The most-frequent question an operations lead is asked, "how was the library used last Friday," is the question they are least often able to answer cleanly. Not because the data is missing, but because getting at the data has historically required someone else: an analyst, a SQL query, a Monday-morning report.
The Talk to your data assistant is the part of the Occivar platform that closes that loop. Plain-English question, charted answer, citation, exportable. The data is no less private for being conversationally accessible.
01 · WHAT THE ASSISTANT IS ALLOWED TO SEEWhat the assistant is allowed to see
The privacy posture of the assistant is the same posture as the rest of the platform: it can see only what the dashboard sees. That means anonymous occupancy counts by zone, dwell distributions, threshold history and the methodology metadata. It does not see, and cannot ask for, faces, demographic inferences, or anything identifying, because none of that exists in the system in the first place.
Conversation history is scoped per tenant. A query against one customer's data is invisible to every other customer. The assistant's responses cite the underlying counts: every chart it draws includes the source identifier, the window and the count totals.
02 · THE KINDS OF QUESTION IT ANSWERS CLEANLYThe kinds of question it answers cleanly
The assistant is tuned to four classes of question:
- Utilisation. How was X used last week? The answer is an hour-by-hour curve with the four-week average overlay, and a one-line summary at the top.
- Scheduling. When should we open this floor? The answer is a recommended weekly schedule built from the trailing-N-weeks demand curve, with the energy saving annualised.
- Safety. Did we cross density thresholds? The answer is a threshold-crossing log with timestamps, durations and egress estimates.
- Reporting. Generate the Q1 sustainability report. The answer is a draft PDF in the format the audit auditor expects, with the methodology cited.
A worked example of each (and how the response is rendered) lives on the Talk to your data demo page.
03 · WHAT THE ASSISTANT WON'T PRETEND TO DOWhat the assistant won't pretend to do
It will not answer questions about people. It will not infer demographics. It will not project an identity onto an anonymous count. If a question can't be answered from anonymous occupancy data, it says so and suggests a question that can.
This is not a moderation layer pasted onto a general model. It is what the model's tools have been scoped to see. There is no identity-bearing data in the conversation context, because there is no identity-bearing data anywhere in the platform.
04 · HOW THE ASSISTANT FITS THE RESTHow the assistant fits the rest
The assistant reads the same anonymous counts as occupancy analytics. It produces the same kinds of output as reports. It hands threshold rules into the alerts layer when the operator asks it to. And it routes scheduled exports through integrations: email, Slack, Snowflake, so the answer doesn't stop at the chat window.
The assistant is the conversational front-door to the rest of the platform, with the same privacy posture as everything else. If your last building-data question got answered by an analyst on Monday, the assistant is the conversation worth having.