A five-day design sprint to validate Gavl — an AI-driven proposal engine and agentic assistant for South Africa's R300B property market.
South African auctioneers are losing deals and capacity because their proposal workflow is manual, slow, and unscalable — while sitting on a data gold mine nobody is collecting. This sprint asks one question: will brokers trust AI-generated proposals enough to present them to sellers?
The template engine is the trojan horse. The data is the real product. The sprint validates the trojan horse.
312,000+ residential sales worth ~R300B in 2024. Auctions capture only 3–8% vs Australia's 17–55%. Massive headroom.
30+ auction houses run manual workflows: Lightstone PDF → Property24 → Word template. No auction-specific proposal tool exists.
The Deeds Office doesn't classify sale method. Nobody tracks auction outcomes. Every proposal generated passively fills this gap.
The risks that could kill the venture — framed as testable hypotheses.
Monday's goal is to build a shared understanding of the entire customer journey — from broker sourcing a property through to seller signing a mandate — and identify where the biggest opportunity lies.
After expert interviews, the team generates opportunities as questions — one per sticky note.
Target Customer: The AC Broker — the person who presents proposals to sellers.
Target Moment: The moment the broker reviews a generated proposal before sending to the seller.
"If the broker doesn't trust the output enough to present it, nothing else matters."
Tuesday opens with Lightning Demos — rapid 3-minute presentations of analogous products that inspire specific features. The afternoon is spent sketching divergent solutions individually.
WhatsApp message → 5-minute AI generation → broker reviews on phone → sends link to seller. Speed is the feature.
Desktop-first pipeline: intake → enrichment → AI draft → broker edits → branded PDF → share link with tracking.
Web-based interactive proposal — seller explores comps on a map, sees suburb trends live, signs mandate inline.
Agentic AI that proactively suggests reserves: "This Sandton property is similar to 3 others — suggest R3.2M."
Outcome feedback loop: after auction, results feed back into the AI model, making each proposal smarter.
The Decider resolves the tension between "Instant WhatsApp" and "Agent Dashboard" — combining both: WhatsApp as the intake, dashboard as the review. The 15-frame storyboard maps the full click-through from broker sourcing to seller signing.
WhatsApp as intake, Dashboard as review. They share 80% of the screens. The prototype tests both entry points.
Chat with "Auction Central AI" — sends property address.
Seller name, reason for selling, preferred date, features.
"Pulling property data… Analysing comps… Writing description…"
"Your proposal for 21 West Road is ready. Review it here →"
Pipeline view with AI-generated proposal card.
Matches AC's brand voice with "Edit" capability.
8 properties within 2km, sorted by distance and similarity.
12-month median price chart with +12.3% growth indicator.
AI reserve with transparent reasoning: comps, trends, cross-checks.
Accept AI recommendation or enter their own reserve price.
AI-generated letter ready to present.
Pixel-match to current AC template.
WhatsApp/email/SMS with view analytics.
"Seller opened 3 times, spent 4 min on comps."
E-signature embedded in web proposal.
Thursday produces the full click-through prototype spanning 15 frames — mobile screens for the broker flow and desktop pages for the proposal document. Everything uses the "Sapphire Archive" design system: deep navy, institutional gold, Inter typography.
Friday puts the prototype in front of 5 real users — power brokers, new brokers, a recent seller, a Tier 2 auction house principal, and a traditional estate agent. The team watches behind a one-way screen, noting reactions moment by moment.
| # | Profile | Why This Person |
|---|---|---|
| C1 | Active AC broker (high volume) | Power user — will they trust AI output enough to present? |
| C2 | Active AC broker (newer) | Less experienced — does AI-assisted valuation help them? |
| C3 | Recent AC seller | Did the proposal influence their decision? How would AI compare? |
| C4 | Tier 2 auction house principal | Would they adopt this tool? What would they pay? |
| C5 | Traditional estate agent | Could auction AI proposals convert traditional agents? |
| Pattern (if 3+ agree) | Signal | Decision |
|---|---|---|
| Brokers trust AI descriptions | ✅ Validated | Ship it |
| Brokers override AI valuations | ⚠️ Expected | Keep override prominent |
| Sellers prefer interactive web proposal | 🔄 Pivot signal | Web-first, PDF as export |
| Brokers won't use WhatsApp intake | 🔴 Kill signal | Web form primary |
| Tier 2 house would pay R250/proposal | ✅ Validated | Proceed to Phase 3 faster |
| Assumption | Verdict | Evidence |
|---|---|---|
| SA auction market is large & under-served | ✅ Validated | ~R300B market, 3–8% auction share vs 17–55% in Australia. 312K+ sales/year. |
| No purpose-built proposal tooling | ✅ Validated | PropData and CMA Info serve agents, not auction houses. No WhatsApp intake, no auction-specific templates. |
| No centralised auction intelligence | ✅ Validated | Deeds Office doesn't classify sale method. SAIA doesn't aggregate. Nobody tracks outcomes. |
Validate broker trust — everything else is secondary.
| Outcome | Action |
|---|---|
| Strong validation | Proceed to Phase 1 build (4–5 weeks). Priority: proposal engine with review dashboard. |
| Mixed signals | AI assists with data + description, valuations stay manual. Still saves 70% of time. |
| WhatsApp rejected | Kill WhatsApp intake. Web-form-first with WhatsApp as future add-on. |
| PDF preferred | PDF-first output. Interactive web proposal moves to Phase 2+. |
| Tier 2 house excited | Accelerate multi-tenant architecture. Run parallel Phase 1 + Phase 3 planning. |
These interactive research deliverables were produced alongside the sprint to validate market assumptions, model unit economics, and map the competitive landscape. Each opens as a standalone interactive document.
Interactive P&L projections: revenue per proposal, cost structure, and path to profitability across 3 phases.
5-year revenue model with scenario analysis: conservative, base case, and aggressive growth assumptions.
Feature matrix and positioning analysis of PropData, CMA Info, Qwilr, and other adjacent tools.
Top-down go-to-market strategy: target auction houses by tier, geography, and outreach sequence.
Decision-maker mapping for the Lightstone partnership pitch: leadership, product heads, and data team.