Five ways to use less AI

Two questions keep arriving in the same week. One is from owners who want to get their arms around AI for governance reasons: they've realised nobody can list, on a single page, where AI is running in their business or who signed off on it. The other is from owners who've read about the energy cost of model calls and want to do less damage. Different motivations, same work. The tools that cut your policy exposure also cut your compute bill and your carbon footprint. Here are five ordinary things to do this quarter.

1. Audit what you already pay for

Most AI didn't arrive as a purchase. It arrived as a default, switched on inside a tool you already owned. Microsoft 365 Copilot in the sidebar of Word. Gemini features in Google Workspace. Notion AI. Generative fill in Adobe. Slack AI summaries. Before buying another AI product, spend a morning listing the AI features that already shipped with the software you pay for. The quickest way to use less is to notice what's already running.

2. Turn it off where you don't need it

Google's web search will stop running the generated summary at the top of the page if you append -ai to the query. Copilot has a settings panel that will disable the inline suggestions and the sidebar. Most SaaS tools (software you rent over the internet, like Slack or Notion) have an "AI features" toggle buried two menus deep in admin. It takes an afternoon to go through them. You will not miss most of what you switch off. Leave a note in the admin settings so a future colleague knows why.

3. Check the training setting on what you keep

The tools you decide to keep using need one more pass. Are they training their models on what your staff type into them? Most vendors default to yes and put the opt-out a few clicks deep. This is the setting that turns convenience into a slow IP leak. It costs nothing to change, and nobody will tell you to.

4. Ask whether a non-AI version would do

For a surprising number of small jobs, a rules-based tool is more accurate, cheaper, and doesn't send your data anywhere. A deterministic spell-check (one that follows fixed rules and gives the same answer every time) catches typos more reliably than a rewrite suggestion from a model. Programmatic document comparison catches changes in a contract more reliably than an AI diff. A calendar invite sent by a human beats a calendar invite drafted by a model and then edited by the human anyway. The question to hold in your head is not "could AI do this?" but "is the non-AI version good enough, and how much does the AI version actually add?" Often: not much.

5. Keep a register

One shared document. Which AI tools are in use, which team uses them, who approved them, and when they were last reviewed. Shadow AI (the tools staff sign up for on a company card without telling anyone) survives in silence. Once a tool has a line in the register, someone owns it, and the review date means someone has to look at it again. Most of what you find won't need removing. It just needs to be visible. A clipboard is fine. A spreadsheet is fine. Do not let anyone sell you a platform to do this.

If you do one thing this quarter

Make the register. Everything else follows from having the list. Here is the short version of the work above, compressed into a morning:

• List every AI feature already bundled in what you pay for.
• Switch off what you don't use.
• Opt out of training on inputs for what you keep.
• Ask whether a non-AI version would do the job.
• Put it all in one shared document with a review date.

None of this is anti-AI. It's about knowing where AI is, and only keeping it where it earns its keep. The business won't feel smaller. The surface area you're responsible for will.