Swap the name and details and ask only general questions. Someone else’s diagnosis isn’t her data. Without the name and details she can understand the topic without exposing her friend.

What does her choice mean?

Chats may be used for training and kept up to 5 years. That’s what Anthropic’s terms say: allow training and data is kept 5 years; don’t, and it’s the usual 30 days.

Where is the best place to do this?

In a local model through Ollama on her laptop. A diary shouldn’t leave the device. A local model is good enough to spot themes and summarise.

What should she do first?

An audit: keep the one plan she actually uses. She barely opens two of the plans. Keeping one saves money and costs her nothing.

Privacy and the cost of agents

Lesson 20 of 22 in the course “AI Agents for Life”. Free, with animations and a quiz.

Where your chats go

Everything you write to an assistant is stored on the company’s servers. The key question: can it be used to train new models? It’s your call, but you need to know where the switch is. In ChatGPT it’s “Improve the model for everyone” under Data Controls: turn it off and new chats won’t go into training. In Claude, since August 2025 everyone chooses in the privacy settings. If training is allowed, data is kept for 5 years; if not, the usual 30 days. Check your assistant’s settings today: it takes two minutes.

What never to send

Even with training off, the chat sits on someone else’s server, and agents also read other people’s texts, which may contain traps. So there’s a short list of things that go to no AI. Passwords, SMS codes and API keys. Card numbers and passport scans. And other people’s secrets: a friend’s diagnosis or a screenshot of someone else’s messages isn’t your data, and you don’t have permission to share it. Need help with a text like that? Replace names, numbers and addresses with made-up ones: the meaning stays and nothing extra leaks.

AI that lives on your machine

There’s another route: run a model right on your own computer. The simplest way is Ollama, a free app for macOS, Windows and Linux. Install it, download an open model such as Gemma, Qwen or gpt-oss, and chat offline. The upside is huge: data never leaves your device and there’s no subscription. The downsides, honestly: these models are weaker than the big cloud ones, you need a powerful computer, and there’s no search or connectors out of the box. Perfect for a diary, drafts and anything you don’t want on someone else’s server.

What agents cost

You can start for free: almost everything in this course works on the free plans of ChatGPT, Claude and Gemini, and Zapier and n8n both have a free way to start. Money shows up in three places. An assistant’s paid plan for higher limits and agent modes. Automation runs: the more often a chain fires, the more it spends. And heavy modes: remember, an agent team burns about 15 times the tokens of a chat. Prices and plan names change often, so check them on the official site on the day you buy, not in a year-old review.

Does it pay off

Agents pile up easily, like subscriptions: a Telegram bot, a morning briefing, a price hunter, three paid plans just in case. Do an audit once a month. Two questions for each agent. How much time does it really save? And do you actually read what it sends? If the briefing has been going unread for ages and checking the report takes longer than doing it yourself, it’s time to turn the agent off. Check access too: an agent you’ve switched off doesn’t need connectors anymore. It’s the subscriptions lesson again, just for your team of agents.

Try it: the agent audit

First, two minutes on settings: check whether each assistant may train on your chats, and choose deliberately. Then list every agent you built during the course: brief, connectors, subscription, how often it runs. Paste the list with the brief below into a chat. The AI sorts the agents into “keep,” “trim” and “turn off” and suggests which access you can take back. You make the decisions; it just asks the right questions.

Check yourself

What should she do?
Correct answer: Swap the name and details and ask only general questions. Someone else’s diagnosis isn’t her data. Without the name and details she can understand the topic without exposing her friend.
What does her choice mean?
Correct answer: Chats may be used for training and kept up to 5 years. That’s what Anthropic’s terms say: allow training and data is kept 5 years; don’t, and it’s the usual 30 days.
Where is the best place to do this?
Correct answer: In a local model through Ollama on her laptop. A diary shouldn’t leave the device. A local model is good enough to spot themes and summarise.
What should she do first?
Correct answer: An audit: keep the one plan she actually uses. She barely opens two of the plans. Keeping one saves money and costs her nothing.

← Traps: when someone else’s text gives the orders When you don’t need an agent →