What I Actually Use for AI Workflows Right Now
I spend most of my day juggling between different AI assistants, and by early 2026 the landscape had shifted enough that my old setup felt obsolete. The tools that kept showing up in my daily work weren't the ones getting the most press coverage either. I started tracking which ones I came back to repeatedly, and that list became something I shared with a few colleagues who asked what I was running on. That search phrase has been coming up more often in my circles lately. People want to know what actually works versus what just got good marketing. I'm going to lay out what I use, why I use it, and where I've hit walls. The biggest shift I noticed this year is that the gap between general-purpose models and specialized tools narrowed considerably. A lot of people still reach for the same three names for everything, but I found myself splitting workflows differently. Code generation, document synthesis, and research lookups each have a tool that does one of them well without mediocre-at-everything tax.
For code, I use Claude for longer reasoning tasks because it holds context better through multi-file changes. I tested a workaround last month when I hit an edge case where it would hallucinate API signatures after about 40,000 tokens of conversation. The trick is to paste the official documentation excerpt directly into the prompt instead of relying on the model's training data, and ask it to cite line numbers. That cuts false positives from about 30 percent down to under 5 percent in my experience. Gemini handles my document summarization because the 1M token window actually works for real PDFs now, not just benchmarks. I fed it a 200-page technical manual once and got a structured breakdown in about 90 seconds. The thing nobody tells you is that Gemini's free tier throttles you silently after about 60 requests per minute, so if you're processing batches at scale you need either the paid plan or a queue script that respects the rate limit without you noticing the slowdown. For research and fact-checking, Perplexity is still useful when you need sources, but I stopped using it for anything requiring citations from papers published after March 2026. Its indexing lags behind arXiv and Nature by roughly three weeks, so recent findings won't show up until the next crawl cycle. I learned this the hard way when I submitted a literature review that cited a paper which didn't exist in their database yet, and the editor flagged it during peer review.
Notion AI handles my meeting notes and project tracking. The integration with their database makes it genuinely useful for structuring information, though the autocomplete features still feel gimmicky for long-form writing. I keep it turned off for anything longer than 500 words because it tends to flatten the tone into something generic. Midjourney v7 is what I reach for when I need illustrative images for internal presentations. The quality jump from v6 was noticeable, especially around text rendering in images. But the copyright situation is murky if you're using generated assets commercially, and I haven't found a clear policy from Midjourney on ownership of outputs past 2025. I use it internally and avoid putting those images on client-facing materials. Obsidian with the AI plugin is my second brain. The local-first approach means I never worry about data leaving my machine, which matters more than most people admit. The tradeoff is that the AI features depend on connecting to an external model, so you're still sending content to a third party unless you run a local model through Ollama, which is slower but keeps everything on-device. I run both setups simultaneously and switch based on whether the content is sensitive.
Get the Full Details
There are tools I tried and dropped this year. You might have heard of some of them. One was a writing assistant that promised to match your voice, but it actually just inserted the same corporate-neutral phrasing everywhere. Another was a coding agent that claimed to write entire modules, but the output needed more manual review than writing it from scratch would have taken. I spent about four hours debugging code it generated in what should have been a thirty-minute task. The common thread across what I actually keep using is that each tool fills a specific lane without overlapping too much with the others. When I try to use one tool for everything, the quality degrades across the board. I learned this from burning two days on a project where I forced a single model to handle code, documentation, and research, and the results were mediocre in all three categories. If you're just getting started, I'd suggest picking two tools and using them well before adding more. Most people overshoot on the number of accounts they manage and end up with context-switching tax that eats more time than any tool saves. I had eight AI tools open simultaneously last spring and averaged maybe forty minutes of actual productive work per day because I kept losing track of which model I'd already asked about a topic.
The Ai Tools 2026 Favorites Google Trend search results will show you what's popular, but popularity doesn't mean fit for your specific workflow. I recommend tracking your own usage for two weeks before deciding what stays. Write down which tasks you reach for each tool, which ones you skip, and where you feel friction. That personal data is worth more than any ranking list. One more thing about cost. The free tiers have gotten tighter this year. I moved to paid plans for the three tools I use daily, which runs about sixty dollars a month total, and that pays for itself if you're doing this kind of work full-time. The real expense isn't the subscription, it's the setup and integration time, which I estimate at about six to eight hours spread across a couple of weeks before things feel smooth. I've also noticed that tool reliability varies by region. Some models process queries faster from certain cloud regions than others, and I rearranged my API routing last quarter to route through us-east-1 instead of eu-west-1, which cut average response time from about 2.3 seconds to 1.1 seconds for the same queries. It's a small detail that adds up over thousands of requests.
If you want a quick reference, here's what I reach for on a typical day: Claude for thinking-heavy tasks, Gemini for processing large documents, Perplexity for sourcing, Notion AI for organization, Midjourney for images, and Obsidian for keeping everything connected. Everything else is situational.
