Short answer
A tenant, engineer, or property manager can now ask a building plain questions — "when does my lease expire," "why is Suite 400 warm," "what's overdue on the rent roll." The answer is only useful if it's traced to a source, staged for a human where it triggers action, and logged. A multi-agent setup that skips those steps is a liability, not a feature.
What does "asking a building a question" actually mean?
Multi-agent AI — several specialized programs that split a task and hand work between them — is being wired into buildings so tenants, engineers, and managers can ask questions in plain language and get data-backed answers. One recent design, described in the AIAEC Digest, uses a "Doorman" agent to route each question to the system that holds the answer: building management for HVAC, the lease file for terms, the rent roll for balances.
The idea is sound. The risk is who trusts the answer, and what happens next.
Why is the source of the answer the whole game?
An answer to a building question is only as good as the record it came from. A number with no traceable source — the lease PDF, the rent roll row, the work-order log — is a guess in a confident voice. For CRE operators, a wrong number is not a typo; it's a misquoted expiration or a rent balance that starts an argument with a tenant.
Ground every answer in a named artifact:
- Lease terms → the abstracted lease, with the clause and page cited.
- Rent balances → the rent roll, as of a stated date.
- A warm suite → the BMS reading plus the open work order.
- Renewal windows → the lease's notice provision, not a summary of it.
If the system can't name where an answer came from, it isn't ready to give it. According to Deloitte's 2024 State of Generative AI in the Enterprise survey, trust and data-quality concerns are among the top barriers to scaling these tools — provenance is the fix, not more model.
When does a building answer need a human before it acts?
Answering is safe. Acting is where governance earns its keep. This is Vantrow's spine: propose, never commit — the system stages an action, a human approves it, and the record lands on an audit trail (a time-stamped log of who saw what and who approved it).
Reading is low-stakes:
- "When does my lease expire?" — read the lease, cite it, done.
- "What's my current balance?" — read the rent roll, show the as-of date.
Acting is not:
- Filing a maintenance ticket against a unit.
- Sending a renewal notice to a tenant.
- Adjusting a rent charge or applying a credit.
Those cross into commitment. The building can draft the ticket or the notice; a manager sends it. Autonomy that skips the approval step is how a plausible answer becomes a wrong action nobody signed off on.
Who is actually asking, and what do they each need?
Different roles ask different questions and carry different stakes. A good building-question system knows the difference and scopes what each person can see and trigger.
- Tenants ask about their space: lease dates, balances, how to report a problem. They should never see another tenant's terms.
- Engineers ask about systems: what's failing, what's overdue, equipment history. They act on tickets, not leases.
- Managers and asset teams ask across the portfolio: expirations this quarter, rent roll exposure, follow-ups nobody tracked. They approve the actions the others propose.
Same building, three permission surfaces. Multi-agent routing without role-scoped access just moves the leak around.
What should a CRE operator require before turning this on?
Treat a talking building like any tool you give access to your leases and rent roll: assume it will be wrong sometimes, and design so wrong answers can't quietly become actions.
- Provenance on every answer — a citation to the lease, rent roll, or work order.
- Read/act separation — answers flow freely; actions stage for approval.
- Role-scoped access — tenants, engineers, and managers see only their surface.
- An audit trail — every proposal and approval logged and reviewable.
- A stated "as of" — rent and occupancy answers carry a date.
Get those five right and "ask the building" becomes a real operating tool. Skip them and you've automated confident mistakes.
FAQ
Q: What is a multi-agent AI system for buildings? A: It's several specialized programs that divide a task — one routes the question, others pull from the BMS, the lease file, or the rent roll. A "router" or "Doorman" agent decides which system holds the answer. The value is accurate routing to a real source, not the conversation itself.
Q: Can a tenant get another tenant's lease information by asking? A: Not if access is role-scoped. Tenants should only reach their own lease and balance. If the system can't enforce who sees what, it isn't ready to face tenants — role-scoped permissions are a baseline requirement, not an add-on.
Q: Should the building be allowed to act on its own — file tickets, send notices? A: It should draft, not send. Under a propose-never-commit model, the system stages the ticket or renewal notice and a human approves it, with the decision logged. Autonomy on actions is where a plausible answer turns into a wrong commitment.
Q: How do I know an answer is correct? A: Require provenance. Every answer should cite the artifact it came from — the lease clause and page, the rent roll row with an as-of date, the open work order. An answer with no traceable source is a guess, however confident it sounds.
Q: Is this different from asking ChatGPT about my building? A: Yes. A general chatbot has no live access to your rent roll or leases and no audit trail. A governed building system pulls from your actual records, scopes access by role, and logs every proposed action for approval.