AI Is Getting More Capable. Who Gets to Say Yes?

Imagine a contractor's invoice arriving on a Friday afternoon. The repair is urgent. The quotation was approved, and the invoice matches the amount. An AI assistant brings the paperwork together in seconds.

Then a later email asks for payment into a different bank account.

In this hypothetical example, the assistant has helped organise the file. The changed banking instruction still needs checking. Before anyone approves payment, they need to know which checks have been completed, which remain open and who has authority to proceed.

That is becoming a more pressing question as AI gains the ability to act through connected systems. An AI agent is software that can use tools to carry out tasks, such as searching records or drafting an email. What it can actually do depends on the access it has been given.

On 28 September 2026, NVIDIA announced its Open Agent Safety Platform, including controls designed to enforce permissions and record agents' actions. Anthropic's accompanying explanation describes limits enforced outside the AI itself.

For trustees and HOA boards, the practical question is close to home: what has your managing agent's AI been allowed to do with your scheme's information, correspondence and money?

Give people a better file to work with

There is useful work for AI before a decision is made. It can help locate an earlier resolution, assemble a repair history or find correspondence explaining why a contractor's appointment is still unresolved.

The value is in making the important information easier to check. A property manager can spend less time reconstructing the history and more time resolving the issue. A trustee can review a clear recommendation with its evidence available.

Give the tool a defined job and the records that job needs. An assistant reviewing a contractor's invoice needs the relevant project file. There should be a clear reason for any access to unrelated resident information or another scheme's records.

Make the authority clear

Preparing a recommendation, authorising it and carrying it out are separate steps. The systems should make those steps visible.

Ask what the actual tool can do: read documents, change a record, send a message, instruct a contractor or release a payment. Some tools only prepare drafts. Others have connections that allow actions. The controls need to match those connections and the scheme's agreed delegations.

A routine internal update and a payment instruction carry different consequences. Where an action needs approval, the system should prevent it until the required authorisation is recorded. A sentence telling the AI to ask first is only one part of that control.

The relevant authority comes from the scheme's governing documents and delegations. It may sit with trustees, HOA directors or a manager within an agreed mandate. It needs to cover the specific action being proposed.

A matching amount does not settle the payment

Return to the invoice. A useful review note would identify the approved quotation and its date, check any payment conditions, and flag the changed banking instruction. It should also say what has actually been checked.

For example:

Each finding should point to the document supporting it.

This is a real business risk. SABRIC's 2025 report describes supplier-payment fraud involving changed banking details and compromised communications. Nedbank advises confirming changed details with a known person at the supplier, using the number normally used to contact them.

Record that verification and follow the scheme's payment procedure, including any bank account verification checks. An attached confirmation letter is only part of the evidence; the changed details still need independent confirmation.

Urgency needs its own clear route: who can respond, what authority they have and which checks are required. That helps an urgent repair move forward with an accountable decision.

Give the reviewer a clear job

'A person reviews it' needs to mean something specific. The reviewer should be able to open the source records, question the conclusion and hold the action while a gap is resolved.

Check that the approval covers this contractor, this work and this version of the quotation. An earlier approval may refer to a different scope. If payment depends on work being completed, identify who confirmed completion and what evidence they used. An invoice alone cannot establish that.

If an approval cannot be found, the assistant should say which records it searched. A missing document leaves a question to resolve; it does not establish that approval was never given.

Keep the final decision, its authorisation and any outstanding conditions with the supporting records. The next trustee or manager should be able to understand why the action was taken.

Five questions for your next meeting

Choose one regular task, such as approving a contractor payment, and ask your managing agent to walk through it:

Ask for a demonstration using fictional records with a changed bank account or an approval that covers an older quotation. Watch where the system pauses and what the reviewer sees. Repeat across several cases; one successful demonstration cannot establish reliability.

Measure the time needed to reach a checked decision, including review and follow-up. Count errors or gaps missed, as well as those found. That gives you a useful basis for judging whether the tool reduces the work.

The standard I would hold us to

At PropAI, we use AI to bring relevant correspondence, recorded decisions and supporting documents together around a scheme issue, with sources people can check. Our team reviews those records and prepares the next step.

I would judge our work by how well it helps the responsible person make and follow through on a decision. Can they see the evidence, understand the uncertainty and identify who must act next?

As AI becomes more capable, a scheme should be able to explain its authority just as clearly as its usefulness. Knowing who gets to say yes is part of good management.