I have a small graveyard of AI accounts.

There’s one tool I joined because it promised to organise every thought I had. Another was supposed to turn meetings into “actionable intelligence,” which is an impressive phrase for writing down who agreed to send the spreadsheet. A third could generate presentations, websites and social posts. I used it twice, forgot the password and moved on.

This is normal in 2026. AI tools are easy to try and strangely hard to keep.

My opinion is that most people don’t need a giant AI stack. They need one general assistant, one tool connected to their most annoying type of work, and perhaps one specialist tool. Anything beyond that often becomes a hobby involving subscription receipts.

The useful tool gets opened next Tuesday

A demo is a terrible way to judge AI software.

Nearly every tool looks clever when it is summarising a clean document chosen by its own marketing team. The useful test is less exciting: when Tuesday becomes messy, do you remember to open it?

I judge tools by repeated friction. If I keep searching through long email threads, that is a real problem. If I prepare a client report every Friday from the same five sources, also real. Wanting to “use more AI” is not a problem. It is a vague resolution wearing a hoodie.

Pick one task that happens every week. Give the tool three chances. If it saves time twice, keep it. If using it creates another place to copy, upload, format and check things, delete the account.

Harsh, perhaps. The industry will survive.

Start with one assistant and stop comparing screenshots

For most people, the first useful choice is a general assistant such as ChatGPT or Claude.

Use it for rough drafts, explanations, awkward emails, planning, file analysis and the first pass through a confusing problem. The important word is “first.” I would not send an AI-written proposal to a client without rewriting it.

ChatGPT’s deep research mode can work across the public web, uploaded files and connected apps, then produce a cited report. Claude’s Artifacts feature puts substantial output—documents, tools, visualisations or small apps—into a separate space where you can keep editing it. Those are different strengths, but you probably don’t need to pay for both during your first month.

Choose the one that fits your normal work. Learn its habits properly. Save a few prompts that worked. Give it examples of the output you like. Ask it to challenge your draft instead of praising everything.

Constantly changing assistants is like switching gyms because another one has newer treadmills.

Research tools are useful when you still check the research

This is the category I use most carefully.

Perplexity’s Research mode searches across multiple sources, works through a question in stages and can export the result. Gemini’s notebooks take a different approach: you can gather files, websites, Drive content and instructions inside a continuing workspace that syncs with NotebookLM.

For current comparisons, a research search tool is handy. For a project with a fixed pile of material—an annual report, interview transcripts, policy documents, product notes—I prefer the notebook approach. It keeps the work grounded in sources I chose.

The mistake is believing the neat report because it took several minutes to appear.

I once used an AI research tool to compare two software subscriptions. The answer confidently listed a feature that one product had removed months earlier. Its citation pointed to an old help page. Nothing dramatic happened; I simply wasted twenty minutes checking what I thought had already been checked.

Use these tools to reduce reading, not remove verification. Dates, prices, laws, medical claims and product limits deserve an open tab.

Meeting AI is boring. That’s why it works

The most useful AI product in some offices is the thing that remembers what everyone promised during a call.

Granola works as an AI notepad and can turn meeting context into organised notes without sending the usual visible bot into the call. Fathom focuses more on recorded meetings, summaries, action items, follow-up emails and searching across past conversations.

Try one only if meetings genuinely consume your week. Otherwise you are paying software to summarise four conversations a month, two of which could have been messages.

Also, tell people when a meeting is being transcribed. “The tool does it automatically” is a strange excuse for forgetting basic manners.

The real win is not the summary. It is asking, three weeks later, “What did we decide about the launch date?” and getting something better than a search through Slack.

Coding assistants are wonderful until you stop reading the code

Cursor and GitHub Copilot have moved beyond autocomplete. Their agent features can inspect a codebase, propose changes across files, run tests and handle larger chunks of work.

For developers, that is genuinely useful. For everyone else, an AI coding tool is not automatically the fastest way to build an app.

You still have to understand what information the app stores, who can access it, what happens when an API fails and why login behaves strangely in Safari. Generating code is easier than owning software.

If you code regularly, try Cursor or Copilot on one real project. Ask it to explain unfamiliar areas, write tests and handle small changes before giving it half the repository. Review the diff. Run the tests yourself.

A green “task completed” message is not a warranty.

The subscription graveyard needs a rule

Here is mine: an AI tool must either save thirty minutes a week or produce work I could not reasonably make without it.

Interesting does not count. Fun for one evening does not count. A viral demo of someone creating a booking app during a taxi ride definitely does not count.

Keep one assistant. Add a research workspace if your work involves sources. Add a meeting or coding tool only when that is where your time goes. Free plans are enough for testing. Pay after the habit exists, not before.

The best AI tool in 2026 may be less impressive than the one filling your feed. It might simply turn a messy Friday meeting into five accurate lines and remind you that you agreed to send the spreadsheet.

That one tends to survive until Tuesday.