Wednesday afternoon, 14:47, at a two-partner managing-agent office off Somerset Road in Green Point. Eleven schemes under management, four hundred and thirty-eight units between them, and the trustee WhatsApp for a twelve-storey block in Sea Point has just spiked. The lift between the seventh and eighth floors has stopped again. The resident in unit 704 wants a trustee to phone the lift company right now. The chair, a retired quantity surveyor who checks his phone on his walk to the shops, has replied twice with a shrugging emoji. In a second window, a levy statement query has come in for the same block: an owner in Milnerton has paid R4,382 with the wrong reference, and it is sitting unallocated in the Nedbank Corporate Saver trust account. On a third thread, a plumber wants to know before he pulls the geyser fittings in unit 604's ceiling whether it is a body-corporate claim or a private one: the sectional plan says one thing, the master policy says another, and the trustees have not been in the office in six weeks.

None of this is what a managing agent studied for. The substance of the job — the STSMA Act, the fiduciary weight of a trustee resolution, the auditor's questions in September — is real work. What consumes the hours is the ledger-and-messaging churn around it. Levy runs. Reference errors. Maintenance tickets in three chat threads. Minutes to distribute. Contractors to chase. The CSOS-adjacent resident who has read enough of the Act to insist on a section 39 order for the geyser in 704.

AI does not fix any of the substance. It will not sign a trustee resolution or approve a special levy. What it earns its keep on is the ledger-tied messaging in between.

Where the hours actually go at a small SA managing agent

The reliable time drains at a body-corporate practice look like this:

None of it is fiduciary work. All of it is why the junior in the office is still writing WhatsApps at 17:40 on a Wednesday.

The levy-query loop, wired against the trust account

This is the single highest-return starting point for most SA managing agents. At a block of forty units, four to eight owners will message on any given day with a variant of "what is my current balance?", "did you receive my payment last week?", or "why has my levy gone up this year?".

The manual version involves opening WeConnectU, MDA or Prokon, pulling the owner's ledger, cross-referencing the trust account for the most recent EFT, checking whether the annual budget resolution moved the levy up, and drafting a reply the owner will actually read. Fifteen minutes per query, four to eight times a day, per block. It does not scale past three or four schemes without hiring a second admin.

Where AI helps. The AI layer reads the incoming message against the scheme it belongs to, pulls the owner's ledger from the property-management software via API or a nightly CSV, confirms the last payment against the trust bank feed, and drafts a specific reply. Not a template. A reply that names the unit number, the balance to the day, the last payment received and its allocation, and — where the levy has changed — a two-line reason tied to the AGM resolution that raised it. The managing agent hits send.

Where it must stay in draft-and-confirm mode. Any query that touches a discretionary call — a request to write off arrears, a section 39 CSOS notice, a threat of legal action, a hardship plea — is flagged for a human, not auto-answered. In my experience most managing agents lose trustee goodwill by trying to auto-reply to arrears queries. Don't. Draft; do not send.

Maintenance tickets and the four-way phone tag

The plumber in unit 604's ceiling is the archetype. A leak from 704 into 604 involves four parties: the resident of 604 (who is standing under it), the owner of 704 (who may live in Constantia, or in London), the trustee (who has an approval limit set in the scheme's spend-authority resolution), and the plumber (who needs to know, before he pulls the fittings, whose insurance is on the hook).

The manual version is four separate WhatsApp threads, none of which record what was agreed. Two weeks later, when the plumber invoices, nobody remembers whether the trustees approved R4,800 or R2,400 for the geyser or whether the owner is paying for the internal pipework separately. The audit trail, at year-end, is nowhere.

A better shape. A maintenance ticket opens from the first message, cross-referenced to the sectional plan and the scheme rules on responsibility (common property versus individual section, with the STSMA's Prescribed Management Rules as the default where the scheme has not registered its own) and to the trustees' approved spend limit. One structured card holds all four sides of the conversation, and drafts the messages each party sees. The trustee gets a note that reads: "Unit 704 geyser leak into 604. Photos attached. Reads as a common-property call under PMR 25(a). Estimated R4,800 by the plumber who did unit 302 in March. Approve, decline, or request a second quote?" The trustee replies from their phone. The record survives to year-end.

What must stay in draft-and-confirm mode here. The classification of common versus private property is a legal call the sectional plan and rules govern; the AI proposes a reading, and a human confirms it before money moves. Get it wrong and the body corporate pays for something an owner should have covered, or the reverse. Both damage the relationship with the trustees, sometimes past repair.

The AGM cycle and the paper trail the auditor asks for

This is where the compression happens.

SA body-corporate rules require the AGM within four months of financial year-end, with at least fourteen days' written notice, notice to CSOS, proxies collected, minutes distributed, and resolutions minuted in a form the auditor can trace nine months later. For a managing agent running eleven schemes, the paper cycle is unrelenting.

AI helps at three points. Drafting the AGM notice pack (agenda, prior-year minutes, trustee-election forms, proxy templates and the audited financial extract) in the scheme's standard format, ready for the agent to review before it goes out. Tracking proxies as they arrive so that, on the day, the quorum position is clear without a scramble at the door. Drafting the minutes from the meeting recording, with resolutions structured in the format the auditor expects at year-end.

None of these is a legal document until a human (the agent, the chair, the auditor at their turn) signs it. But the drafting time drops from a full day per scheme to under an hour.

Compliance renewals and the diary that does not drop

Lift certificates lapse annually. Fire-equipment inspections need a competent person to sign off under the OHS Act. Generators (many of them installed during the worst loadshedding stretch between 2022 and 2024) need scheduled servicing that most trustees stop tracking after year two. The body-corporate insurance revaluation is due every three years.

Each has a date. Each date, missed, becomes either a compliance risk or a claim the insurer rejects when the geyser next bursts. Manual tracking on a spreadsheet works for one scheme. It does not work for eleven. AI schedules the diary against the compliance calendar, sends the trustee a WhatsApp thirty days out, drafts the request-for-quote to the last contractor used, and only escalates to the managing agent if the trustee has not responded within seven days.

Not glamorous. Just a diary that does not drop a date because someone was on leave the week it fell due.

What the bot never touches

A short list. This is where trust is either kept or lost.

Approving a spend without a trustee resolution behind it. The trustees hold fiduciary responsibility under the STSMA. The AI drafts requests. It does not authorise expenditure, no matter how small the number.

Writing off arrears or granting a levy waiver. Both are trustee decisions with real financial consequences for every other owner in the scheme. A bot that concedes on arrears has just cost the body corporate money it cannot get back, and set a precedent it will pay for at the next AGM.

Handling a CSOS dispute or an Information Regulator complaint. Both are legal channels with formal response requirements and time limits. Every message on those threads is flagged and routed to a human, immediately.

Editing the sectional plan or the scheme's rules. These are legal instruments filed with the Chief Ombud and, for the plan itself, the Surveyor-General. Nothing about them is automatable.

Storing owners' banking details or ID numbers on any consumer AI service. POPIA and FICA both bear here. Data stays inside the managing agent's controlled infrastructure and the scheme's property-management software, full stop.

Where to start

Not everything at once. The mistake most managing agents make is to plug an AI layer over the whole practice on day one and end up with a half-configured tool that the trustees do not trust and the owners do not recognise the voice of.

The reliable order is four steps.

  1. Levy-query loop first, on one scheme. Wire it against the trust-account feed and the property-management software. Draft-and-confirm for the first month, then move the routine queries to auto-send only once the agent trusts the drafts.
  2. Maintenance ticket triage next, on the same scheme. Structure the four-way messaging into one record and one card per repair. Keep everything in draft-and-confirm.
  3. AGM notice pack and minutes drafting on the third pass, timed to the scheme's actual year-end.
  4. Compliance diary last, once the other three are stable enough that the agent can supervise a new loop without dropping the ones already running.

Then, and only then, roll the same four loops out to the second scheme.

Managing agents in Cape Town, Sandton, Umhlanga and the smaller schemes across Bloemfontein and Port Elizabeth who take this path report the same two things. The trustees stop feeling that everything is falling through the cracks. And the agent gets to Friday afternoon without seventy unread WhatsApps in a chair's group thread. That was the point of running eleven schemes with two partners.