In short
When several AI agents read and change your data, you need to know which one did what. Here's what's happening (Notion, the AEPD) and what a good activity log should tell you.
In this article
In short
With a single assistant, knowing who changed a piece of data is easy. Once you connect several AI agents to your documents, your CRM or your inbox, it isn't. That's why tools are starting to show which agent made each change, and why the Spanish Data Protection Agency (AEPD) asks for a history of what an agent does on your behalf. Before giving an agent access, make sure you'll be able to see what it read, what it changed and when, and undo it.
Why does it matter now which agent made each change?
Because agents no longer just answer: they write. An MCP client (the way assistants like ChatGPT, Claude or Grok connect to other apps) can create a page, move a record or rewrite a note. If your team has three people and four agents working on the same data, "somebody changed it" stops being an answer.
On 29 October 2026, Notion announced that its "Last edited by" label now shows which agent, MCP client or automation made a change, including outside agents connected through MCP. Notion pitched it as a bit more accountability for agent edits. Microsoft, for its part, records in its Purview audit log when an agent is invoked and when it runs a tool.
What does the AEPD say?
In its ten-point guide to AI assistants and agents, published on 27 October, the AEPD recommends choosing an assistant that lets you check what it accessed, when, and what it changed in each task, and undo any changes you don't agree with. It also advises reviewing the activity history and checking in the affected apps what was sent or modified.
And among the things whoever provides the agent should offer, it lists "a history of what the agent has done on the user's behalf" and an immediate way to stop it. We covered the guide in what your AI agent can see, remember and do.
What should a good activity log tell you?
If a tool says it keeps an activity log, check that it answers these six questions:
- Who did it. Which specific agent or client acted, not just "the AI".
- On whose behalf. Which person started it or approved it.
- What it touched. The exact record, document or field.
- What it did. Reading, creating, editing and deleting don't carry the same weight.
- When. The date and time of every action.
- Whether it can be undone. Going back to the previous version without relying on someone's memory.
It's worth logging blocked attempts too: an agent trying to touch something it shouldn't is a warning sign, even if it fails.
What if my data is spread across ten tools?
That's the real problem. Each app has its own history, or none at all, and working out what happened on a Tuesday afternoon turns into opening ten tabs. The more copies of your data sit in different tools, the harder it is to know who saw what.
That's why at Traiker we start from a different idea: one secure place where you keep your data and connect your agents, instead of scattering it. Every person or company gets its own encrypted database, we don't train models on what you store, and Traiker is designed so your assistants reach it through MCP only when you connect them. If you're aiming to become an AI First company, start by knowing where your data goes.
Frequently asked questions
Is it enough to ask the agent what it did?
No. The agent can get it wrong or summarise badly. The log has to be kept by the tool where the data lives, not by the conversation.
Does an activity log replace permissions?
No. The log tells you what happened; permissions decide what can happen. Ideally you give the minimum access and can still review everything afterwards.
What should I do if I find a change I didn't approve?
Stop the agent, revoke its access, undo the change if you can, and check where you keep your MCP credentials.
If you want your data in one secure place with your agents connected to it, sign up for Traiker.
