The great illusion of technological productivity gains (including AI)

For more than half a century, each technological wave has been accompanied by the same promise: to produce more, faster and at lower cost. We heard it when microcomputers arrived in the 1980s, during the massive rollout of ERP systems in the 1990s, when the internet and then mobile technology burst onto the scene, and again today with artificial intelligence. It’s always the same promise, always the same narrative of a historic breakthrough that is supposed to transform productivity.

And yet history tells us something else. The gains appear later than hoped for, in smaller proportions than announced, and often benefit players other than those who made the investment. And AI is no exception to this scenario. Behind the enthusiastic rhetoric, the figures show that the promised productivity often remains just that: a promise.

The aim here is not to criticize the technology gratuitously, but to recall the lessons of the past that we refuse to learn and to show that, ultimately, there are constants that impose themselves on us. This is therefore a broader reflection on how organizations approach each wave of technology and why productivity, far from being an automatic given, is a much more complex subject than we are willing to explain.

It is a subject that is complex at its core and that human ingenuity has even ended up making complicated. That is why, rather than writing an article even longer than the ones I usually write, I have, as you have probably seen, dealt with different pieces of this puzzle in previous articles, and this article will serve as a “meta-article“”” with a summary of each point raised. Feel free to read the relevant article for a more detailed analysis.

To conclude this preamble, however, I would like to emphasize one point. With every technological breakthrough, we are constantly told about revolution, that nothing will ever be the same again, that everything that was valuable before no longer works, and recent history reminds us once again that this is not true and that there are laws which, much like the law of gravity, are constants that cannot be overcome.

Yes, technology changes, but people and human nature do not change, and so it is no surprise that history keeps repeating itself and will continue to do so for a long time to come, in my opinion.

Change and, beyond that, productivity cannot be decreed and do not happen “automatically” with technology. They are the product of a system, and it is within that system that they are born, evaporate, move, or do not necessarily benefit those we think they will.

[Note written after finalizing the article]: This article is still very long, with ideas that emerged as I was writing it, which clearly shows that the subject of the link between productivity and technology cannot be dealt with simply by taking easy shortcuts.

In short:

• The promises of productivity gains linked to technological innovations, particularly artificial intelligence, are repeated with each new wave but rarely materialize in a significant or fast manner.
• Productivity does not automatically result from the introduction of a technology, but depends on a transformation of practices, organizational coherence, and an overall strategy.
• Apparent productivity gains can shift or be diluted, sometimes benefiting other actors (suppliers, customers, technostructure) rather than the companies that invest.
• The indicators used to measure the effects of technology are often insufficient or biased, making it difficult to truly assess the benefits obtained.
• AI, like previous technologies, is only effective if it is used by competent people in a well-designed framework; otherwise, it increases complexity and disappointment.

• The unfulfilled promise of AI
• How much credence should we give to measurement methods?
• Productivity evaporates in your dreams
• Productivity is not revenue
• Technology does not automatically create productivity
• Productivity gains spill over, and not just to yours
• Local optimum does not mean global optimum
• You generate value, but others reap the benefits
• Your productivity gains are captured by your employees and managers
• Technology creates new uses and tasks
• Transferring costs does not mean increased productivity
• AI in the wrong place, for the wrong people
• Bottom Line
• To answer your questions

The unfulfilled promise of AI

Artificial intelligence is described as the next industrial revolution, an unprecedented driver of productivity that is set to transform organizations. Executives present it as essential, consultants as a historic breakthrough, and analysts as a necessary step to remain competitive. However, the reality is quite different.

The figures published over the last two years are clear: 95% of generative AI pilot projects fail to produce a tangible impact (MIT report: 95% of generative AI pilots at companies are failing). Behind the hype, use cases remain confined to prototypes that, although attractive, are difficult to integrate into everyday processes. Projects fail for simple reasons: costs are too high, results are too unpredictable, integration into existing practices is difficult, or there is resistance from business units that do not see the value of tools they did not choose.

Independent analyses agree. Most businesses are unable to transform AI into measurable benefits (Why AI Disappoints At Productivity – But Excels At Ambition). When there is a gain, it is one-off: a text generated more quickly, a synthesized search, but this gain is immediately offset by the time spent checking, validating, and correcting.

We could talk about the paradox of a technology that raises expectations and investments while producing so little measurable value for those who adopt it, but in fact, it’s just history repeating itself, and AI is just another episode in a series we know well.

How much credence should we give to measurement methods?

When trying to measure a gain or even a loss, two things must be measured: the situation before and the situation after. This may seem obvious to you, but it is not obvious to everyone, since in most cases things are not measured before the technology is deployed, which means that all the figures you read on the subject must be taken with a pinch of salt, regardless of whether they are good or bad.

Worse still, we don’t even measure things after deployment, which is ultimately logical because we have no reference point for comparing the figures (How can we measure the productivity gains of AI?).

So most of the figures you read are based on feelings and declarations. We’ve seen better ways to take a rational approach to a subject.

Productivity evaporates in your dreams

The discourse on AI is saturated with announcements that far exceed the actual capabilities of the technology and, above all, presuppose the existence of conditions for success that rarely exist in many businesses today. There is also a tendency to confuse predictions with forecasts, the former being based on intuition, personal beliefs, or media posturing, while the latter are based on data, methods, and analytical rigor. In public debate, this distinction disappears and the promises of AI productivity are based on risky predictions dressed up as scientific forecasts (AGI, employment, productivity: the great bluff of AI predictions).

Adding to this confusion is the weight of marketing that no longer settles for touting benefits, but uses fear as a selling point. Businesses are told that if they do not immediately adopt a particular technology, they are doomed to disappear (If you don’t buy my products and services, you’re all going to die.). Productivity and its consequences (job losses) are then presented as an imperative for survival, for lack of anything else to sell.

Under this pressure, businesses invest heavily, not based on a clear-headed assessment of their needs, but because they fear missing out on a supposed revolution, the famous FOMO (Fear of Missing Out). They are engaged in projects that are driven more by urgency and fear than by value.

So we end up with executives who don’t understand the subject but put pressure on their teams to “do stuff with AI” with unrealistic expectations. The moral of the story: it goes in all directions and doesn’t produce much (Under Pressure: Engineering in the Age of AI and Why tech can fail in the last mile: The devil is all in the detail).

And, as in the past, we are also making the mistake of believing that because people have adopted AI in their private lives, they will naturally and usefully do so in their professional lives (Why the widespread adoption of AI by consumers says nothing about its future in the workplace).

AI therefore suffers from the same problem as many technologies before it. It’s not that it doesn’t work, that it doesn’t do what it’s designed to do, but that we expect too much, too quickly, and in the end we are disappointed and throw the baby out with the bathwater.

Productivity is not revenue

A study by MIT is often cited, which tells us that 95% of AI pilots fail, in order to heavily criticize the technology (MIT report: 95% of generative AI pilots at companies are failing). But aside from the fact that the study’s sample is debatable, and even though I am convinced that AI will never live up to its promise, I am surprised that no one has noticed that this statement is based on a huge misinterpretation. Indeed, we are told that businesses are not seeing an increase in revenue, when the promise of AI lies in productivity!

This may be the biggest mistake made when evaluating technology in business, and AI is no exception to the rule: we implement technology to improve productivity and then measure revenue to see if it has worked.

Unfortunately, productivity and income have never gone hand in hand, and never will. To put it another way, if that’s not clear enough: revenue is not a function of productivity!

Even for a freelancer, where one might think the link is mechanical because they work alone, are both the input and output of the workflow, and therefore there is no risk of loss, this does not work. Productivity gains are easier to measure than in a business (e.g., delivering 20% faster), but they do not automatically translate into revenue.

Consultants paid by the hour deliver faster but bill less, and it is their clients who benefit from the savings. Graphic designers paid a flat fee maintain their price, but they only earn more if demand follows. Productivity then becomes a resource that they can turn into revenue, provided they find new clients (For freelancers, productivity does not always pay off).

In a business, the situation is even more complex. Individual gains are diluted in processes, come up against bottlenecks, and disappear in organizational friction. Producing faster does not mean selling more, and selling more does not mean earning more if you have overestimated your capabilities or if your clients ask you to cut margins that they suspect are on the rise.

In other words: to maintain revenues, you need to find commercial opportunities equivalent to the increase in productivity observed (for example: 10% increase, 10% more customers), otherwise revenues will decline.

Productivity is a valuable commodity, but it is not income. It can only be converted into revenue if sales and even marketing follow suit, and it is highly dependent on the market to transform the time saved into value.

Technology does not automatically create productivity

As old pots make the best soup, I can only remind you of economist Robert Solow, who summed up the situation with a now famous phrase: “We see computers everywhere, except in productivity statistics“. Forty years later, the phrase is still relevant.

Businesses are investing heavily in technology, but the impact on overall productivity remains invisible (You can see the computer age everywhere but in the productivity statistics (Robert Solow)). It is believed that productivity automatically results from the accumulation of technical resources, but the reality is that technology produces nothing if it is not accompanied by a transformation of uses and practices.

Here again, history is repeating itself. ERP systems were supposed to streamline management and improve processes, but in organizations that had not reviewed their operating methods, they mainly added rigidity and costs. AI is unsurprisingly following exactly the same trajectory.

It is worth remembering that it was not electricity that increased factory productivity. In water-powered factories, the machines had to be arranged in a line, aligned with the transmission shaft connected to the mill, and when electricity arrived, many manufacturers simply replaced the water wheel with a central electric motor, keeping exactly the same layout. The result: no significant gains. It was only when it was understood that an individual motor could be installed for each machine and the organization of the factory completely redesigned that productivity really took off.

Productivity gains spill over, and not just to yours

If Solow took us back to 1987, we will now go even further back in time, to the 1950s.

It was during this period that Alfred Sauvy noticed a simple but rather counterintuitive phenomenon: productivity gains do not disappear, but neither do they remain where they are produced. They move, like a burden that we think we have eliminated but which reappears elsewhere.

Let’s take a concrete example. A business decides to automate its customer service. On paper, it’s a success: fewer advisors, lower costs, “streamlined” processes. But the problem hasn’t gone away. It has simply been transferred to the customer, who has to search for information on their own in an online portal, explain their problem ten times to a chatbot, or waste time navigating menus. What looks like a productivity gain for the business is actually a waste of time, energy, and satisfaction for the user (Are you familiar with Sauvy’s spillover theory on productivity gains?).

The same phenomenon occurs within organizations. To lighten the load on support functions, operational employees are delegated tasks that they did not have to manage before: entering expense reports, booking travel, filling out reports. Support departments then post better ratios, but the time “saved” is lost twice: it eats into employees’ productive time and increases their mental load.

This transfer gives the impression that productivity has increased, when in fact the work has simply been shifted, and this shift comes at a cost. When it weighs on customers, it damages the commercial relationship. When it weighs on employees, it undermines their engagement and ultimately reduces collective performance.

What Sauvy demonstrated remains true today: technology often makes it possible to get rid of a local problem, but it does not automatically create overall value, as we will see later. It shifts costs, and that always ends up being paid for somewhere.

Local optimum does not mean global optimum

Eliyahu Goldratt pointed out that the performance of a system is limited by its bottleneck. Until this issue is addressed, optimization elsewhere is useless (Local optimum vs. global optimum and the theory of constraints: why your productivity gains are sometimes useless).

In a business, automating or speeding up certain tasks makes no difference if the bottleneck lies in validation, decision-making, or process integration. Local gains only create a queue of tasks and to-dos that are delayed elsewhere, create waste, or shift the problem.

Productivity is not the sum of micro-optimizations. It is a global dynamic, limited by the constraints that determine the capacity of the system, and “augmenting” certain employees with AI in no way guarantees that the productivity of the business will be increased (AI in the workplace: going beyond augmentation to actually transform).

And for those who are closely following the “revolution” that is supposedly underway at Moderna, which emphasizes the concept of flow, I think that the challenge will be to monitor and measure gains “from end to end” and not just locally (Thinking of work as a flow: appealing, but is it realistic?). Now that we have sufficient feedback, we can see that individual appropriation contributes little compared to a workflow approach (How to Make Enterprise Gen AI Work).

Besides, I don’t know whether to laugh or cry, but it still took McKinsey over a year to discover that if you don’t start with a process-based approach and a redesign of work and workflows, silver-based AI will only bring disappointment (One year of agentic AI: Six lessons from the people doing the work).

And even without a process or activity-based approach, it is collective ownership alone that can truly commit the business to a process of improvement and progress (There’s no (A)I in Team).

You generate value, but others reap the benefits

The businesses that invest in technology are not always the ones that reap the benefits. Carl Shapiro and Hal Varian have shown that in a value chain, productivity gains are often captured by other players, those in strategic positions.

We saw this with personal computing: user businesses invested, but it was Microsoft and Intel that captured most of the value. We are seeing the same thing today with the cloud and AI: organizations are spending, but it is Amazon, Microsoft, Google, and Nvidia that are reaping the margins (Are you familiar with the “value chain squeeze” or how your suppliers and customers are stealing your productivity gains?).

Productivity gains do exist, but they are not reaped by those who bear the effort and the cost.

Your productivity gains are captured by your employees and managers

Even when productivity gains exist, they do not always translate into better economic performance. They can be absorbed within the organization itself, by employees, by managers, or more broadly by systems.

The first form of capture comes from the employees themselves. The time freed up by a tool or automation is transformed into comfort: a little more flexibility here, a few informal breaks there, more time for training. This is not lost on the employee, but it is less obvious for the business because it is not visible in financial indicators or quantifiable by the business. This is Parkinson’s law: “Work expands to fill the time available”. In other words, any time saved ends up being filled (Are you familiar with Parkinson’s law on how your employees manage their own time and productivity?).

And if it’s not your employees who reap these gains, it will be your managers who will reclaim them, and again, not in a productive way: they become a pretext for adding control tasks, reporting, and meetings, as Galbraith described so well in his analysis of the technostructure. Organizations have a natural tendency to become more complex and generate administrative work that justifies their own existence, and what we thought was a net gain will end up dissolving into bureaucracy (Are you aware of the technostructure that devours all your efforts to improve?).

Technology creates new uses and tasks

In the 19th century, William Stanley Jevons observed that the more efficient a resource became, the more its consumption increased. What he described with coal now applies to digital technology.

Tools introduced to reduce workload often have the opposite effect. Instant messaging, which was supposed to speed up communication, has saturated organizations with interruptions. Automation systems, which were supposed to simplify processes, have multiplied control points and layers of complexity. Generative AI, which was supposed to reduce production time, multiplies content and creates an additional need for validation. (Are you familiar with the Jevons paradox at work? When efficiency leads to inefficiency.).

Unfortunately, the logic is relentless: every apparent gain opens up new possibilities for use, which translate into increased consumption of time and resources, and not always for useful and productive things. Efficiency becomes overload.

An excellent example of this is what is known as “AI Workslop”. This refers to work produced by AI that looks clean and professional but, in reality, lacks substance and shifts the burden to the recipient to understand, improve, rephrase, or even completely redo it. Rather than simplifying everyday life, it creates extra work, generates frustration, wastes time, and reduces trust between employees. In terms of efficiency, it weakens collaboration and negates much of the productivity gains expected from AI (AI-Generated “Workslop” Is Destroying Productivity).

Transferring costs does not mean increased productivity

Let’s go back even further, to 1937. That was the year Ronald Coase explained that the existence of businesses can be explained by transaction costs. When it is simpler and less expensive to coordinate an activity internally, the business integrates it. Conversely, when these costs decrease, it outsources.

Digital technology has profoundly changed this equation. By reducing transaction costs, it has encouraged outsourcing, but this outsourcing often results in a simple transfer of burden.

Towards the customer, with self-service: automatic terminals in airports, self-service checkouts, banking apps. Towards the employee, with administrative tasks shifted to operational staff: expense reports, reporting, reservations. The business shows gains, but the work has only changed hands. On the one hand, customers may become weary, and on the other, operational staff find themselves inundated with administrative tasks that only benefit support functions. (Are you familiar with Coase’s law on transaction costs and its impact on capturing productivity gains? and Employee self-service: how far to go before you go too far).

In a similar vein, remember that the introduction of email led to the dismissal of many secretaries, resulting in some of the secretarial work being transferred to managers, who gained in workload but lost in productivity in their profession (If the AI Industry Fails, It Could Take the Rest of Us Down With It).

AI in the wrong place, for the wrong people

AI adds no value in itself, but amplifies the value of those who already master their subject, because they are able to evaluate, validate, and supervise the results produced ([FR] Add AI where there are skills. Not where there is a lag.)

In the right hands, it speeds up work, enriches analysis, and frees up time, but in the hands of people without sufficient skills, it adds no value. On the contrary, it creates uncertainty and dependency, and forces the organization to put in place new layers of validation.

The mistake is to believe that AI can compensate for a skills gap, but in reality, it widens the gap between those who know and those who don’t.

Bottom Line

The same story keeps repeating itself, but each time some people still manage to be surprised: technology does not bring productivity in itself. Gains shift, dilute, evaporate, or benefit others. AI is no exception.

It can speed up certain tasks, but without transforming practices, without a comprehensive approach to workflows, and without a clear strategy, it repeats the scenario of disappointed (over)promises.

The real challenge is not to produce faster, but to know how to appropriate these gains, what to do with them, and how to transform them into revenue and growth. It is indeed a little too easy to run a pilot, see that the numbers are not there, and condemn the technology.

Were the objectives realistic? Where did we go wrong in the deployment? What have we learned to start over and do better?

But the truth is that more often than not, we observe, condemn, and move on without learning anything, when this learning loop should be at the heart of AI projects (95% of Enterprise AI Pilots “Fail”–Just Like Lean? Not So Fast).

Once again, the problem is not the technology, which we are often quick to blame when it disappoints, but rather our expectations of it, what we do with it, and our understanding of the macro logic that explains that what appears to be a problem is sometimes just the nature of things.

To answer your questions

Image credit: Image generated by artificial intelligence via ChatGPT (OpenAI)