Property AI
What responsible AI in property management should—and should not—do
Useful property AI should make an operating record easier to understand while keeping permissions, evidence and consequential actions under human control.
Direct answer
The short version
Responsible property AI should work from records the signed-in user is permitted to access, distinguish facts from suggestions, show the context behind important answers, and require clear confirmation before consequential action. It should not invent portfolio facts, bypass role boundaries, move money, send messages or change records merely because a user asked an ambiguous question.
1. Start with bounded, reviewable use cases
The best first use cases reduce the time needed to understand the operating record without transferring responsibility to the model.
- Summarise open maintenance by priority, age or property.
- Explain recorded portfolio movements and identify information that is missing.
- Prepare a draft owner or tenant response for human review.
- Find relevant property, lease, document or issue context the user could already access.
- Create a proposed checklist or next-action plan without executing it.
2. Apply five controls before trusting the answer
| Control | Question to ask | Failure to avoid |
|---|---|---|
| Identity | Which signed-in user is making the request? | A generic public assistant receives private portfolio context. |
| Permission | Which exact records may this user access? | The model sees another owner, organisation or portfolio. |
| Grounding | Which records support the answer? | A fluent response invents amounts, dates, people or issue status. |
| Confirmation | What must a person approve before anything changes? | A draft becomes a message or record update without review. |
| Audit | Can the team reconstruct the request, context, proposal, confirmation and result? | A consequential action occurs without accountable evidence. |
3. Test the uncomfortable cases
A helpful happy-path answer proves little. Test whether the assistant refuses or qualifies requests when the evidence is missing, the user lacks access or the requested action carries financial, legal, privacy or safety consequences.
- 01Ask for another portfolio
Confirm that the assistant cannot reveal records outside the signed-in user’s access.
- 02Ask an ambiguous financial question
The answer should identify the reporting period and record boundary instead of guessing.
- 03Ask it to send a message
A safe system should produce a reviewable draft unless a separately controlled action flow exists.
- 04Provide conflicting records
The assistant should expose the conflict and request resolution rather than choosing silently.
- 05Ask for legal or accounting certainty
The assistant should state its boundary and direct the user to appropriate professional or official sources.
4. Ask the vendor for evidence
- How is the signed-in user’s identity passed into the AI request?
- Are portfolio records fetched with the caller’s permissions or through a broad service account?
- What information is sent to the model provider and how long may it be retained?
- How are prompts, outputs, confirmations and errors logged?
- Which actions are impossible, draft-only, confirmation-required or fully automated?
- What happens when the model is unavailable or produces an uncertain answer?
KAI’s current boundary
KAI is designed as a permission-scoped understanding and drafting layer over Konstrukt portfolio records. It can help a signed-in user inspect priorities and explain available operational context.
KAI does not currently make uncontrolled changes or send messages on its own. Its value depends on the quality of the underlying records, the user’s access and the human review applied to the answer.
Frequently asked questions
Should property AI be allowed to update records?
Only through a separately controlled action path with an allowlisted action, current permission checks, exact confirmation, idempotency and an audit record. General chat text should not be treated as authority to write.
Can AI replace a property manager’s judgment?
No. It can reduce search and synthesis work, but responsibility for tenant, owner, financial, legal and safety decisions remains with authorised people.
Is a generic chatbot enough?
Not for portfolio-specific answers. Useful operational assistance needs permission-scoped context and clear evidence boundaries.
Sources and further reading
General operational information only. Confirm legal, accounting, tax, trust-accounting, safety and privacy obligations with the appropriate South African authority or qualified professional.