specific context, where the biggest opportunities are and how it should go about pursuing them.
If the current situation is that 80% of the workforce doesn't have enough time or energy to do their work, as Microsoft reported in 2025, AI should be used to bring that number down. If the capacity that AI frees up leads to layoffs, the people who remain are back at full utilisation, with no slack for surprises and no time to explore what AI could do next. Keeping part of the gain in reserve, and investing some of it in exploration, gives the organisation a chance to keep benefiting from AI as things develop and we discover new and better ways to use it.
Limit work-in-progress and reduce context-switching
Limiting work-in-progress and reducing context-switching reinforce each other. The more things you have going on at the same time, the more likely it is that you try to work on them simultaneously, shifting your focus from one task to another multiple times before getting any of them done. It may feel like you're on top of things, but in practice every shift in attention between contexts causes delays that could be avoided by finishing one thing before moving on to the next one.
This is because our working memory has a limited capacity to hold information. We can also direct our conscious attention to only one thing at a time. Each shift means having to reorient ourselves, and to bring back to working memory the information that is relevant to whatever is now in focus. The result is that things stay unfinished for longer, and as new work keeps arriving, the pile of things in progress grows.
I'd say this is where the sustainable pace framework raises the biggest concerns about how we use AI at an individual level.
In summer 2025 I decided to see if I could use Claude to develop an iPhone app. I had some programming experience in Swift and a passing familiarity with Xcode. I described what I wanted the app to do and created the graphic design for it, while Claude wrote the code, which I then put into Xcode. The whole process was iterative. I would explain the features I wanted to add, run tests, and report back to Claude what didn't work or needed changing. It didn't take long to have a working prototype. Getting all the kinks ironed out took longer, but at no point did it seem impossible. Eventually I published the app and it was available in the App Store for a year. It works flawlessly, and I still use it myself. The reason you can't get it any more is that I didn't want to continue paying the annual developer licence fee to Apple. It was, after all, just a hobby project.
Whenever Claude was busy writing code, I was doing other things. A few minutes later, I'd get a notification that Claude was done and was now waiting for me. It felt really productive to make progress on multiple things, but at the same time I was starting to lose my own ability to stay focused. A few months later I had an even stronger experience of this, when I was running multiple AI tasks simultaneously. It felt like I was getting far more things done than normal, but I also became stressed and distracted. It was like having multiple colleagues, each of whom wanted something from me every few minutes. Doing any other work, especially work that required focused attention, became impossible. And because there was no way to tell when Claude would be done and needed my input again, I often ended up just sitting there and doing nothing. Until the notification bell rang and, like Pavlov's dog, I would perk up and go and see what treat awaited. When working with AI, it felt pointless to even think about doing things that would require my full attention.
My experience is not unique. In the UC Berkeley study mentioned earlier, people ran AI processes in the background while reviewing code, drafting documents or attending meetings, and some ran several AI agents at once. While prompting and iterating, they talked about momentum and a sense of expanded capability, but when they took a pause to look at their work more broadly, they described feeling busier and more stretched. Short-term productivity gains won't get you far if they cannot be sustained in the long term.
To be fair, AI can make each switch between contexts less costly. Much of the effort of returning to a task — especially a complicated one — goes into bringing the details back into working memory: where you left off and what still needs doing. AI is good at reconstructing them, whether by summarising a long thread or reminding you what you were working on before the interruption. That lowers the cost of each switch, but there is still a cost. It also does nothing to reduce how often you switch.
To make matters worse, modern work is already built for interruption. According to Microsoft's own data from 2025, people working in Microsoft 365 are interrupted every two minutes by a meeting, an email or a chat message. That's 275 interruptions a day. When an alert can arrive at any moment, checking easily turns into a habit. "Ding!" Maybe it's something important. Better check.
It’s the same principle of unpredictable rewards that makes slot machines and other forms of gambling so addictive. Your mind becomes conditioned to them and starts craving interruptions and distractions. Over time, we start to lose the ability to be still and stay with one thing for extended periods of time. In the absence of external interruptions, the mind interrupts itself with distracting thoughts, ideas and impulses. The ability to maintain focused attention is like a muscle. It atrophies if we don't use it.
There's a real danger that AI becomes one more source of constant interruption, and a harder one to resist. Nobody feels guilty about muting social media during working hours. Muting the notification that tells you an agent has finished a task is another matter, because it can feel like neglecting your work. Especially if the organisation you work for is pushing hard for everyone to use more AI.
The implication here is that if we get accustomed to using AI for many different tasks simultaneously, and our work gets reduced to checking outputs and giving the next set of instructions, we start losing the ability to just be still, to think and reflect deeply, and to sit with a problem or a piece of work for hours instead of minutes. On top of that, as we become more dependent on AI doing the work for us, we may even start to become worse at checking and verifying what the AI does, as I mentioned earlier.
Besides increasing the amount of context-switching, having a lot of work in progress also means we end up wasting more time on overhead work that could otherwise be avoided. As an example, every project usually comes with its own status meetings, messages and coordination for as long as it stays unfinished. If each of five parallel projects has a one-hour weekly status meeting, that’s five hours a week spent talking about the work instead of doing it. And because each project now gets fewer hours a week, it takes more weeks to finish them. The result is even more status meetings per project.
AI can be used to reduce this "overhead tax", as Cal Newport calls it in Slow Productivity, but it would still be more effective, less wasteful, and also feel more effortless to do fewer things at the same time — with or without the help of AI — and to get them done faster from start to finish.
What AI does not reduce is the part of the work that requires human attention. It can even exacerbate the problem. Running multiple AI agents in parallel means having more things in progress, and each one eventually comes back to you for review, feedback or a decision. Once the agents produce work faster than you can review it, adding more of them only lengthens the list of things that you need to pay attention to. You become the bottleneck, and as illustrated by Eliyahu M. Goldratt in his book The Goal about the theory of constraints, the bottleneck sets the pace for the whole system.
On the one hand, AI is pulling us towards becoming more distracted and working on multiple things at the same time. But the bigger the piece of work that an AI produces, the more it takes time and focused attention to properly check it. So on the other, the quality of the work we do with AI would probably benefit from us being unhurried and able to focus better.
Yet it's awfully tempting to manage a multi-agent workforce. To see the AI agents scurry off and get things done while you check and supervise their work. I'm not saying you should stop doing this altogether. What I'm saying is that you should be aware of the potential harm that comes from this way of working.
Setting the boundaries for how many tasks you allow yourself to have in progress at a time can also be applied to AI agents: run only as many agents in parallel as you can actually review, and finish what's in progress before starting something new. Turn off the notifications that tell you an agent has finished. Instead, check on them at times you choose. This is slower but I'd argue that in the long term it's more sustainable. It also helps preserve your own ability to do focused work by keeping the periods of instructing and reviewing separate from those of deep thought and concentration. You set the pace that feels comfortable and decide when to shift between the work only you can do, and that which you delegate to AI.
Finally, if you feel like you've already lost the ability to stay focused, you need to start practising it. Turn off all sources of interruption (email, group chat, phone) and set a timer. See if you can stay with one thing for 15 minutes without becoming distracted. When the timer rings, take a breather. Practise these 15-minute intervals and then gradually increase the time as staying focused starts to get easier.
Make work visible
Knowledge work is intangible. You can't see at a glance how much of it is piling up, so you have to make it visible before you can plan it realistically. That applies to the work itself, and also to the information it depends on. In order to use AI effectively, both matter.
Start with the work. Much of what we now do with AI remains invisible. Writing and rewriting prompts, reviewing outputs, correcting errors and running the same task again with better instructions don't appear on any task board or in any timesheet. In a Workday survey of 3,200 employees who use AI, nearly 40% of the time AI saved was lost again to fixing errors, rewriting and checking. Only 14% of employees consistently got clear, positive net outcomes from AI.
This is different from the occasional half hour lost to a tool that won't cooperate or a task that takes a little longer than expected. It's a steady cost that comes with almost every task where AI plays a part. But since nobody sees it, expectations are based on a misconception of reality, and everyone ends up permanently behind schedule. It's easy to imagine that surely AI helps you get more things done. However, our view of work often lacks the nuance that only becomes visible in the moment the work is being done. I dare say we make this error all the time even with our own tasks (which is why things often end up taking longer than expected), and it just gets worse when we try to imagine what someone else's work entails.
As an example of this kind of invisible work, imagine a task that takes about 8 hours, but can only be done effectively when you can focus on it without distractions and interruptions. If you can work on it in 2-hour time slots, and finish them all within a week, it's likely that the task will be done close to the estimated time. However, if you can only work on it haphazardly for 15-30 minutes at a time before having to do something else, the task can easily consume significantly more time than estimated. The key factor is not the content of the task, but how much time is wasted by context-switching. The latter never shows in any project plan because it's dependent on working culture, not the work itself.
Then the information. In order to use AI to help with your tasks, or delegate some work entirely to AI, it needs to understand what the task is about. Often this means that AI needs information regarding the context of the task, and the better it understands it, the better job it will do. In this sense AI benefits from exactly the same things a self-directed human professional would: clear tasks, roles and responsibilities, up-to-date documentation and easy access to the information that is needed to do their work well.
Publicly available knowledge is rarely the problem. The bigger hurdle is making sure that internal information — whether that's internal for your organisation, team, project, or just information stored in your own notes and documents — is available to the AI. Everything should also be organised and formatted in such a manner that AI can decipher and process it easily. That way the AI doesn't have to wade through irrelevant material to find what it needs, which helps preserve both tokens (less money wasted) and context window (AI retains more capacity to hold relevant information).
It's somewhat ironic that missing, outdated and disorganised information has been causing unnecessary cognitive load and dysfunction in human work all along. Perhaps organisations will finally be compelled to pay attention to documentation practices, because good documentation is also needed to make better use of AI. At the very least we can use AI to help us organise information and improve the way things are described and documented.
As a personal example, I have been taking notes from books, research papers, trainings and various articles in Evernote since 2008. During the past few years, I've slowly migrated these notes to Markdown files, and shifted my note-taking to systems that store information locally in non-proprietary file formats. I've had Claude create instructions for itself on how to navigate these files and find what is relevant without having to process thousands of pages every time. It now has access to all the information I've found worth saving or making a note of during close to two decades of working life. This has completely changed how I work. Knowledge I had forgotten has become more accessible and usable when it is needed.
I am now evaluating my use of other software tools as well. To see if I could free up their data. In many cases, a piece of software is just an interface for processing the data within. When the data is freed, AI can be used to process it directly. As needed, AI can also be used to conjure the necessary interfaces for accessing and manipulating the data, such as forms, tables, queries, reports and dashboards. Except that the interface will be tailored to my preferences, instead of being dictated by a software vendor whose design philosophy I might not agree with. And I am of course also free to add more features and make changes to those interfaces when the need arises.
This makes licence-fee-based SaaS platforms vulnerable. Especially the ones used for a single process in large organisations, such as the travel expense claims I mentioned earlier. Building and maintaining an in-house replacement with the help of AI is becoming a realistic option, as long as there is someone who knows how to take responsibility for security, integrations and compliance. And considering the level of UI/UX in many enterprise tools, the bar for a better and more user-friendly design is set rather low. Tools that are easy to use and automate the tedious parts of a process will also impose less extraneous cognitive load on their users.
As I am writing this, different companies are starting to release their autonomous personal AI assistants (e.g. Muse by Meta, Dots by OpenAI and Claude Duet by Anthropic; there are also numerous startups and open-source projects on this). The idea is that with access to the same information you have, AI can proactively do some of the work for you, remind you of things that need your attention, and take care of menial tasks such as scheduling. Want us to find time for a lunch meeting? My AI assistant can talk directly to your AI assistant and they can work out the details. It's like having trusted and personal human assistants at work and at home. What was once reserved for corporate executives and the ultra-rich is becoming accessible to everyone.
The usefulness of these assistants, like that of human assistants, is severely limited by the information they have access to. Can they see your email and chat messages? Can they access your calendar? What about your task and knowledge management systems, and other tools that you use to get things done? If the answer to all of these is yes, an AI assistant could do something no single tool has managed so far: bring together the commitments scattered across your inbox, chat, calendar, project plans and task lists, and help you make better sense of them all.
Whether or not you should trust all this data to a vendor that operates the AI is not really a topic of this essay. Suffice to say, it's plausible that in a few years' time we have powerful enough AI running locally on our own hardware, which would also help keep our data more secure.
Working well with AI requires visibility in two places. The AI needs access to contextual information in order to do good work. And you need to make explicit the work that AI creates for you, including all the checking and correcting, in order to plan realistically and make sure you can maintain a sustainable pace.
Manage your energy
In the sustainable pace framework, I use energy as an umbrella term for the state your mind is in at any given moment. It covers how alert or tired you feel, and also how easy or hard it is to think, concentrate, get started, remember things and keep your emotions and impulses in check. If you’ve slept badly, your mood is sour and nothing seems to hold your attention, it's going to affect both the efficiency and the quality of your work.
In knowledge work the effort we put into a task does not in itself make the task any more worthwhile. Small things can have a big impact. Similarly, huge undertakings can turn out to be a complete waste of time. When stressed and exhausted, it becomes difficult to see which is which, to make good decisions and to be thoughtful. In order to spend your limited attention on things that make a difference, and to do your work well, it pays to maintain good, positive energy.
For most of us, work consists of tasks that vary in their scope and cognitive demands. From short, few-minute things to those that take hours or even days to finish, and from those you can jump right into to those that take time to get oriented before you can begin to make any kind of meaningful progress. Ideally, a workday should contain variety between different types of tasks instead of spending it all in "cognitive middle gear" as put by James Hewitt of Hintsa Performance. Alex Soojung-Kim Pang also argues in his book Rest that most people can sustain intensive, focused work for only about 3-4 hours a day.
The thing to be aware of when it comes to using AI is how it can change this mix of work. If AI takes over the quick and easy tasks, what's left for you are the things that require judgement: evaluating outputs, making decisions, giving directions, and spotting and fixing errors. That reduces variety and affects how exhausting your work feels.
For example, the bread-and-butter corporate trainings I do on Sustainable Pace in Knowledge Work take 3 hours. That's high-intensity work for me. I can usually ride the rest of the day on the adrenaline rush that comes from being in front of an audience, but the following day is a different story. I feel knackered, and I've learned over the years that on those days I should do only low-effort work. But when AI has already taken care of all the easy tasks, all I have left is the work that my energy levels are not suited for, which then impairs my recovery. And if I do nothing I feel like I am failing to earn my keep (thank you, Lutheran upbringing and work ethic!). I would also not consider checking the work that AI does as low-effort, as it often takes focused attention and requires judgement.
I also have a tendency to worry. Especially when I feel exhausted, and even about things that I can't control or influence. That saps energy, which means there's less of it available for my family, myself, and for doing the work that matters. You can probably relate if you're worried, for example, about AI and how it will affect your future. Unfortunately, it's precisely these kinds of disconcerting moments of transition, where the future of work is being explored and created, that would benefit from creative thinking, optimistic opportunism, and experimentation. When low on energy, it's difficult if not impossible to operate in the mindset that would actually help you deal with the situation.
The tendency to worry is a built-in feature of our survival mechanism. It's not easily overwritten. Here in Finland the winter always comes. So you had better be prepared for it. We have survived for generations in this climate because our ancestors worried about having food and shelter to last through the long cold darkness.
A sense of security is needed to keep the survival mechanism from taking over. Now imagine if the leadership of an organisation talks about AI bringing a fundamental shift to how work is being done, while also emphasising the need to cut costs. That leaves people to wonder: when are we going to see changes, who is affected, and how exactly? Will there be layoffs? This only creates more uncertainty and worry about the future. People's minds go into survival mode instead of the playful, optimistic, creative and explorative mode. As a result, the whole organisation is worse off when it comes to making the most of the AI transition.
As I mentioned earlier when discussing the Embrace uncertainty principle: many lean transformations failed because people were laid off after the first improvements yielded results. If the CEO and the leadership team openly stated that no person will lose their job due to AI-driven improvements in productivity, it would create a sense of security and support the conditions needed for exploration and experimentation. You'd also be more likely to get those productivity improvements as a result. Nobody can know yet how the change will play out, but leadership can still decide what they commit to while it's underway.
In addition to needing safety and security, we also need to be operating from a state of positive energy in order to be creative and try out new things. Maintaining that energy means paying attention to rest and recovery, and one of the most important factors there is the ability to detach from work. To fully disconnect, and to direct our thoughts somewhere else. Will always-on AI assistants become just another burden on our cognitive resources? Will AI be one more thing sitting alongside Teams, Slack, Outlook and the like on our phones? Will we end up checking AI outputs late into the evening just so that it can continue working while we're asleep?
The pull is already there. In the UC Berkeley study, people sent prompts during lunch, before meetings or in the evening when an idea came to mind, which dissolved some of the natural stopping points in the workday. They described being less able to fully disconnect. Left unchecked, that kind of strain builds towards exhaustion and burnout: even familiar tasks start to require more effort, the ability to recover becomes impaired, you're prone to making more errors, self-regulation suffers, and making simple everyday decisions can start to feel overwhelming.
Kevin Kelly has written that productivity is for robots. Humans, in his words, "excel at wasting time, experimenting, playing, creating, and exploring". The value of that kind of work bears little relation to the hours put into it, and it's the kind of work AI should be making more room for. If we're not careful and don't set boundaries, it can go the other way. Instead of directing the machines so that we can do more of the work that is uniquely human, we end up spending our evenings in their servitude just to keep them productive. AI becomes a drain on the very energy it was supposed to free up.
I'm not saying this will happen. I'm just saying that it's one plausible scenario. We're currently in the middle of learning to work with AI, co-creating the future of work where AI is as omnipresent as the internet. In doing this we are bound to face issues that have to be problem-solved on the fly.
Managing energy on an individual level, in teams, and in organisations is what makes it possible to approach this transition with wisdom and thoughtfulness. You can't have effective, creative, innovative, self-directed, assertive and productive knowledge workers if they're consistently feeling stressed, worried, exhausted and overwhelmed.
Finland, 2011
Conclusion
Looking at AI through the six principles of sustainable pace in knowledge work, the same few patterns keep coming up. The capacity AI frees up stays free only if protected. The work AI leaves to us can be demanding and exhausting: instructing, clarifying, judging, checking and deciding. The more things AI generates for us, the more our attention, judgement and energy become the bottleneck.
AI tends to amplify what’s already there. Where there is slack, a limited amount of work in progress and well-organised information, it can help individuals, teams and organisations to spend more of their time and energy on work that matters. To create better results, impact and value. But insert AI into an already overloaded system, and the overload is likely to grow with it.
How an organisation uses the gains that come with AI adoption also matters. If people see productivity turned into layoffs, or fear the possibility, they stop exploring what AI could do, which is like shooting oneself in the foot after the first few miles of a marathon.
The framework used here looks at the use of AI from a human perspective, while also taking into account productivity and effectiveness in organisations. The two must go together if we want to create a future of work where the goals of employees and employers are aligned.
The dynamics behind these patterns are not new. The efficiency trap, the need for slack, the overhead caused by work in progress, and the limits of attention were all there before AI. The principles of sustainable pace were developed to deal with them. I'm certain there is much about AI that falls outside the perspective taken in this essay, but as long as humans are needed for knowledge work, I dare say the six principles remain a good guide for thinking about where and how AI should be used.
If you want to apply this to your own work, here are a few questions to start with:
Has AI helped you spend more time and energy on valuable work, or has it led you to just produce more things that aren't really needed?
Assuming AI has saved you some time, where has that time gone? And has any of it been kept free for dealing with surprises?
Has AI increased or reduced the amount of work in progress? What would happen if you imposed stricter limits?
Is the amount of time and effort required to supervise, check and correct AI-generated work visible anywhere?
How often does a notification from an AI tool pull you away from something that needed your full attention?
With AI doing some of the work, what is left for you? How much variety is there? Do you find your work interesting and worth your while, or has it turned into monotonous drudgery?
If you are in a leadership position, how do you talk about the use of AI? Whether intentionally or not, do you provoke fear and uncertainty? Or are you creating a safe space where people can explore and experiment with new tools and technologies?
I hope that you got some insights and ideas on what to consider when it comes to the use of AI in your own work and organisation. If you'd like to discuss this more, please leave a comment.
Thank you for reading all the way to the end. If you found the essay worth your while, please share it with someone you think would find value in it.
Helping organisations make knowledge work effective and sustainable is what I do for a living. I deliver trainings, workshops and talks. I also coach people and teams. If you're looking for someone with whom to think these questions through in your organisation, I am available for advisory work. You can find out more at samipaju.com, or just send me a message.
Sources
Burkeman, Oliver (2021). Four Thousand Weeks: Time Management for Mortals. Vintage.
Goldratt, Eliyahu M., & Cox, Jeff (1984). The Goal: A Process of Ongoing Improvement. North River Press.
Junni, P. P., Sarala, R. M., Taras, V., & Tarba, S. Y. (2013). Organizational ambidexterity and performance: A meta-analysis. Academy of Management Perspectives, 27 (4), 299-312.
March, James G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, Vol. 2, No. 1, 71-87.
Newport, Cal (2024). Slow Productivity: The Lost Art of Accomplishment Without Burnout. Penguin Random House.
O’Reilly, Charles A., & Tushman, Michael L. (2004). The Ambidextrous Organization. Harvard Business Review, April 2004.
O’Reilly, Charles A., & Tushman, Michael L. (2013). Organizational Ambidexterity: Past, Present and Future. Academy of Management Perspectives, Vol. 27, No. 4, 324-338.
Paju, Sami (2026). Kestävä työtahti. Tuuma Kustannus.
Pang, Alex Soojung-Kim (2016). Rest: Why You Get More Done When You Work Less. Basic Books.
Raisch, S., & Birkinshaw, J. (2008). Organizational Ambidexterity: Antecedents, Outcomes, and Moderators. Journal of Management, 34 (3), 375-409.
Taleb, Nassim Nicholas (2007). The Black Swan: The Impact of the Highly Improbable. New York: Random House.
Torkkola, Sari (2015). Lean asiantuntijatyön johtamisessa. Talentum Pro.