Before an AI agent reads your inbox, ask these four questions
AI agent makers are racing to promise privacy, but the fine print still decides who sees your customers' data and what the agent can spend.
What happened
The Verge reports that the companies building AI agents are now competing on privacy. An AI agent is software that can take steps on its own, like reading your messages or making a purchase. At OpenAI's DevDay, CEO Sam Altman said the company wants to "set a new standard for privacy in frontier AI."
The outlet says Meta made similar promises a couple of months earlier for its agent, Muse. Mark Zuckerberg called it "built from the ground up for privacy and security." But the report says Meta can still access the data in a user's isolated virtual computer, and that Muse defaults to letting Meta train models on what you put into it (you can opt out). An Inc. reporter also complained that Muse read his private messages without his asking.
The Verge adds that OpenAI pitched its own agent, Dots, as the safer option. Altman showed a rule that stops it from buying anything over a set dollar amount. For business customers, OpenAI also described stronger data controls and a zero data retention option, meaning nothing is stored on OpenAI's servers.
Why it matters
Agents only work if they can see a lot. TechCrunch describes agents you can simply text that connect to your apps, send email, and shop online for you. Every one of those connections is a door into your data.
The Verge notes that even its own reporter, Allison Johnson, felt uncomfortable typing in her bank information when an agent asked. And a privacy promise on a launch stage isn't the same as a setting you can check.
Microsoft CEO Satya Nadella made a related point. In a post on X, The Verge writes, he argued that an authorized person should always be able to pause or shut down a model mid-task.
What it means for your business
- Clinics: Patient messages and records are sensitive. Don't connect an agent to the inbox or calendar that holds them until you know where the data goes and whether it can be used for training.
- Law firms and accounting firms: Client files carry confidentiality duties. Ask about zero data retention on business accounts, and keep client work out of consumer-tier agents.
- Retail shops: If an agent can pay for things, set a purchase cap first, like the spending rule that The Verge says Altman demonstrated for Dots.
- Agencies: Some tools may default to training on your input; The Verge reports that Muse does, with an opt-out. Look for that setting before you connect client accounts, and consider a separate email login for the agent.
One thing to try this week
Pick one agent you've been tempted to try and write down four answers from its settings and terms: What can it see? Who else can see it, including the company? Does it train on my data, and can I opt out? What spending limits can I set?
If you can't find an answer, treat that as a "no" for now. Test it on a throwaway account with fake data before it gets anywhere near your customers.