AI can produce more work in less time. The harder question is whether it gives people and organisations the capacity to understand more, decide better and contribute what only they can bring.
We keep being shown the same promise.
A report that once took hours can be produced in seconds. A presentation appears from a conversation. A film that once required equipment, specialist software and a production team can be made by describing what you want to see. The demonstration ends with the finished thing on the screen and the time saved beside it.
It is compelling because the capability is real. I have used it. I have taken thoughts that would once have remained scattered in my head and worked through them in conversation with an AI model. I have used AI to search, challenge, organise, build, write and create. I have also watched it misunderstand what I meant, produce something polished but wrong, lose a decision we had already made and confidently take work in a direction I had explicitly rejected.
Both realities can be true at once.
Sometimes the saving is substantial. Sometimes the work merely moves into prompting, checking, correcting and rebuilding. My question begins where a genuine saving remains: what should that time become?
AI is an extraordinarily powerful tool. We are still behaving as though the only important thing it changes is how quickly something can be produced.
That feels like a remarkably small ambition.
We are living through two loud conversations about AI. One asks whether it will replace people, take control or eventually threaten humanity itself. The other promises that it will save time, reduce headcount and generate more work with fewer people. Somewhere between the fear and the sales demonstration sits a quieter question that may matter much sooner to most organisations.
What do we mean by productivity now?
The definition measures what is easiest to count
The conventional definition is straightforward. The OECD describes labour productivity as total output per hour worked.[1] More output from the same number of hours means higher productivity. The same output produced in fewer hours does too.
That works when the output is visible and reasonably consistent. If a factory produces more units from the same machinery and labour, there is something concrete to count. If a farmer grows more food from the same land and resources, the improvement can be measured. The calculation becomes less comfortable when the output is a recommendation, an investigation, a strategy, a hiring decision or an understanding of why three people left the same team.
In knowledge work, the completed artefact is often mistaken for the outcome.
A report has been generated, but is the organisation better informed? A summary has been written, but did it preserve the disagreement that mattered? A presentation has appeared, but does the person presenting it understand the argument? A decision arrived more quickly, but was it the right decision, and what will it cost to correct if it was not?
If the unit we count is the report, then producing it in two minutes looks spectacularly productive. If the real output is a reliable decision, we also have to count the evidence that was missed, the time required to verify the claims, the corrections made later, the consequences of an answer shown to the wrong person and the organisational trust lost when confidence outruns understanding.
The economic definition does not need to be thrown away. We need to become more honest about what we place on each side of it.
For AI-assisted knowledge work, the input is not only the time taken to generate the first answer. It includes the human attention required to frame the question, examine the sources, test the answer, recognise what is missing and decide what may follow. The output is not simply the document that appeared. It is the value, reliability and consequence of what the organisation does with it.
More output is not automatically more productivity. Sometimes it is simply more output.
We have mistaken acceleration for transformation
Much of the current marketing around AI stops at the moment of production.
Look how quickly the report appeared. Look how easily the video was made. Look how many tasks the agent completed. Look how little technical knowledge the person needed before they could begin.
There is real value in that. Removing technical barriers can allow somebody with a story to tell to make a film without first becoming an editor, cinematographer and sound engineer. Removing hours of data assembly can allow a manager to ask a question while it still matters. Making specialist capability more accessible can give smaller organisations opportunities previously reserved for those with larger budgets and teams.
I do not want to diminish any of that. I want to ask why our imagination so often ends there.
We keep treating a jackhammer as though it were merely a faster hammer. The real opportunity is not to drive more nails. It is to break through walls that were previously beyond us.
I recently spent a considerable amount of time trying to direct a film through an AI system. I could describe the world, the movement, the feeling and the story in ordinary language. The system could produce images and sequences I could never have made alone. Some of them were stunning. It also changed things I had not asked it to change. Fixing one element could rebuild another. A sequence could look beautiful and still communicate the wrong idea.
The physical production barrier had fallen. The need for direction had not.
My work moved. I was no longer operating a camera or editing every frame, but I still had to know what I meant, recognise when the result contradicted it and keep hold of the whole while the system responded to one fragment at a time. The technology made production possible. It did not supply the reason the film deserved to exist, the judgement required to shape it or the responsibility for what it eventually said.
That distinction reaches far beyond creative work. AI does not always eliminate work. Sometimes it relocates the work from production into direction, examination and judgement.
If organisations measure only the time removed from the first part, they will miss the value required in the second.
The time saved has to go somewhere
I remember research before search engines became an ordinary part of life.
You found the library. You worked out which catalogue might contain the subject. You located the books, found the relevant chapters, followed references and tried to determine which account deserved your confidence. It was slow. Sometimes the information you needed was not available at all. Sometimes you discovered something important because you had to remain with the question for longer than you wanted.
AI changes that relationship with information. It can bring a vast body of knowledge within reach and offer a useful first synthesis in seconds. That knowledge is not complete. The model may be out of date, the source material may be wrong, and the answer may flatten genuine disagreement into a smooth paragraph. Even so, the expansion of access is extraordinary.
That should create more room for thought.
If finding the initial information no longer takes three days, those three days do not have to disappear. Some of that time can be used to ask better questions. What does this answer assume? Which evidence supports it? What would change the conclusion? Whose experience is absent? Is the relationship the system identified meaningful, or merely available? What happens if we act on it?
AI should reduce the time we spend finding information. It should not remove the time we spend thinking about it.
Research is beginning to identify the same tension. Studies have found measurable gains from generative AI in professional writing and customer-support work, although the effects differ by task and worker.[2][3] A 2025 study of 319 knowledge workers found that higher confidence in generative AI was associated with less reported critical thinking, while greater confidence in their own ability was associated with more. The researchers also found that critical thinking did not simply disappear. It shifted towards verifying information, integrating responses and stewarding the task.[4]
That last point matters. The human contribution does not become unnecessary because AI completes part of the work. The contribution changes. Whether it becomes more valuable or merely more passive depends on what the person and the organisation believe the tool is there to do.
The danger begins when the efficiency at one end becomes an excuse to abandon the work at the other. The answer arrives quickly, reads well and appears plausible. It becomes tempting to use it because it sounds like a good answer. The person stops examining what is in front of them.
That is what I mean when I say that any tool in the wrong hands can be dangerous. I am not speaking only about malicious people. The wrong hands can belong to a perfectly responsible person who has stopped being conscious of the tool they are using.
Assistance becomes reliance. Reliance becomes habit. Eventually the habit becomes delegated judgement.
I use AI without handing it the conclusion
That distinction is the reason I can use AI as extensively as I do.
I trust the conversation to help me articulate something that is not yet fully formed. I trust the process enough to ask questions, test an argument and expose gaps in my own thinking. I can talk through an idea, receive a response and realise that the response has misunderstood me. Explaining why it is wrong forces me to become clearer about what I actually mean.
I do not transfer ownership of the conclusion.
An AI model can produce a persuasive answer. Persuasion is not evidence that the answer is true. It can find a connection I had not noticed. The connection still needs examination. It can draft something in my voice. I still have to recognise whether the thought belongs to me. It can tell me that a claim appears well supported. If that claim matters, I want to know where it came from.
This is not distrust in the sense of refusing the technology. It is an active relationship with it. I am using the model as a tool for thought while retaining responsibility for what I allow that thought to become.
We already understand this principle in other forms of knowledge. Reading one book does not require us to accept its author as the final authority on the subject. We read another. We compare accounts. We notice where they agree, where they conflict and what each author could see from their position. Scientific knowledge depends on methods, evidence, challenge and the ability of others to test what has been claimed. A result earns confidence through more than the fluency of the person describing it.
AI has made synthesis easier while making the source of that synthesis less visible. That does not make critical thinking old-fashioned. It makes it one of the central capabilities of the AI age.
People need more than prompt-writing lessons. They need to understand when AI is appropriate, how to question what it produces, how to distinguish a source from a summary, how to recognise uncertainty and when to stop using the tool and seek another form of expertise. They also need enough confidence in their own experience to disagree with an answer that arrives carrying the weight of a machine.
The question is whether organisations will give them permission to do so.
The organisation determines whether people can think
It is easy to turn this into a lecture about individual responsibility. Read the output. Check the sources. Do not become lazy.
Individuals do carry responsibility. They do not work in a vacuum.
If an organisation introduces AI with a promise that everybody will produce twice as much, it has already told people how the saved time will be used. If success is measured by response volume, completion speed and lower cost, pausing to investigate a plausible answer can look like poor performance. Critical thinking becomes an unofficial extra undertaken by conscientious people after the measurable work is complete.
The organisation cannot praise judgement in its values and punish the time required to exercise it.
Compliance will remain necessary. Some activities need rules, controls and records because the consequences demand them. Yet compliance represents the minimum condition for responsible operation. It does not describe the full opportunity. A compliant organisation can still be incurious. It can follow every required process and make a poor decision because the people involved never had the time, context or permission to ask what the information meant.
The larger opportunity is organisational intelligence: the ability to bring relevant information together, preserve the relationships between it, recognise what remains uncertain and allow experienced people to make better-informed decisions.
That requires selective friction. Nobody needs a committee to approve every spelling correction or routine administrative action. Bounded, low-consequence work can and should move quickly. The need for human examination increases with ambiguity, irreversibility and consequence. A benefits calculation affecting somebody's income deserves a different level of scrutiny from the formatting of an internal agenda. A statement entering an employee's record deserves more care than a first draft that will never leave its author.
Good organisational design does not slow everything down. It makes clear where speed is useful and where speed can conceal damage.
The time released by AI therefore becomes an organisational resource. Leaders decide whether to consume it immediately through higher output expectations or invest it in stronger judgement, deeper customer understanding, learning, experimentation and the development of their people.
AI can create a productivity dividend. Organisations will decide whether it becomes only a financial dividend or a human one.
The colleague whose title was Typist Three
In the early 2000s, I worked at an organisation with a colleague whose formal title belonged to another technological era. They were a typist. There were pay grades behind it: Typist One, Typist Two, Typist Three.
The title made sense in the history of the organisation. Work had once depended on typewriters and early word-processing systems. Producing accurate documents quickly was a recognised specialist skill, and they were one of the fastest typists I have ever known.
Technology changed around them. They learned with it.
By the time I knew them, the value they brought could no longer be contained by the title on the pay scale. They became the executive assistant and personal assistant to the chief executive. The typing did not suddenly become worthless. It stopped defining the boundary of what they could contribute.
Think about what sits inside that progression. An executive assistant does not merely produce words more quickly. The role requires context, discretion, anticipation, relationships, memory and judgement. It requires an understanding of the person, the organisation and the significance of what is happening around both. Technology absorbed or simplified part of the task. Their career grew towards the human capabilities the task had never fully represented.
That outcome was not guaranteed.
The organisation could have treated new technology solely as an opportunity to remove a typing post. They could have been denied the time or support required to learn. Somebody could have looked at the title, concluded that the task had disappeared and missed the person standing behind it.
Instead, their contribution expanded.
Technology can remove the task without removing the person. It can reveal that the person was capable of far more than the task for which the organisation originally employed them.
The learning curve is central to that possibility. It is often discussed as a temporary inconvenience on the way to adoption. In reality, the learning curve is the human investment that determines whether technological change produces progression or displacement. People need the opportunity to acquire the tool, question it, develop alongside it and discover where their own value now sits.
Their story also prevents us from romanticising the past. I am not arguing for the return of slow machinery or for people to spend their lives completing work a tool can do safely and accurately. I am asking whether we can be more ambitious about the person whose capacity is released when that work changes.
What might today's equivalent of Typist Three become if an organisation looks beyond the disappearing task?
Productivity can expand the person
The current discussion often treats employment as a contest between a human and a machine. Which tasks can the machine perform? How many people will the organisation still need? How much cost can be removed?
Those questions belong in business decisions. They are not the whole decision.
AI can give a junior employee access to explanations, examples and initial guidance that previously depended on the availability of a more experienced colleague. It can help somebody explore a field that once felt closed to them. It can make technical production available to a person with vision but without years of specialist training. It can allow an experienced professional to examine more evidence than one person could reasonably gather alone.
These uses do not make expertise irrelevant. They change how expertise develops and where it is applied.
There is a genuine risk here. People often learn through the apparently ordinary work that comes before senior judgement. The junior researcher learns by finding the material. The new analyst learns by building the first model. The coordinator learns how the organisation works by preparing the meeting and following what happens afterwards. If AI removes every developmental task, organisations may discover that they have also removed the route through which future experts acquire context.
The answer cannot be to preserve unnecessary labour for its own sake. It is to redesign development consciously. If the machine completes the first draft, the person can learn by examining its choices, testing alternatives and understanding why one answer is stronger than another. If AI assembles the data, the person can spend more time learning how to interpret it, while still understanding enough about its origin to recognise when the assembly is wrong. If the tool makes a film possible, the emerging creator can concentrate on story and direction without being allowed to mistake technical ease for creative mastery.
Organisations now have an opportunity to create roles around judgement, connection and growth. That requires more imagination than deleting the tasks automation can perform and dividing the remainder among fewer people.
It also requires a different way of recognising potential. The colleague I remember was not valuable because their original role had survived unchanged. They were valuable because they could grow as the work changed. In a period when tools will continue to move quickly, learning capacity may become one of the most important forms of organisational value.
AI may remove the old route into expertise. Organisations will have to build the next one deliberately.
But giving people room to develop judgement means little if the system has already decided what they are allowed to see, what remains disputed and which answer appears settled.
A human in the loop can still be absent
The phrase human in the loop appears everywhere in discussions about responsible AI. It is reassuring. A person remains involved. Somebody will approve the answer before anything consequential happens.
The presence of a person does not guarantee the presence of judgement.
If the system has already selected the evidence, compressed the disagreement, written the conclusion and presented it as settled, what exactly is the person being asked to approve? If the reviewer cannot see what was excluded, does not understand how the answer was formed and faces pressure to move quickly, the approval may be human only in the biological sense.
This is where the language of accuracy can become too narrow. Of course an AI answer should be accurate. Accuracy still leaves questions. Accurate according to which evidence? Current at what time? Appropriate for which purpose? Visible to whom? Recorded as an observation, an interpretation or an authorised finding? What should happen when new evidence changes the organisation's understanding later?
Consider the confidence we place in a share price. No single commentator makes a price true by announcing it persuasively. Behind the number sits a system of recorded transactions, timestamps, market rules and calculation methods. We can disagree profoundly about why the price moved or what it will do next. We can still trace the conditions under which the quoted number came into existence.
An AI-generated sentence can arrive without an equivalent chain. Its coherence becomes a substitute for provenance. The organisation knows what the answer says without being able to see clearly how that answer acquired the status it now carries.
Responsibility can disappear in the same gap. The provider created the model. The organisation deployed it. An employee entered the request. A manager approved the result. If everybody is described as responsible in general, nobody may be accountable for the particular consequence.
The responsibilities are different. Providers must be responsible for the capabilities, representations and protections surrounding what they release. Organisations must be responsible for where the technology is deployed, what information it may use and what authority it is given. Individuals must remain conscious of how they use it. The person authorised to make a consequential decision must retain responsibility for that decision.
Technology should make those boundaries easier to see. It should not allow them to dissolve between a prompt and an outcome.
A button is not authority. Authority requires the information and freedom necessary to make a real decision.
Why I built TURI
I did not build TURI because I believe AI should be stopped. I built her because I believe something this powerful should be used without requiring people to surrender judgement, authority or ownership of what follows.
TURI began with an organisational problem I had seen long before generative AI arrived. Organisations collect fragments of reality from people, systems, documents and decisions. As those fragments move, their relationships can disappear. An account becomes a category. A possibility becomes a finding. The newest version overwrites the history that explains why it changed. Information reaches somebody because it is technically available rather than because it is appropriate for them to see.
AI accelerates that old weakness. It can turn partial evidence into a complete-sounding answer before the organisation has determined what the answer is allowed to mean.
TURI is designed to keep the journey connected. Evidence remains tied to its source. Relationships, competing perspectives, uncertainty and revisions remain available to the people permitted to examine them. Identity, access, factual status, memory, wording and visibility do not collapse into one blanket approval. The person with the relevant authority remains responsible for what becomes consequential.
An AI model can participate where interpretation or drafting is useful. It does not grant itself permission, decide that its proposal is an organisational fact or determine who should see the result. It is optional and replaceable. TURI can govern a bounded transaction without an AI model when interpretation is unnecessary.
This matters because meaningful human authority needs more than a final click. The authorised person needs to see enough of the connected evidence to understand what they are being asked to decide. They need to know what remains disputed. They need to distinguish what the model proposed from what the organisation has actually established. If the understanding changes later, the organisation should be able to revise it without pretending the previous version never existed.
That is why I resist describing TURI only as an AI guardrail. Governance is part of what she produces, but the deeper purpose is connected, revisable organisational intelligence. The goal is not simply to prevent a model from doing something harmful. It is to help organisations use evidence and intelligence in a way that supports better-informed human decisions.
TURI is my response to the same question running through this essay. Once technology has given us access to more information and removed some of the labour required to assemble it, how do we ensure that the human contribution becomes deeper rather than thinner?
The answer cannot live only in training. People need the skills to question AI, but organisations also need systems that preserve what those people require in order to question it meaningfully.
What we choose to do with the gain
I do not know whether AI will eventually become the existential threat some of its pioneers fear. I do know that organisations are making decisions about it now.
They are deciding what productivity means. They are deciding which tasks disappear, which roles develop and whether the time saved belongs entirely to the output target. They are deciding whether a polished answer is enough, whether a human approval is meaningful and whether people will be taught to work with AI without handing their judgement to it.
None of this requires us to deny the economic reality. Organisations need to compete. Efficiency matters. Some roles will change profoundly and some will disappear. Technology has always altered the shape of work, sometimes creating opportunity and sometimes leaving people to carry the cost.
When I put the library, the film and the colleague whose title was Typist Three together, they show three different ways technology can change the work without deciding what the gain will become.
The library reminds me how much of the old labour lay in reaching the information at all. AI can remove much of that labour. The question is whether the time it releases is used to test sources, follow disagreement and ask better questions, or whether the first synthesis becomes the final answer simply because it arrived quickly.
The film reminds me that removing a production barrier does not remove the need for direction. A beautiful sequence can still communicate the wrong idea. Technical ease can widen access to creation, but it cannot decide what deserves to be made or take responsibility for what the finished work says.
Typist Three reminds me that a task can shrink while a person's contribution grows. Technology changed around them. They learned, and the organisation gained access to more of what they could bring. That is not a sentimental footnote to technological progress. It is one possible definition of progress.
Together, those stories expose the limit of measuring productivity only by output per hour. In knowledge work, the finished artefact may be evidence of activity without being evidence of value. The value may lie in the quality of the decision, the strength of the understanding, the ability to revise it when new evidence arrives or the growth of the person trusted to act.
The difference will not be determined by the AI model alone. It will be determined by the environment around it: what the organisation measures, what it rewards, what it preserves and whether its people have the information, confidence and authority to ask why.
AI's greatest productivity gain may not be the number of things it allows us to produce. It may be the time and information it gives people to contribute more of what only they can bring.
But time saved does not automatically become time for thought, and human potential does not automatically become human progression. Organisations have to choose what the gain is for.
Once the tool has saved the time, what will the organisation make possible for the person?
References
1. Organisation for Economic Co-operation and Development. Labour productivity and utilisation. Accessed 11 September 2026. Source
2. Shakked Noy and Whitney Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science 381 6654. 2023. Source
3. Erik Brynjolfsson Danielle Li and Lindsey R Raymond. Generative AI at Work. Quarterly Journal of Economics 140 2. 2025. Source
4. Hao Ping Lee and others. The Impact of Generative AI on Critical Thinking. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 2025. Source