The AI Productivity Paradox

That was Robert Solow in 1987. Nearly four decades later, we're hearing the same thing about AI.

The Economist argued last week that AI's (current) effect on productivity is "thin (subscription required)." The same week, a survey of thousands of CEOs showed no measurable impact on employment or productivity. If you only read the headlines, you'd conclude the tools aren't working.

I disagree. The tools are working, extraordinarily well, for a small fraction of people who've figured out how to use them effectively.

Erik Brynjolfsson, the Stanford economist who literally coined the term "productivity paradox" back in 1993, now argues the tide is turning. His analysis suggests US productivity grew roughly 2.7% in 2025, nearly double the decade average. But buried in his research is the more revealing finding: most businesses still use AI for narrow tasks like summarisation and email drafting, while a small cohort of power users are compressing weeks of work into hours by automating workstreams.

That's not a technology problem. That's an adoption depth problem.

The 90/10 Split

Asana's research across 9,000+ knowledge workers found that just 10% of the workforce, the "super productive," save 20 or more hours every week using AI. The other 90% are getting modest gains at best. And ManpowerGroup's 2026 Global Talent Barometer revealed something even more telling: regular AI use increased 13% last year, but confidence in the technology dropped 18%. More people are using AI, and more people think it doesn't work.

That confidence drop isn't because the tools are bad. It's because people are using Claude and ChatGPT the same way they used Google: type a question, get an answer, move on. If that's your entire workflow, then yes, AI feels like a slightly better search engine. Not worth the subscription.

If you read my issue on The AI You're Paying For vs The AI You're Using, you'll remember the three levels of AI use. They're the key to understanding this paradox.

Level 1: AI as Assistant. Question in, answer out. Email drafts, meeting summaries, basic research. This is where most people sit and where The Economist's "modest impact" lives. It's also where you'll stay if nobody shows you what else is possible.

Level 2: AI as Thinking Partner. You're working with AI on sustained problems. Iterating on a strategy document across multiple sessions. Using it to stress-test your thinking, not just execute tasks. This is where the 10-25% productivity gains from the research actually show up.

Level 3: AI as Builder. You're creating things that didn't exist before. Not doing existing work faster, but expanding what you're capable of doing. Prototyping products in a day. Building internal tools without an engineering team. This is where weeks compress into hours.

Moving from Level 1 to Level 2 isn't a technology upgrade. It's a relationship change. You stop querying and start collaborating.

The new paradigm is pair-learning with an AI build partner.

Five Shifts That Change Everything

For anyone willing to push through the learning curve, the capabilities are available right now. But they require a different mental model. Here are five mindset shifts that separate the 10% from the 90%.

1. Start with vision, not task. Start with the big picture: here's the project, here's the stakeholder, here's what I'm trying to achieve, here's what good looks like. If you read my issue on why context beats prompting every time, this is that principle in action. Context-first prompting saves time downstream because AI doesn't have to guess what you actually need.

2. Think out loud. AI isn’t judging your lack of coherence. Most people wait until their thinking is polished before engaging AI. That's backwards. Your AI partner doesn't need perfectly formed thoughts. Much of its value is in helping you work through ill-formed ideas (I know…), spotting gaps you haven't consciously caught, and playing your thinking back to you in a way that sharpens it. The messier your input, the more room AI has to add value. Advanced Voice Mode in ChatGPT is great for this.

3. Challenge AI. Every. Single. Time. AI says everything confidently. Challenge it. Ask it to challenge you. The best results come from productive friction, not mutual agreement. If you're nodding along to everything your AI produces, you're not getting the full value.

4. Let AI draft, then curate. Society needs to get comfortable. First drafts should be written by AI. Take advantage of near-infinite output capacity. Don’t limit yourself to one output when you can ask for five approaches, three structures, ten headlines. Then react, edit, combine. This is fundamentally faster than crafting from scratch, but only if you stop treating AI's first response as the final answer.

5. Manage AI like an intern. AI will allow you to pull on any thread you’re willing to entertain. You are the architect of the conversation. Decide what matters now versus later. Manage the session like you'd manage a project, because that's exactly what it is.

Companies: Stop Saying "Go Figure It Out"

Here's what frustrates me most about the current state of AI adoption. Most organisations have bought the licences. They've sent the "AI is now available" email. And then they've told employees to go figure it out.

That's not a strategy.

The companies that will pull ahead aren't waiting for AI to get better. They're investing time now for their people to learn how to work differently. Not a one-off lunch-and-learn. Dedicated, protected time to experiment, fail, and build new habits. This is non-negotiable if you want to be competitive in 18 months.

And the baseline isn't complicated. These three things puts your company ahead of your peers:

1. Get the right licences in place.

2. Help people understand what the tools can actually do (most are unaware).

3. Provide customised, function-specific training.

A product manager's AI workflow looks nothing like a finance analyst's. A marketing team needs different patterns than an engineering team. One generic training session won't move the needle.

This needs strategy, thought, and guidance. It's something I work on directly with organisations through Atomic Theory Consulting. If your company is stuck at Level 1, or you're watching adoption plateau after the initial excitement, I'd welcome a conversation.

What's Next

The people who embrace this new way of working will shape the next generation of productivity. The paradox isn't that AI doesn't work. It's that it works incredibly well for the people who've changed how they think about it, and barely registers for everyone else.

The Atomic Builder is written by Faisal Shariff and powered by Atomic Theory Consulting Ltd — helping organisations put AI transformation into practice.