Originally presented at Wroclaw AI Meetup before React Universe Conference. A deep dive into the mismatch between AI investment and actual developer productivity.
Key Takeaways
- Coding is only 16% of developer time
Most work involves planning, documentation, reviews, and communication- 60% of teams see no AI productivity boost
Despite heavy investment, most organizations report minimal gains- Trust drops as usage increases
46% distrust AI accuracy (up from 31%), frustrated by subtle bugs- Focus AI on the other 84%
Documentation, presentations, and tool integration offer bigger wins- Model Context Protocol changes the game
Connect AI directly to Jira, Notion, GitHub without copy-paste
The Productivity Paradox
AI coding assistants are everywhere now. GitHub Copilot, Cursor, Claude Code. Most developers use at least one. Companies pay for enterprise licenses. Everyone expects massive productivity gains.
Researchers tracked 617 engineering teams in 2025. 60% of leaders said AI hasn’t significantly boosted productivity. Meanwhile, developers report something strange: they save 10+ hours weekly with AI but still lose 6+ hours to the same old problems. Meeting overload. Documentation debt. Seven different places to update when you ship a feature.
Developers spend only 16% of their time writing code. Even if AI made coding twice as fast, you’d only improve overall productivity by 8%. The real opportunity sits in the other 84% of the job.
What does that look like in practice? Build your next presentation in React instead of wrestling with PowerPoint. You get version control, component reuse, and no more 2 AM slide formatting. Turn meeting notes into documentation by recording yourself explaining the system, then feed the transcript to AI with your writing style guide. Connect AI directly to Jira and Notion through Model Context Protocol. No more copy-paste between browser tabs.
Building Presentations Like Software
This presentation itself demonstrates the approach. Instead of slides, it’s a React app with Vite and TypeScript. Industrial sci-fi aesthetics, glitch transitions, typewriter animations. Arrow keys navigate, N toggles speaker notes, and it scales from phones to projectors.
Why build presentations this way? You already know React. You already use Git. Components are reusable. Animations are customizable. Most importantly, you can generate and iterate on slides with AI using the same workflow as regular code. No more exporting, importing, or fighting with slide masters.
Check out the live version at ai-at-work.raf.dev. The source code shows how AI-assisted development works in practice. Most components came from iterating with Claude Code. Describe what you want, refine the output, commit to Git.
Try This Tomorrow
Start with one non-coding task that eats your time. Documentation backlog? Build a simple template and let AI fill it in from your meeting recordings. Presentation coming up? Code it instead of using slides. Weekly status updates? Automate them with MCP connected to your project management tools.
The tools already exist. The workflows are straightforward. Pick one thing, automate it, measure the time saved. Then pick another. The compound effect beats any coding assistant improvement.
What 2025 Surveys Reveal
LeadDev surveyed 617 engineering leaders. 60% said AI hasn’t significantly boosted their teams’ productivity. Only 6% reported major gains. These are companies with GitHub Copilot enterprise licenses, Cursor subscriptions, and AI training programs.
Atlassian found something equally puzzling. Developers using AI save 10+ hours per week (up from 2024). Yet 90% still lose 6+ hours weekly to inefficiencies. Half lose over 10 hours. A full day gone, even with AI helping.
Stack Overflow’s data explains part of it. Trust in AI is collapsing. 46% of developers don’t trust AI accuracy, up from 31% last year. Only 3% highly trust AI output. The “almost right” problem frustrates 66% of developers. The code compiles, passes basic tests, then breaks in production with edge cases the AI didn’t consider.
Meanwhile, 63% of developers say leadership doesn’t understand their actual pain points. Not the coding. The meetings about meetings. The documentation nobody reads but everyone demands. The context switching between eight different tools. These problems existed before AI and persist after it.