Own your data in the AI era: ownership means little without control
Aperisa ยท 7 min read
A house you own, with locks you can't change, is owned in name only. Every key ever cut still opens the door. That is roughly where most people stand with their data, and AI agents just started copying keys.
To own your data in the AI era, you need control, not a title. Control means three things: one record, kept in one place, that you decide about; access you grant agent by agent; and access you can take back. Our view: ownership you can't act on is not worth much, and the agent era is when that gap starts to cost something.
Checked on 2026-10-29.
What does owning your data mean when agents read it?
On paper, you own your email, your documents and your bank history. In practice, copies of them sit with every service that touched them, and you have little say over who reads those copies or for how long.
That was tolerable when the readers were mostly people and mostly slow. Agents change the reader. A coding agent or an assistant on Telegram reads your data on demand, at machine speed, and can call tools based on what it read. If you can't say which agent sees what, "owning" your data is mostly a word.
So the useful definition is practical. You own your data when you can get a copy of it, including what has been built from it, and decide who reads it.
Ownership without control
Our beliefs page makes this argument in plain terms. In a thin, technical sense you already own your data. What the market withholds is control: the ability to say who can read it, for how long, and for what. The page compares it to a house whose locks you aren't allowed to change.
The gap shows up in familiar places. Your details spread across many businesses, and inside each one, anyone with access can read them. Opt-outs differ from service to service. Privacy laws exist, but they mostly react after something has gone wrong, and they were not written to be used day to day by the people they protect.
None of that is new. What is new is who the readers are.
Do you own what an app builds from your data?
Owning your data shouldn't stop at the raw files. When an app reads your email and documents, it builds something new from them: a list of your bills, the renewal dates it found, a running total of what you paid a vendor. Often it mixes in other data too, like categories or exchange rates.
That new layer is where most of the value is, and it is usually the part you can't take with you. Export your email and you get your email back. The organized version, the part that took the work, stays inside the app.
A fair rule: if it was made from your data, for you, you should be able to take it with you in a form another tool can read. It shouldn't be locked in one app's walled garden.
So when you pick a tool for your data, ask what its export includes: only the originals, or also the records and answers it built from them.
Why does the agent era raise the stakes for data ownership?
Because agents read across sources and act on what they read. A person who opens your inbox reads some emails. An agent with inbox access can search years of mail, join it with your calendar and your files, and pass the result to another tool.
Simon Willison named one version of this risk the lethal trifecta in June 2025. His three legs: "Access to your private data", "Exposure to untrusted content", and "The ability to externally communicate in a way that could be used to steal your data". Combine all three in one agent and an attacker can trick it into sending your data out. The first leg is the one this post is about. It is the one you decide when you connect your data.
The Model Context Protocol, an open protocol agents use to reach outside tools and data, is clear about who should be in charge. Its specification says "Users must retain control over what data is shared and what actions are taken." It also says "MCP itself cannot enforce these security principles at the protocol level." The protocol states the principle. Your setup has to enforce it. (New to MCP? Start with what an MCP server is.)
There is one honest limit you can't engineer away
Whatever an agent reads also goes to that agent's model provider. The agent runs on a model, the model needs the text in its prompt, and the prompt goes to whoever runs that model. Connect your bank history to an agent and the transactions it pulls travel with your question to that provider.
Revoking access later stops future reads. It doesn't recall what was already read. Our privacy policy says the same about authorizations: revoking one prevents future access, and information already shared before that is governed by the third party's terms.
So control does not start at revocation. It starts with choosing what each agent gets in the first place. A revoke button is the backstop.
What does control over your data look like in practice?
Three habits, none tied to a product.
One record, in one place, that you decide about
Today your facts sit in Gmail, Drive, a calendar, a bank portal and a dozen apps. Each connection you give an agent is another set of credentials, and each has its own idea of what "access" covers. When the record lives in one place, you set access rules once, and you can see in one view what any agent can reach.
One place does not mean you hold it on your laptop. It means one service you chose, with access rules you set, and a copy you can export when you want to leave.
Access granted agent by agent
Give each agent its own credential, scoped to its job. A bookkeeping agent needs transactions, not your personal email. A scheduling agent needs your calendar, not your contracts. Some jobs need far less detail than the full record: an agent that reports on spending can work from totals instead of individual transactions.
Separate credentials also tell you which agent read what. One shared credential across five agents tells you nothing.
Access you can take back
Credentials leak, agents get retired, and side projects get abandoned with tokens still live. You should be able to cut off one agent without touching the others, and the cutoff should take effect quickly. Ask any service you connect: how long after I revoke a key does it stop working?
For the same pattern across several agents sharing one record, see a context layer for AI agents.
Where does Aperisa fit?
Give each agent its own key and set what it can reach: which data areas, work or personal, and how much detail, from full records down to totals only. Aperisa enforces those limits on the server. Revoke an agent key and it stops working within 2 minutes.
Sensitive data, including document text, email bodies, contacts, calendar details, financial accounts and transactions, identity numbers and chat history, is encrypted at rest with a key unique to you. Aperisa operates the key service, so this is not zero-knowledge. Export a copy of your data, or delete your account from Settings at any time. Aperisa's export includes the records it builds from your documents, such as your bills with their due dates and amounts, as a file other tools can read. Identifying details like account numbers stay encrypted.
Ownership is a starting point
Owning your data matters less than deciding who reads it. Agents make that decision more frequent and more costly to get wrong, and once an agent has read something, its model provider has it too. Decide what each agent gets before it connects, and make sure you can cut any agent off later.
Aperisa connects as a remote MCP server over Streamable HTTP with a Bearer key, and each key can be limited by data area and level of detail. Try it out.
FAQ
What does it mean to own your data in the AI era? In practice it means two things: you can get a copy of your data, and you decide who reads it. Agents make the second part matter more, because they read across sources and act on what they find. Ownership you can't act on doesn't change who reads your data.
If I revoke an agent's access, is the data it already read gone? No. Revoking stops future reads. What the agent already read has been sent to its model and may sit in its context, logs or memory, and its model provider's terms govern what happens to it there. That's why the choice that matters most is what each agent gets in the first place.
Should every AI agent have its own credential? Yes, as a general rule for any setup. A separate credential per agent lets you limit each one to the data its job needs, see which agent did what, and cut one off without breaking the others.
Is my data private once it is encrypted at rest? Encryption at rest protects stored data. It doesn't decide who can read that data through a working access key, and it doesn't keep it from whoever operates the key service. Ask who runs the encryption keys, and what each agent's access lets it read.
Sources
- Aperisa: Beliefs
- Aperisa: Privacy Policy
- Model Context Protocol specification, 2025-11-25: Security and Trust & Safety
- Simon Willison: The lethal trifecta for AI agents (June 16, 2025)
Questions
What does it mean to own your data in the AI era?
In practice it means two things: you can get a copy of your data, and you decide who reads it. Agents make the second part matter more, because they read across sources and act on what they find. Ownership you can't act on doesn't change who reads your data.
If I revoke an agent's access, is the data it already read gone?
No. Revoking stops future reads. What the agent already read has been sent to its model and may sit in its context, logs or memory, and its model provider's terms govern what happens to it there. That's why the choice that matters most is what each agent gets in the first place.
Should every AI agent have its own credential?
Yes, as a general rule for any setup. A separate credential per agent lets you limit each one to the data its job needs, see which agent did what, and cut one off without breaking the others.
Is my data private once it is encrypted at rest?
Encryption at rest protects stored data. It doesn't decide who can read that data through a working access key, and it doesn't keep it from whoever operates the key service. Ask who runs the encryption keys, and what each agent's access lets it read.
Published
Keep reading
- Why your AI agent keeps acting on a note that was never true
An agent can act on a memory note that was never true or has gone stale. Why it happens, how to fix stale agent memory, and where Aperisa fits.
- The lethal trifecta in AI agents: reduce what can leak
Simon Willison's lethal trifecta: private data, untrusted content and a way out. You can rarely cut the last two, but you can shrink what each agent can read.
- How to limit AI agent data access: control what your agents can see
Limit AI agent data access below the OAuth scope: split by data area, work or personal, and level of detail, then give each agent its own key you can revoke.