The AI productivity paradox: why AI’s real value isn’t productivity, and how to capture it

Every wave of software has been sold on productivity: do things faster, do more with less. So when AI arrived, productivity was the obvious promise, and an understandable one.

Except AI isn’t another phase of software. It’s more powerful than that.

We think that if you judge it by how fast a task gets done or how much more you can produce, you’re underestimating what it can do for you. OpenAI’s chief scientist calls it an alien mind. We’d call it intelligence on demand: something that lets a company solve problems that used to be out of reach.

This post is about the AI productivity paradox: what AI productivity should actually mean, how to measure it, and the one thing most companies have to fix before any of it pays off.

What does the data say about AI productivity?

It depends on who you ask and what they measured. The studies that measure a task find big gains. The studies that measure a company find small ones or none.

And to be fair, it’s understandable that the data tell different stories. First, AI as a mainstream tool is still new. The bigger shift, autonomous agents doing real work end to end, only arrived this year. Every new technology takes time before work gets reorganized around it and before we see any significant incremental gain.

Second, productivity is notoriously hard to measure, and AI makes it harder. It’s very hard, and very tempting, to think about productivity in absolute terms: we used to produce five books, now we produce ten. We don’t think organizations can easily draw the line between output productivity and input productivity.

“You can see the computer age everywhere but in the productivity statistics.” Robert Solow, Nobel Prize-winning economist, on the productivity paradox of the computer age.

Why does AI make people busier?

Ask the people using AI and the answer is consistent. It saved them time, and they’re busier than before.

Since AI lets us do things that we haven’t been able to do before, we’re now stretching ourselves even thinner and managing even more tasks. For example, a product manager who used to write specs now, with AI, writes specs, builds the front end with an AI coding tool, and manages three agents running in the background. The organization delivered the product faster. The product manager did a job they should never have had to do, and did it for twenty hours a day.

“The organization delivered the product faster, but the individual actually felt busier. That’s the paradox of productivity with AI.”

People who use AI are working more. They spend more hours on the computer and more hours doing more things, because their cognition is now holding more projects at once. Their capacity for projects went up, and they became the ones responsible for every one of those work streams. Organizational productivity is much higher. Personal productivity is the opposite. We don’t think many people using AI heavily would say they’re spending more time with their family because they finished early.

So yes, AI gives us more capabilities, but it also makes us busier than ever:

• Workscope expands: Each person’s cognition is now holding more projects at once, and they own every one of those work streams. More capacity means more responsibility, and more hours.

• Constant context switching: This is a pain every people manager already knows. When you do a one-on-one with someone who’s been deep in a project for a week, you can’t absorb their context in thirty minutes. Now everyone working with agents feels that pressure, because every agent is a direct report with its own context. As a species, we’re not naturally good at this. It’s a skill, and we’re all learning it at once.

What does AI productivity actually mean?

The technical bucket: how is the AI performing?

The technical bucket is how well the AI itself performs, and it’s easy to measure. This includes:

• Tokens per second: How fast the model generates output. An absolute unit; higher is better.

• Latency: How long a job takes end to end. Every task should finish in a relatively short time.

• User turns per task: When you submit a task, how many back-and-forth iterations does it take before the session is done? The ideal is one. You say what you want, the agent does it, the artifact is created. You don’t even say thank you.

That third one is the one most teams don’t track and the one that matters most. Every extra turn is a human checking, correcting, or re-prompting. If your agents need five turns to finish what should take one, you don’t have autonomy. You have a very fast assistant that needs supervision.

The organizational bucket: is the company using it?

The organizational bucket is whether the company is actually using AI, and what it can now produce that it couldn’t before.

Here we’d measure tokens per employee. We think it’s one of the most important numbers a CIO can watch. If your employees aren’t consuming tokens, they’re using calculators where everyone else is using Excel. And to be honest, in 2026, that’s… weird.

Tokens per employee tells you adoption depth, not just adoption. A company where 90% of staff “use AI” but token consumption is flat has a lot of people asking ChatGPT one question a week. That’s not an AI-native company. For more on how we think about top-down and bottom-up adoption, see our enterprise AI strategy for adopting agents top down and bottom up.

It’s not a coincidence that these curves rise together. Google Trends shows interest in “AI-native” climbed from February 2026 and peaked in April and May. “AI-first” rose over the same months. And “tokenmaxxing” didn’t register as a search term until March 2026, then peaked at the end of May. Companies started asking how to become AI-native and how to use more tokens in the same window.

The value of AI is solving problems you couldn’t before, not just productivity

In a recent interview on The Ezra Klein Show, Bill Gates pushed back on the usual comparison. People say this is like the PC, he said, and it isn’t: “This is like evolutionary history. This is like the aliens really are here and have come.” His example of what that means: “Read every medical journal and see if there’s anything that we didn’t spot? The AIs do that today.”

If intelligence is no longer the scarce thing, something else is. Alexandr Wang, who runs Meta’s superintelligence lab, says it’s vision and ambition, and calls this “a once in a civilization opportunity to be a dreamer and to have a vision and to have ambition.”

The purpose of a company is to solve problems. So the purpose of AI inside a company isn’t to be a productivity tool. It’s to solve more problems. That includes the ones nobody has been able to take on: a disease that affects a thousand people worldwide never gets a drug program, because no one has the capacity to spend years on it. With AI, someone can. For the first time, you have intelligence on demand.

So the question for a CIO changes. Not “how productive can we be with AI” but “what problems can we solve with AI.”

“It’s not about productivity. It’s about problem solving. For the first time ever, you have intelligence on demand.”

How can companies capture the full potential of AI?

The bottleneck to using AI to its full potential is a lack of trust, and most companies respond to that by over-engineering a solution to a problem that could be fixed faster. Here’s a better way to work with AI:

Give AI an ambitious problem

Take a problem every company has: churn. It’s big, it’s ambiguous, and nobody knows the cause. It could be support, the product, too many bugs, a competitor. So companies go top-down. VP of Support, raise customer satisfaction. VP of Marketing, increase MQLs. VP of R&D, lower the bug count. The churn goal gets broken into departmental goals and nobody owns the whole thing.

Now imagine instead you give one agent the actual goal: find the root cause of churn. You have access to every database and every department. Work horizontally. We’d bet you’ll find AI is very good at siloed data. It can move between marketing data, sales data, R&D data, and engineering data in a way no leader can, connect to those systems on its own, and go down the rabbit hole of what happened. It might decide to go through every support case from the last two years. Nobody configured it to do that. Nobody told it to read the tickets.

“People are not giving ambitious tasks to AI.”

Make it safe enough to try

Companies don’t run that experiment because they don’t trust AI with that much access, and the trust is missing because nothing in their stack makes it safe. Handing an agent the whole company sounds reckless. It isn’t, if the agent works in read-only mode, inside a sandbox, under governance, with a kill switch.

This is the point we keep coming back to on governance: it doesn’t reduce autonomy, it makes autonomy possible. Most companies resolve the tension between control and enablement by letting agents do only low-stakes work, which is the trust problem. The way out isn’t to shrink what agents may do. It’s adding the right controls that let you trust them with more: an access policy set once at the executive level and enforced on every agent, deterministic human-in-the-loop for the handful of actions that need a named approver, an audit trail that shows who triggered what, and a way to test an agent against realistic data before it touches production. We laid out that view in our enterprise AI strategy for adopting agents top down and bottom up.

Fix the infrastructure first

Existing enterprise infrastructure wasn’t built to run AI at scale. Companies lack governance tools. They lack development tools. They can’t deploy and test different models (Claude, Codex, OpenCode, GLM, Kimi or DeepSeek) against the same task. And AI needs access to their systems, which it doesn’t have. So the biggest problem, before trust and before anything else, is that the systems aren’t AI-ready and there’s no infrastructure for letting AI run securely and efficiently.

You download ChatGPT, it isn’t connected to your systems, and it doesn’t work: it can’t answer your clients, and it can’t go through your Slack and notice something you haven’t asked for. If you had the infrastructure, you wouldn’t need to take our word for any of this. You could run the experiment. Motivation and top-down commitment matter too, but assuming you have those, that’s it.

Building those layers yourself takes anywhere from eighteen months to two years, in our experience, before you can see the first use case. Buying them separately doesn’t help, because then you’ll need to figure out how to make those tools work together.

A real example: how AI changed bug fixing at xpander

Every software company has bugs. By definition, you don’t plan engineering capacity for them, because features make money and bugs are things customers already paid for that broke. So bugs eat capacity nobody budgeted. And about 95% of the work of fixing a bug is finding it: reproducing the exact configuration, the exact click, the exact user state that triggers it.

The first generation of AI tools, a year ago, helped with that the way a smart colleague would. You’d paste the error, ask it to help reproduce the bug, and it would find the problem faster and maybe propose a fix. Real productivity gain. Still a human in the loop at every step.

Today, inside xpander, the process looks different. A bug report comes in. Our bug-fixing agents take the screenshot, scan every scenario that could affect it, and spin up a swarm of agents to reproduce it. Once the root cause is found, they propose a fix and run the analysis a human would never get to: this broke because of an unrelated change two weeks ago. All of this happens overnight. We come in and the agent manager reports that it fixed seven bugs, found two more that nobody had reported but that could have surfaced, and fixed those too. It’s already in the product.

This is the difference between a productivity tool and a problem-solving one. Using AI to debug faster is productivity. Letting agents own the bug queue is solving a problem the company had stopped trying to solve.

What should a CIO do first?

Stop measuring hours saved and start measuring whether the company can solve problems it couldn’t before. Then fix the infrastructure, because that’s what decides whether you can give AI anything worth solving.

Here’s how it looks in practice:

• Measure AI usage: Tokens per employee (adoption depth) and user turns per task (autonomy). If the first is flat, people aren’t using AI. If the second is high, your agents aren’t autonomous.

• Pick one ambitious, cross-silo problem: Churn, onboarding time, incident volume.

• Check whether you could run that experiment today: Read-only horizontal access, sandbox, audit, budget cap. If any of those is missing, that’s your first project, not the agent.

It’s not about productivity. It’s about problem solving. For the first time, you have intelligence on demand. The companies that get value from it are the ones that give it a real problem and a safe place to work.