Somebody on your team has already pasted lease data into a chatbot. That's not a scandal — it's what adoption looks like when the official tools lag the useful ones. The productive response is to know where the actual risk lives and put a policy around it.
What happens to a rent roll you paste into a chatbot?
It depends entirely on the tier of service. Consumer chatbot tiers may use conversations to improve models unless you opt out; business and API tiers from both OpenAI and Anthropic commit — in their published enterprise terms — not to train on your data. So the first governance question is boring but decisive: which tier is your team actually on? More than 90% of leading real estate firms call AI a strategic priority (PwC / ULI, Emerging Trends in Real Estate), yet many run that strategy through personal chatbot accounts.
What's the real risk — reading or writing?
Reading risks are manageable: they're contract terms and access questions, the same class of risk as any vendor. Writing is where the damage compounds. An ungoverned tool that updates records directly can:
- overwrite a correct rent figure with a hallucinated one, silently;
- act on a misread date across every downstream report;
- leave no trail of what changed or why.
A wrong answer you read is a mistake you can catch. A wrong answer that commits itself is corruption of the record.
What five questions should you ask before adopting an AI tool?
- Where does our data go, and on what terms? (Named tier, in writing.)
- Can the tool write to our records directly — or only propose?
- Is every suggestion reviewable next to its source?
- Who approved what, when? (Audit trail, not vibes.)
- Can we turn it off without losing our data?
Tools that pass all five are rarer than the demo circuit suggests.
What does governed AI look like in practice?
The pattern is simple to state: read → propose → approve → audit. The system reads your documents and messages, drafts the update, and stages it; a person commits it; the decision lands on a ledger. We've written about why this beats fully autonomous agents — for a rent roll, it's the difference between an assistant and an unsupervised intern with edit rights.
Trust, it turns out, isn't a feeling about the model. It's a property of the workflow around it.