This is a Tautoru view, not a settled position.
Until recently, putting something into an AI tool posed one question: is this safe to type in? That question remains, but it is now the smallest of a set. The tools have stopped being a box. They connect to the systems you already run (email, files, calendar, document store) and increasingly act across them on your behalf. With each connection the documents you must track, and the obligations you carry, multiply.
The providers are not making this easier. Anthropic's latest changes are a fair example. On 24 June 2026 it emailed users about an updated Privacy Policy for its consumer plans. The email did not say when the change takes effect; you have to open the policy to find it is dated "Effective July 8, 2026". Nor did it say what the change is, in any usable way. It lists four headings (connected apps, identity verification, studies, marketing) and points to the full document. The one thing it states plainly is scope: consumer plans (Free, Pro and Max), not the business ones. You are told something important has moved, then left to reconstruct it yourself.
Open the policy and the substance is larger than the email suggests. For the first time it sets out what a connected, agentic tool does with your data. Claude "may send your Inputs, Outputs, and instructions to Third-Party Services to perform actions on your behalf". With your permission, some outputs "may result in actions with effects outside the Services, such as sending communications, modifying files, or interacting with third-party services". Set that against our framework. The grid tells you what is safe to put into a tool, on a plan. It cannot tell you where a connector sends it next. Every service you connect brings its own terms, privacy policy, retention and jurisdiction, and you answer across the chain, not only for the provider you began with. The same update reshapes another of these questions: its wording on when Anthropic may go to the authorities is taken up in our human-review note.
The duties you worked through on the other questions re-engage at every connection, and are easy to miss, because no one pasted anything. A connector runs both ways. It pulls in more than the line you typed: point it at a mailbox or folder and it can draw on everything there, including other people's information you never meant to expose. And it pushes out, handing your data to a third party to finish a task. Pulling a client file into the tool is a disclosure decision. So is pushing a draft out to be sent or filed. Where the material is personal information, the cross-border rule (IPP 12) can be triggered by that onward flow, even when the AI provider sits on the right side of it. A no-training plan keeps your inputs out of the model. It says nothing about where a connector sends them.
The agentic turn sharpens this. When the tool only answers, a wrong answer is a sentence you can ignore. When it acts, it sends an email, edits a document or files a form, and a wrong action is far harder to undo. So treat switching on a connector as a disclosure decision, not a convenience. Before you enable it, know what it can reach. Scope each connection to the minimum. Keep the authority to act apart from the authority to read. Keep consequential actions reversible. And read the terms at the far end of a connector, not only the AI provider's. In a connected chain, the weakest link governs your data.
Our view
This is convoluted, and getting more so. The single-box model the early advice was built on is gone. In its place is a thicket of overlapping documents and obligations, and the providers are content to leave you to untangle it, announcing change without saying when or what. The answer is not to retreat from the tools, nor to wish the complexity away. It is to treat your plan, your connectors and their limits as part of basic competence, and to demand of the provider, and of yourself, a clear account of what each tool can see and do. The question is no longer "can I put this in?" It is "what can this reach, what can it do, and whose rules govern it once it is there?" How that applies to your systems and your matter is your judgment, on your facts.