Productivity Paradox
Have we reached the AI tipping point?
In 1987, something caught the attention of economist Robert Solow. "You can see the computer age everywhere but in the productivity statistics."
The technology was on every desk, but it wasn’t reflected in the numbers. This became synonymous with the productivity paradox. Almost four decades are we looking at AI as another one?
Individuals are getting faster.
There's good evidence supporting that AI makes individuals more productive. McKinsey’s August 2026 State of AI survey found, 80% of respondents say it has improved their personal productivity, but only 37% report a positive contribution to their organisation's EBIT.
It’s not necessarily that AI isn't working. It seems that making a person more productive and making a business more profitable are two different things. An employee might write a report in half the time, but what happens to the time saved? Does it go into higher-value work, better service or more customers? Or does the business simply produce more work because work has become easier to produce? Most of us have experienced what happens when it gets easier to send emails.
Then the billing arrives
For an individual, AI maths is simple: pay for a subscription and see if it helps. This is more complicated when a business deploys agents that do work autonomously.
Take Microsoft's Copilot Cowork. The underlying Microsoft 365 Copilot licence is still a predictable per-user cost, but Cowork adds usage-based pricing. Tasks consume Copilot Credits, depending on the models used, the context retrieved, the runtime and tools involved. Pay-as-you-go credits are currently $0.01 each. That seems reasonable, since an agent doing more work should cost more. The difficulty is predicting it. Microsoft has improved cost visibility. Users can type /cost in Cowork to see the approximate credits a task has consumed so far, while administrators can set spending limits, receive alerts and analyse consumption by user, group, agent or service. This is not the same as knowing the price upfront: Microsoft states users cannot check the cost of a task before running it,
/cost is an approximation rather than a billing record. Businesses can now contain unexpected spend more effectively, but they still cannot fully predict it.
Imagine a business of 50 people. Around 30% become heavy users, 30% occasional users and 40% barely touch it. That group of heavy users drives most of the consumption and the bill climbs quickly. We've already heard of businesses running up thousands of dollars of usage in a matter of days. That may well be money well spent. The point is that the business needs to know up front, otherwise what is the point of budgets, financial planning. If AI spend doubles in a month, someone has to be able to explain why and pay the unexpected bill. It’s not feasible for most people to run a business on the basis that a technology bill might double and you'll work out the reason and its value add afterwards.
Price isn't the real issue
The conclusion isn’t that AI is too expensive. A $10,000 monthly bill is excellent value if it creates $20,000 of measurable benefit. Whereas a $1,000 bill is poor value if nobody can say or see what it achieved.
That echoes the original paradox. Businesses had to do more than install computers; they had to reorganise work around them and train people to use them. With AI, people need to be trained to use it securely and accurately, we need to decide which work suits AI and which should stay with people, how workflows should change and how the value will be measured.
Deloitte's 2025 research found that only 6% of respondents achieved payback on a typical AI use case within a year, with two to four years the most common timeframe. It could be that we are still learning how to measure its return on investment. It’s often quoted in business it generally takes three years to turn a profit.
Uncertainty can also suppress experimentation. "We think this could save 100 hours a month, but we can't say what the AI will cost" is a difficult sentence to say out loud to a finance director.
So, what's next… who knows
For the last few years the conversation has been about adoption, training and FOMO. Getting people to use AI and choosing the tools. The next phase is making it useful, measurable and financially predictable, which means understanding usage, setting sensible limits and tying AI activity back to business outcomes.
Microsoft has added spending limits, usage alerts and reporting at tenant, group and user level.
The goal isn't to use less AI or more of it, but to make the most of it: to understand where AI creates value, what that value costs and how to scale it responsibly.
The real tipping point may be when a business can answer: what is our AI doing, what is it costing us and what benefit is it actually delivering?
Further Reading
Robert Solow 1987 We'd Better Watch Out Wikipedia
McKinsey's 2026 State Of AI Report
