Architecture Studio
AI observability for architecture firms
Your firm can see some of what an AI assistant does and not the rest. This is what Architecture Studio records, what it can't show, and where review still belongs to your people.
Seeing that an assistant did something is not the same as knowing the work is right. Observability tells you what happened. Your people still decide whether it was good enough to issue.
When an assistant helps with project work, a principal reasonably asks: what did it actually do, and how would I know? Observability is the practice of recording enough about a system's activity that you can answer that question after the fact.
Add to Claude
Add to ChatGPTThis page explains what Arch Studio records today, what it cannot see, and why a record of activity is not a substitute for review. It is written for the people who answer for the firm's work.
Two different questions
Observability asks: what happened? Which tool ran, when, and whether it finished.
Evaluation asks: was it any good? Whether the answer was right, complete and fit for the project.
A record can answer the first question without touching the second. A search that ran and returned results shows that a search ran. It does not show that the search found the right products, or that it missed none. Treat any activity record as evidence of activity, and nothing more.
What Arch Studio records today
The hosted connection records usage data by default. It notes which Arch Studio tools, skills and references are used, and when. Each event is linked to your account by a pseudonymous account ID and to your firm by its organization ID, so it isn't anonymous.
It never includes your prompts, conversations, tool inputs, project files, generated outputs or access tokens.
How long it is kept:
- Raw usage events are deleted after 30 days.
- Account activity milestones are deleted after 90 days without activity.
- Aggregate report snapshots are deleted after 397 days.
You stay in control of it. Tell your assistant “turn off usage data” while Arch Studio is connected, and a firm admin can say “turn off usage data for my firm”. While it is off, no usage events are recorded. In the open-source plugin, set AS_TELEMETRY_DISABLED=1. Because the data is linked to an account, it can be deleted: delete your Arch Studio account, or email help@architecturestudio.ai.
Collection is best effort. A missing event means coverage is unknown, not that nothing happened.
Today this data supports the people who run Arch Studio. It does not yet come with a report your firm can open.
What a record can't tell you
Five limits apply to any AI activity log, ours included. Knowing them keeps a log from being mistaken for assurance.
A connector only sees what reaches it
Arch Studio sees the inputs a skill explicitly sends to a hosted tool, and the usage events described above. It doesn't see your conversation, the files you attach, or what you do in your drawing and documentation tools. If activity never reaches the connector, the connector has no record of it. That is a design choice, and it is why your prompts and project files stay with your provider.
Finding a skill or source is not following it
A record can show that a skill or reference was retrieved. It cannot show that the assistant applied it correctly. Delivery of instructions or resource pages does not prove that an architectural task was completed. Whether the work follows your firm's rules is checked in review, against your acceptance criteria.
A tool-call record is not a design decision
Knowing the assistant called a tool says nothing about what an architect then used, changed, rejected or issued. The record ends where the human judgment begins. Decisions that matter are kept as dated approvals by named people, as described in AI governance.
A full chat history still leaves most of the work unseen
Even a complete transcript of every conversation records what was said to the assistant. The drawing, the specification and the model are produced afterwards, in other tools, by people. A transcript doesn't observe that work. It shows how a question was asked, not how the answer was used.
Research and issued work are not linked automatically
It is tempting to assume that research done with an assistant can be traced to the drawings and specifications that were issued. Nothing connects them on its own. A firm that wants to know whether an issued package reflects the research behind it has to review the package itself. Observability can point you toward where to look. The package is what you check.
Sign-off stays where it matters
More logging does not mean more approvals. Asking someone to sign off on every AI interaction would bury the checks that count and doesn't match how a practice works. Put review where it already belongs: at the deliverable. Each one has a named preparer, reviewer and issuer, and the assistant prepares work that your people review and release.
Use records to make that review better informed. Use them to see which sources were consulted and where to look first. Don't use them to replace the reviewer.
FAQ
Can Arch Studio show me what the assistant did on a project?
Not in the form of a firm report today. Arch Studio records which tools, skills and references are used and when, for the people who operate the service. It does not record your conversation or the content of your work.
Does usage data include my prompts or project files?
No. It never includes prompts, conversations, tool inputs, project files, generated outputs or access tokens. It is linked to an account and a firm, so it isn't anonymous.
Can we turn usage data off?
Yes. Say “turn off usage data” to your assistant while Arch Studio is connected. A firm admin can say “turn off usage data for my firm”. In the open-source plugin, set AS_TELEMETRY_DISABLED=1.
Does a log prove the assistant followed our standards?
No. A log can show a standard was retrieved. Whether the work meets it is a matter for review.
Do we need sign-off on every AI interaction?
No. Review the deliverable, and keep named responsibilities and dated approvals for the decisions that carry professional responsibility.
How long is usage data kept?
Raw usage events for 30 days, account activity milestones for 90 days without activity, and aggregate report snapshots for 397 days. Retention and deletion →
Start with one question.
Connect Arch Studio to the Claude or ChatGPT account your firm already uses, then decide where in your own process a person checks the work.
Read AI governance → · Privacy and data handling → · Questions? Contact us →