AI is making us less productive

The headlines have been piling up over the past year:

“80% of companies see no productivity gains from AI”: an NBER study of nearly 6,000 executives found that the vast majority of companies see no measurable effect of AI on either productivity or employment (Fortune, Feb 2026).
“AI intensifies work instead of reducing it”: Harvard Business Review reported that employees supervising AI tools experience increased mental fatigue and decision exhaustion (HBR, Feb 2026).
“AI-generated workslop is destroying productivity”: 41% of employees have encountered AI-generated material that required nearly two hours of rework per instance (HBR, Sep 2025).
“Time on email has doubled, focused work down 9%”: companies rolling out AI saw workloads increase instead of decrease (Fortune, Mar 2026).

So AI is overhyped? It doesn’t work?

No. But the research shows something important: it doesn’t work to just throw tools at people.

The problem isn’t AI, it’s how companies introduce it

MIT Sloan has identified an “AI adoption J-curve” — companies that simply buy AI tools without adapting workflows and training staff see worse results initially (MIT Sloan). BCG confirms it: “AI transformation is a workforce transformation” — the companies that succeed have the most ambitious training programs, not the most expensive tools (BCG, 2026).

The usual pattern:

Leadership decides: “We’re going to use AI now”
Everyone gets a ChatGPT license
No training, no adapted workflows, no follow-up
Employees copy-paste a bit, think it’s “alright”, and drift back to the old way
Six months later: “AI gave us nothing”

The problem was never the tools. The problem was that nobody actually learned to use them. Giving the team ChatGPT without training is like buying Photoshop for the whole department and expecting everyone to produce great design.

This is about business development, not technology development. Which workflows can be improved? Which repetitive tasks can be automated? How do roles need to change? Those questions take time, training and change management, not just a new license.

The difference: copy-paste vs. work partner

It really starts with a fairly simple question: How do you use AI?

Most of us probably started the same way: open ChatGPT, type a question, copy the answer somewhere. That’s how AI is presented in most contexts, and that’s where many people get stuck. It’s also part of why the surveys above look the way they do.

We started out with things like “write a text about UX design”, got a fairly generic answer, and figured there wasn’t much to it. Then it sat unused for a while.

There was nothing wrong with the AI. We were just using it like a search engine with extra steps.

Chatbot vs. work partner

There’s a real difference between chatting with AI and working with AI:

Chatting with AI:

You write a prompt, get an answer, copy it somewhere
You do everything manually, AI is just a text generator
Every new question starts from scratch, with no context
You’re still in the browser, copying and pasting

Working with AI:

AI has access to your files, your project, your context
You describe what you want to achieve, AI does the work
AI remembers what you’ve worked on and builds on it
The result lands directly where it should, no manual steps

It’s a bit like the difference between Googling “how to build a shelf” and having someone next to you who can hold the drill while you steady the board.

What is an AI agent?

An AI agent acts, not just answers. It can:

Read and understand your files — it sees your whole project, not just what you paste
Create and edit files — it writes code, text and documentation directly in your files
Run commands — it can start servers, run tests, install tools
Plan and execute — you give it a goal, the agent breaks it into steps and carries them out
Learn your way of working — through config files it understands how you specifically want to work

Why we landed on Claude Code

We’re UX/AI designers, not developers. But Claude Code has become the tool we use most during the workday.

Before that we tried a few: Lovable, Figma Make, Cursor, Antigravity, Manus, Codex, Augment Code. They all have their strengths, and which one fits best probably depends on what you do. For us, Claude Code is the one that feels closest to a work partner rather than just a tool.

A few things we got hooked on:

1. Agents: Claude Code doesn’t just run one prompt at a time. It can spin up several agents in parallel that work on different parts of a problem at the same time. That can save time, especially when the task splits naturally.

2. Skills: You can build your own “skills” that teach Claude Code how you work. We have some for UX work, some for brainstorming, some for communication and admin. Over time it learns how we specifically want things done.

3. Context: Through a simple file (CLAUDE.md), Claude Code understands your project, your code structure, your preferences. It doesn’t start from zero every time.

A typical workday with AI

Here’s what a day can look like. Not every day flows like this, but the pattern is familiar:

08:00: We open Claude Code in our current project. It reads CLAUDE.md and gets a baseline understanding of what the project is about.

08:15: We need to sketch a new landing page. Instead of jumping straight into Figma, we ask for three variants of a hero section, shown as simple wireframes. Not all of them are good, but it’s enough to find a direction to work from.

09:00: We pick a variant and ask for a quick HTML prototype. It’s not perfect, but good enough to show and get reactions to.

10:00: After some feedback, we convert the prototype into a React component that follows our design system. It takes a few tweaks, but the foundation is in place.

13:00: After a client meeting we need to send a follow-up. We let a skill we built go through our notes and suggest a structure. We rewrite it in our own tone before it goes out.

15:00: A bug in production. We describe the symptoms, Claude Code looks at the logs and suggests a likely cause. Sometimes it’s right, sometimes we need to dig further. Today it was quick.

We spend very little time copy-pasting text between windows, and much more time actually thinking.

It’s not about coding

The most common objection we hear: “But Claude Code is for developers, right?”

Not really. Claude Code is just as much for anyone who wants AI to help with concrete tasks, not just answer in a chat. You don’t need to know how to code to say “create a prototype”, “summarize this meeting” or make a presentation.

And you don’t even need to use the terminal if you don’t want to. Claude Code is now available as a desktop app for Mac and Windows. That’s probably the easiest way to get started if you’re not used to terminals — it looks pretty much like an ordinary chat window.

Next step

In the next module we’ll go through the tool map: which AI tools exist, what they cost, and which are worth a look depending on your role.

This piece was co-written with Claude. We know what we want to say, but we’ve used Claude to draft suggestions that we then read, adjusted and made calls on. The phrasing, the opinions and the choices are ours.