I've tried a lot of AI tools.

Some became essential. Many didn't. A few were actively counterproductive—tools that promised productivity gains but delivered context-switching overhead and mediocre results.

Here's what I stopped using and why. Maybe it saves you the same experimentation.

The "AI-Powered" Note-Taking Apps

I tried several: Notion AI, Mem, Reflect, and others promising to "surface insights" from my notes.

Why I stopped:

The AI features felt bolted on. Summarizing notes I just wrote? I know what they say—I wrote them. "Surfacing connections"? Usually obvious or irrelevant.

The core problem: good note-taking is about thinking while writing, not having AI think for you afterward.

What I use instead: Plain markdown. My brain does the connecting.

AI Meeting Assistants

Tools that join your calls, transcribe, summarize, extract action items.

Why I stopped:

Privacy concerns. Asking everyone in a meeting to consent to AI recording changed the dynamic. Some people clammed up.

The summaries were... fine. But I could take notes faster than fixing AI-generated summaries that missed context.

The real issue: most meetings shouldn't happen. AI making bad meetings slightly more efficient doesn't solve the actual problem.

What I use instead: Fewer meetings. When I do meet, I take manual notes for what actually matters.

Code Review Bots

Automated PR reviewers that comment on every pull request.

Why I stopped:

Noise. So much noise. The bot would flag style issues I didn't care about, miss actual problems, and clutter every PR with generic comments.

When everything is flagged, nothing is flagged.

What I use instead: Claude for focused code review when I need it. On my terms, for specific concerns.

AI Writing Assistants (for Code Comments)

Tools that auto-generate documentation and comments.

Why I stopped:

The generated comments described what the code did, not why. "Loops through array" is obvious from reading for item in array. The useful comment would explain why we're looping, what we're looking for, what happens next.

AI can't know the why unless you tell it—at which point, you might as well write the comment.

What I use instead: Writing documentation with AI, but for real docs—READMEs, API documentation—not inline comments.

General-Purpose AI Agents

Autonomous agents that supposedly complete tasks end-to-end.

Why I stopped:

They don't work reliably. Great demos, frustrating reality. The agent would confidently do the wrong thing, requiring more cleanup than doing it manually.

Maybe in a year or two. Not yet.

What I use instead: AI as a tool I direct, not an agent I delegate to.

Specialized Coding AIs

Various tools claiming to be "better than Copilot for X language" or "built specifically for Y framework."

Why I stopped:

Context switching. Learning another tool's quirks. Managing another subscription. The marginal improvement over general-purpose tools wasn't worth the overhead.

What I use instead: Claude, ChatGPT, and Copilot. Three tools that cover everything well enough.

The Pattern

Looking back, the tools that failed shared characteristics:

Automation without understanding. They did things automatically that needed human judgment.

Solutions to non-problems. Optimizing tasks that shouldn't exist or adding AI to workflows that worked fine.

Demo-ware. Impressive showcases that didn't survive real usage.

Overhead exceeding value. The time spent configuring and correcting exceeded time saved.

What Actually Stuck

The tools that survived the cull:

  • Claude for thinking and complex generation
  • ChatGPT for quick lookups and multimodal
  • Copilot for inline completions
  • Local models for private work

That's it. Four tools. Everything else was noise.

Before You Subscribe

Questions to ask:

What specific problem does this solve? Not vague productivity—specific tasks.

How does it compare to tools I already use? Is it better enough to justify the switch?

What's the overhead? Setup, learning curve, context switching, subscription management.

Can I trial it on real work? Not toy examples—actual projects.

Most tools fail these questions. The few that pass are worth paying for.

Related Reading