Before AI reads the next property email
The build has moved a little this week. We have continued with the foundations for version 4 and published a new article about why property work gives AI something to build on. The shared inbox is still ahead of us.
This week’s signals concern AI doing work across tools, the information it reaches and what happens to that information along the way. A property email is a useful way to read all three: it can contain a task, a building’s history and details that should only reach certain people.
FROM THE LAB
On Friday, we recorded decisions about the next workspace and tested a small foundation change on staging, our separate test copy. Nothing from that session went into production. The existing hardened core remains in use, and the shared inbox is next. We are keeping the changes small enough to check before the next layer depends on them. There is no new live workflow to present this week.
SIGNAL
From answering questions to preparing work
What happened. In a report published on 14 October, OpenAI says Codex and ChatGPT Work account for more than half of output tokens in its EU enterprise sample. Tokens measure model-generated text here. The finding concerns OpenAI customers and usage, rather than the share of European businesses using agents or the productivity they gained.
For an operator. The useful possibility is an assistant that helps prepare a piece of work across records and tools. For example, a request to arrange an inspection might become a draft brief with the relevant history attached. That is a possible use, not a result demonstrated by this report. The benefit depends on whether the brief is accurate and saves work after checking.
Our position. We would judge the finished handover and the effort needed to trust it. More generated text can mean more material to review. A proposed next action, its source and a person responsible for approving it give us a more useful basis for deciding whether an assistant helps.
European storage does not settle where AI processes the message
What happened. On 12 October, Lovable announced EU inference for Enterprise teams: supported model requests can be routed to designated European endpoints. Inference is the processing that produces an AI answer. The setting covers supported requests while building and through Lovable’s built-in AI service in published apps; direct calls to other providers are excluded.
For an operator. A fault report can be stored in a European database while its contents are processed somewhere else. Storage and model processing need separate answers. Lovable also says its web tools still use US services, and some model endpoints can be in European countries outside the EU. The feature’s name alone is not enough to establish the path a particular request takes.
Our position. When assessing a proposed AI connection, we would want to know which model processes a property message and whether any additional tools receive its contents. That gives us a concrete basis for deciding what information belongs in the request.
A connected account comes with a particular set of terms
What happened. On 16 October, Google added a clarification to an older Workspace post. It says optional connections to other Google services can use selected Workspace content for personalisation and, depending on settings, model training. The linked Search help page describes personal Google accounts and the use of summaries, excerpts and inferences from connected content. This is not a blanket change to business Workspace terms.
For an operator. Using AI inside a business service and connecting a personal inbox to another service are different arrangements, even when the products share a brand. If property correspondence passes through a personal account, the relevant question is which connection was enabled and what that service is permitted to do with the material.
Our position. A permission to read an email needs more context than a familiar sign-in screen. We would check the account, receiving service and stated uses together. This is a reason to read the terms of the specific connection, without assuming that every Google AI feature has the same data treatment.
IN THE JOURNAL
Article / 004: Why property work gives AI something to build on. Recurring work, lasting assets and a history that becomes useful again give an operator somewhere to begin. The article follows a letting workflow into the decisions and information behind it, then draws on our own document and maintenance work. It is written for someone who knows the properties, whether or not they ever build software.
The first letter is also in the archive, with the background to our build and the database test that changed the guarantees we could claim.
TRY THIS
Take one AI connection you already use. Write down which account supplies the information, which service receives it and whether it can act or only prepare a draft. Check the provider’s description against your answers. An unknown belongs on the list as an unknown; there is no need to connect another account to find it.
Reply and tell us what you run and what breaks most often. We read every answer.
RENAVID · renavid.com
Free. Sundays.