Do not switch your whole agency onto AI lead nurturing in one move. Roll out one workflow through five gates: map a fenced pilot slice, set permission limits, replay past inquiries with every outbound message held for approval, run a live pilot with a named owner, then expand only when the evidence holds. At each gate you decide to proceed, pause or roll back to your manual process. Your agents keep every real conversation; the AI handles first response and qualification only.
The five gates at a glance
| Gate | What you do | Exit signal |
|---|---|---|
| Map | Fence one slice and document the workflow | Every stage has a named owner |
| Configure | Set approval tiers and name approvers | You know which limits software enforces |
| Shadow | Replay past inquiries, outbound held for approval | Held messages match what you would send |
| Pilot | Go live on the fenced slice, review daily | Pilot measures stay within agreed limits |
| Expand | Extend, pause or redesign on evidence | Each new slice restarts at map |
Gate 1: map one fenced slice
Pick the safest slice, not the busiest. One inquiry source, one branch or one property category. A fenced pilot is one you can switch off without touching the rest of the business. Simon's Bulawayo real estate AI assistant use case tells the story of a deployment like this; this article is the method underneath.
Before any automation, the workflow must exist on paper: trigger, reply rules, stage owners and fallbacks. If you have not mapped how an inquiry currently reaches an agent, start with the eight-stage property inquiry workflow. The manual versus automated nurturing comparison covers the earlier decision to automate at all.
The AI workflow readiness audit sets the fencing rules: one slice, a named owner, human approval on outbound messages, a rehearsed rollback and a review date agreed before launch.
Gate 2: set the permission lines
Decide which messages the AI may send on its own, which wait for approval and which a person must always handle. The approval rules for AI lead follow-up define the three tiers and the four-commitment test for spotting a message that binds the agency; apply them to your slice rather than rebuilding them. Instant acknowledgements, qualifying questions on budget, timeline and location, and internal tagging and scoring of the lead sit in the automatic tier; anything that commits the agency waits for a named approver.
Name the after-hours approver now, because inquiries arrive at 9pm. An acknowledgement can go out automatically, but any commitment waits for that person or a holding message promising a human reply.
Then ask TechTribe during onboarding which limits the Platform enforces and which depend on people following procedure, and write the answers down.
Gate 3: replay past inquiries before going live
Shadow is an operating practice you agree with TechTribe during onboarding. Take recent real inquiries from your pilot slice, run them through the configured workflow and hold every outbound message for human approval; the approval rules article uses twenty past inquiries as the working set.
Review each run against the five run-log fields set out in the guide to monitoring an AI workflow after launch: date and time, input reference, action or output, human involvement, outcome. The property lead follow-up logging method applies those same fields to property inquiries. Without that record you review from memory, and memory flatters pilots.
Look for wrong availability answers, invented details, mishandled tone and missed escalations. Adjust and replay until the held messages are ones you would have sent yourself.
Gate 4: run the pilot with a named owner
Go live on the fenced slice only. One person owns the pilot. Every day they work the exception queue, which holds the conversations the workflow could not finish; they approve held messages; and they hold the authority to pause. Whether that queue is a screen in the Platform or a shared label your team maintains is one of the answers you wrote down at gate two. The weekly evidence review in the monitoring guide is a separate, slower pass.
Agree the pause and rollback triggers in writing before launch; the monitoring guide calls these switch-off triggers. The landing state is your manual process, so keep it documented and rehearsed for the pilot slice.
Simon's guide to designing the AI-to-human handoff covers where humans belong in any AI workflow; this article stays inside this one.
What agents do during the pilot
Inside the pilot slice, three things change for an agent.
They receive routed leads with context. When the AI marks a lead warm and qualified, the assigned agent gets the conversation summary in the central inbox: budget, timeline, location and what has already been said. The agent continues the conversation rather than restarting it.
They pick up escalations. When the AI cannot answer, a lead asks for a person or a stop condition trips, the conversation reaches the exception queue with full context and the pilot owner assigns it.
They stop replying to pilot-slice leads from personal chats. Parallel private conversations kill pilots quietly; if a lead contacts an agent directly, the agent handles it, then records it in the workflow.
Nothing else about their day changes. Keep it that way: a rollout that redesigns everyone's job gets resisted.
Handle the awkward cases before the pilot starts
Opt-outs. WhatsApp's Business Messaging Policy requires opt-in permission before you message someone, plus clear opt-out instructions that you honour. Replying to an inquiry a lead sent you is one thing; adding them to an automated follow-up sequence later is where opt-in bites. When a lead says stop, automation for that lead stops; a person decides any further contact.
Unavailable properties. A confident wrong answer about a sold house costs more than a slow right one; availability replies sit in the needs-approval tier.
Duplicate leads. The same person often inquires through both the website and WhatsApp; routing rules and the central inbox give one lead one owner.
Failed messages. A message that never sends is a system failure, tracked separately from wrong outputs. Watch for absence too: a day with abnormally few runs in the log needs checking.
After-hours inquiries. Gate two set the approval rule but not the cover. Name a backup for when the after-hours approver is on leave or unreachable, and agree how long a held commitment may wait.
Before you lean on automated lead scoring
Controller duties under SI 155 of 2024 apply to your agency regardless of who inspectors visit first. DLA Piper Africa's commentary on the Cyber and Data Protection Act notes that controllers answer for the acts of their agents, and that a decision based solely on automated processing needs consent or another legal basis. Veritas's Bill Watch summary of POTRAZ Regulatory Notice 2 of 2026 set 1 September 2026 as the start of mandatory compliance inspections on a risk-based model, and does not list real estate among the priority sectors. Take advice from your lawyer or POTRAZ before a score decides who gets called back.
Gate 5: expand on evidence, not relief
That brings you to the last gate. Read the four pilot measures from the readiness audit: completion rate (inquiries the workflow finished without a person stepping in), exception queue volume (conversations it could not handle), time to first acknowledgement, and correction rate (the share of AI-assisted replies a person had to fix before they were usable).
Expand when completion holds, exceptions stay boring and corrections are rare. Pause when exception volume climbs or corrections cluster around one failure type, and fix the configuration first. Redesign when the workflow map itself was wrong; no configuration repairs a badly drawn process.
Extending to a new branch, inquiry source or property category is a new mapping decision with its own fence.
Sequencing several AI initiatives across the agency is executive territory; Simon's 90-day AI rollout plan covers that layer above this single deployment.
Where the Platform fits
This workflow runs on TechTribe's Real Estate Platform. The AI answers inquiries within seconds around the clock, qualifies on budget, timeline and location, scores and tags each lead, and routes it to the right agent with a summary. Leads from listing platforms can feed the same workflow when connected. The Platform is a layer on top of a real estate website, and TechTribe's pricing page separates the two and carries the current figures: the basic website is a once-off build with no monthly fee and does not include this workflow, while the Platform adds a setup fee and a monthly fee that pays only for the AI and operations layer. It runs month to month, cancellable with 30 days notice, and your website stays live if you cancel. Priority support with a 24-hour response is included.
Basing replies on your own listing data lowers the error rate without removing it, which is why the approval tiers and the run log exist. The Platform does not manufacture a process; it automates the one you mapped at gate one.
Start with the slice you could switch off tomorrow. Write down who owns each stage of it, and you have gate one.
Put AI lead nurturing to work without losing control
TechTribe's Real Estate Platform answers inquiries within seconds, qualifies leads and routes them to your agents, with your team in charge at every gate. The Platform layer runs month to month with 30 days notice, current figures are on TechTribe's pricing page, and your website stays live if you ever cancel.
About the author
TechTribe Team
The TechTribe team writes practical guidance about websites, lead capture, and digital growth for businesses in Zimbabwe.
