After a Year of Using AI Daily
I’ve used AI tools — primarily ChatGPT, Claude, and Gemini — as part of my daily workflow for over a year. This isn’t an experiment or a review; it’s how I actually work. The perspective that produces is different from a product review: you see not just what the tools can do in a demo, but what they’re like on Tuesday afternoon when you’re tired, when the task is ambiguous, when the output needs to be trusted in a professional context.
The honest assessment is more nuanced than either the enthusiasts or the skeptics would have you believe. These tools are genuinely useful in ways that compound over time. They’re also unreliable in ways that matter, and the failure modes are non-obvious to newcomers.
Where AI Actually Saves Time
The highest-value use case for most knowledge workers is what I’d call “first draft generation with editing.” Writing a proposal, a summary, a document, an email — providing the AI with context and a clear goal and getting back something that’s 70% of the way there, which you then edit to completion. The time saved is real: a first draft that might take an hour to start from scratch takes fifteen minutes to produce and edit from an AI starting point.
The second high-value use case is explanation and learning. When I encounter something I don’t fully understand — a technical concept, a regulatory requirement, a statistical method — asking the AI to explain it in accessible terms, with specific examples, and then to answer follow-up questions, is significantly faster than a Google search and more interactive than reading documentation. The caveat: verify anything that matters before acting on it.
Where It Fails Reliably
Specific factual claims are where AI tools fail most consequentially. Not in an obvious “this is clearly wrong” way, but in a “this is confidently stated and plausible and subtly incorrect” way. Citations that look real but don’t exist. Statistics that are in the right ballpark but not actual figures. Names and dates and details that are consistent with what might be true but aren’t. For any professional context where facts are load-bearing, every specific claim needs independent verification.
Novel reasoning is the second failure mode. AI tools are very good at applying patterns from their training data to new situations. They’re much weaker at genuine novel reasoning — working through a problem that requires combining concepts in a genuinely new way, or catching the specific flaw in an argument that requires domain expertise to identify. The output looks like reasoning. Whether it is reliable reasoning depends on the domain and the quality of human review.
The Skill of Working With AI
Getting good results from AI tools is a skill that takes time to develop, and the people who dismiss the tools as useless have usually not developed it. The key components: providing enough context that the AI understands the specific situation, being explicit about what a good output looks like, iterating through multiple exchanges rather than expecting one prompt to produce final output, and maintaining the judgment to recognise when the output is unreliable.
The reverse error — outsourcing judgment entirely to AI — is equally common and arguably more dangerous, because it produces confident-looking outputs that the user hasn’t verified. The effective relationship with AI tools is collaborative, not deferential.
An Honest Daily Assessment
| Task | AI usefulness | Caveat |
|---|---|---|
| Writing first drafts | High | Always edit; voice and accuracy need human review |
| Explaining concepts | High | Verify specific facts before using professionally |
| Generating options/ideas | High | Quantity is high; quality of individual ideas is variable |
| Factual research | Low–Medium | Hallucination risk is real; treat as starting point only |
| Complex analysis | Medium | Good at structure; unreliable at subtle domain-specific judgment |
| Final professional outputs | Low alone | Never submit without thorough human review and verification |
Key Takeaways
- AI tools genuinely save time on first drafts, explanation, and option generation — the value is real and compounds
- Specific factual claims from AI require verification — the failure mode is confident-sounding incorrectness, not obvious errors
- Effective AI use is a skill: providing context, iterating, and maintaining judgment about output quality
- Outsourcing judgment entirely to AI is as problematic as dismissing it — the relationship needs to be collaborative
- The people who get the most value are those who use AI as a drafting and thinking partner, not a replacement for their own expertise
Sources
- Mollick, E. (2023). Co-Intelligence. Portfolio/Penguin.
- Brynjolfsson, E. et al. (2023). Generative AI at Work. NBER Working Paper.
- Marcus, G. & Davis, E. (2019). Rebooting AI. Pantheon Books.