A year of working alongside AI tools: an honest assessment
It's been roughly a year since ChatGPT launched and I started actually integrating AI tools into daily work. Time for an honest review of what's changed and what hasn't.
November 2022 to November 2023. One year since ChatGPT launched and started a conversation that hasn't stopped.
I've been deliberately using AI tools throughout that year, not as a curiosity but as working tools integrated into my daily work. Here's what actually changed.
What genuinely changed
Writing is faster. Not just the mechanics of writing, but the cognitive resistance to starting. I now begin almost every significant written output (proposals, documentation, emails requiring careful wording) with an AI-assisted draft. I rewrite substantially. The quality of my output hasn't deteriorated; I think it's marginally improved because I spend more time editing and less time staring at blank documents.
Scripting is different. I write PowerShell, a bit of Python, some Graph API calls. My approach to this work changed. I now start with a described intent, get a working scaffold from ChatGPT or Copilot, then review, adapt, and test. The time to working first draft has halved for routine automation tasks. The verification step has not gone away; it's more important than ever.
Research is faster. For high-level summarisation and orientation on unfamiliar topics, AI tools have substantially replaced my starting-point Google search. For precise technical verification, I still go to primary sources.
What hasn't changed
Technical judgment. The hardest parts of my job (deciding on architecture, assessing risk, reading a room, knowing which battles to fight), AI doesn't touch these. It might help me think through options, but the judgment is mine.
Relationship-based work. Conversations with customers, internal politics, managing expectations, delivering difficult news. These are human problems.
Novel problem solving. When I'm genuinely in new territory (a weird edge case, an undocumented behaviour, something that doesn't have a clear prior art), AI tools are less helpful. They pattern-match on training data. Truly novel problems don't have training data.
The tools I actually use
- ChatGPT Plus (GPT-4): Daily driver for writing tasks, code scaffolding, and research
- GitHub Copilot: Integrated into VS Code, useful for scripting
- Bing Chat: When I need current information with citations
- Claude (Anthropic): Occasionally, for comparison. The Claude 1 model showed up earlier this year and has a noticeably different feel to GPT-4
What I've learned about integrating this well
- Be specific in your prompts. Vague questions get vague answers.
- Treat all factual claims as unverified until you check them.
- Use these tools to augment the slow parts of your workflow, not replace your judgment.
- Stay current with what's changing. The landscape six months ago is different from today.
Where I think this is going
The progression from GPT-3.5 to GPT-4 to GPT-4 Turbo (which OpenAI just announced at their Dev Day earlier this month) shows that models are improving faster than most people expected. Microsoft 365 Copilot went GA this month, after a long preview period, and is now available to enterprises willing to pay the licensing premium.
We're past the "is this real" phase. We're in the "how do we actually use this well" phase. That's a more interesting and more productive question.
My recommendation after a year: if you haven't built a genuine working relationship with these tools yet, start now. Not because you'll be left behind if you don't (though possibly), but because the tools are good enough now that you're leaving value on the table by waiting.
