Which AI Tool, Actually

Your card statement has four AI tools on it

Open last month’s card statement. There’s the big assistant you use every day. There’s a second one you signed up for because someone on LinkedIn said it was better at writing. There’s a transcription tool that emails you notes from meetings you were already in. And there’s the free trial you meant to cancel.

Three of those four probably do a job you could already do somewhere else. I don’t think that’s carelessness. It’s that we tend to buy by tool name, and tools don’t map neatly onto jobs, so the collection grows and the overlaps stay hidden.

So I’d flip it. Start with the job, then look at the tools. Below are eight jobs that come up again and again in small businesses, what to use for each, and where the doubling-up usually sits.

There are no affiliate links on this page and nothing here is sponsored.

Sort by the job and the overlaps show up

Thinking and drafting with your context. This is home base, and it’s where most of the value is. The big general assistants (ChatGPT, Claude, Gemini, Microsoft Copilot) are close enough at everyday drafting that your context matters more than which one you choose. So pick one, load your voice file and context files into it, and stop comparing. Most tools now let you save standing instructions somewhere, so check where yours keeps them. A model that knows your business will beat a slightly stronger one that doesn’t. If you really can’t decide, run your own writing through two of them for a week and keep the one you argue with less.

Long documents. Contracts, tenders, a year of board papers, a 90-page transcript. What matters here is how much the tool can hold at once, and whether it’s working from the actual file or from a summary of it. I wouldn’t trust any of them until you’ve tested it: drop in your longest document, ask for a fact from page 60, and check the answer against the page. Tools differ a lot more here than they do at writing an email.

Images. Three different needs tend to get muddled together. For art-directed images where the look really matters, a dedicated generator like Midjourney. For a quick image in the middle of a conversation, the one built into your assistant, if it has one. And for images that sit in a layout with your fonts and brand colours, a design tool like Canva or the Adobe apps. Check the licence terms for commercial use before anything ends up on a client invoice.

Transcription and meetings. Dedicated note-takers like Otter, Fireflies, Fathom and Granola, and the recording built into your video platform if your plan includes it, will all give you a transcript you can work from. Once the transcript is usable, two other things matter more. How it captures the call: a bot that joins as a participant, or an app running on your own computer. And whether the summary lands somewhere you’ll actually see it.

Whichever you pick, everyone on the call needs to know they’re being recorded. I’m not a lawyer, and recording rules differ between Australian states and territories. Some require everyone’s consent. So get consent before you record: say so at the start, give people an easy way to say no, and check the rules that apply to you and the people you’re recording.

Automation and connecting things. Zapier, Make and n8n all do this, and the automation guide covers how to choose between them. But before you reach for any of the three, check whether the tools you already pay for can do it natively. It’s surprisingly common for a paid automation to be re-creating a feature your CRM already has.

Research. Perplexity, and the research modes most of the big assistants now have, do the same basic thing. They search, read and summarise, with links to their sources. The source is the useful bit, not the summary. Open it before you repeat the claim in a client meeting.

Coding and building. For software you’ll need to maintain, developer tools like Claude Code, Cursor or GitHub Copilot. For a prototype you want to click through by Friday, the browser-based builders. The catch is that anything you build this way still has to be hosted, updated and kept secure by someone, and that someone is you.

Scheduling and social. Buffer, Later, Metricool, Hootsuite. Choose based on which platforms they post to and whether the calendar view suits the way you plan. I wouldn’t choose based on the AI caption feature. A caption tool that hasn’t been given your voice will write generic captions whichever scheduler it lives in, so write them in your assistant with your voice file loaded, then paste them in.

Most of the waste is two tools doing one job

Job	Use one of	Then you don’t need
Thinking and drafting	One general assistant	A second general assistant, any “AI writer” tool
Long documents	The assistant that holds files well	A separate PDF-chat tool
Images	A generator plus your design tool	Standalone “AI headline image” apps
Meetings	One note-taker	A second note-taker, an AI summariser for the transcript
Automation	Zapier, Make or n8n	The other two
Research	Your assistant’s research mode, or Perplexity	Paying for both
Social	One scheduler	A paid AI add-on on top of it

The most common duplicate is two general assistants. The second one is nearly always bought in a moment of frustration with the first, and when you look closely, the frustration was usually a missing context file, not a weaker tool.

The next most common is an AI writing tool that sits on top of a model you already pay for, and adds little more than a menu of templates.

Any one of these is reason enough to cancel

• You haven’t opened it in 30 days and can’t name the thing it does that your main assistant can’t.
• It duplicates a feature in a plan you already pay for. Check your CRM, your design tool and your email platform first.
• It’s a wrapper around a model you already have access to, with no data of its own.
• You’re paying for seats nobody uses. Count who actually logged in this month and, where the plan lets you, pay for that number.
• It was set up by someone who no longer works with you, and nobody has the login.

When not to bother

Don’t switch tools to fix a process problem. If the brief is vague, the output will be vague, and a new subscription just gives you a fresh vague output in a different font.

Don’t chase every release, either. Something new will be announced next month. Put two review dates in the calendar each year and let the noise go by in between.

Don’t consolidate in the middle of a launch or at end of financial year. Moving everyone’s tools during the busiest fortnight is how you end up with a shadow spreadsheet.

And I wouldn’t cancel the thing one person has built a real working habit around just because it’s technically a duplicate. Retraining them will cost more than the subscription.

What you’ve got now

Every AI subscription you pay for, sitting next to the job it does. The ones doing the same job, each with a reason to cancel you could say out loud. And a clearer line between a tool problem and a process problem, which is where most of the money tends to go.

Get that card statement out again. Write the job next to every AI charge, and cancel the first one you can’t fill in.