The pattern to hold above everything else
Send one message to one named human and find out. Doctrine without buyer contact is journaling. End each working day on a name and a sent message, not another saved artefact.
The mental model in one paragraph
WilliamDavid is becoming the most credible African travel intelligence layer on the continent. We sell six different containers of the same expertise. The supplier graph is free at entry and buyer-funded above it; the brain is Carmen's reasoning running on Claude; the moat is verified African ground truth plus a named voice nobody else can claim. The Day-91 target (17 Aug 2026) is one paying Ella Embedded customer at R15–25k/mo. The Feb 2028 destination is paid advisory at R200k+/mo to airlines and boards.
The four 10-year-old definitions
Claude
A super-smart computer friend who can read, write, and think about almost anything you ask it. Careful and honest. Uses tools — calendars, notebooks, calculators — properly.
OpenAI
The big company that built ChatGPT. Wants their robot to help everybody in the world with everything. A giant supermarket selling every kind of food to every kind of person.
Ella
A super-smart friend who only knows about one thing — really, really well — and that thing is travel across Africa. She remembers every good hotel, guide, road, and supplier, and helps grown-ups plan amazing trips in ten minutes instead of weeks. She is not trying to be good at everything; she is trying to be the very best in the whole world at this one thing.
Surge AI
A small company that hires really good writers and thinkers to teach AI robots how to give better answers. Quiet, founder-led, has become one of the most important companies in AI without anyone really knowing about it.
The difference, said simply: Claude is a brain you can rent. OpenAI is a supermarket. Ella is the one person who knows African travel better than anyone — and uses Claude's brain to do her job faster.
The six SKUs · never blurred
| SKU | Buyer | Price |
|---|---|---|
| Blueprint · consumer PDF | Leisure traveller | R299–R499 once |
| Lane C Triage · WhatsApp triage AI | Small SA tour operators | R15k setup + R5k/mo |
| Ella Embedded · intelligence layer | Surviving / larger operators, call-centre businesses | R15–25k/mo |
| AVN Verified · free supplier entry | Suppliers Carmen has booked / stayed at | R0 (buyer-funded) |
| AVN Verified Plus · opt-in amplification | Suppliers already in | R499/mo opt-in |
| AVN Buyer Access · Trust / Intelligence / Access | Corporates, DMOs, embassies, productions, agents, airlines | R20–400k/mo or per-engagement |
| Travel Desk · internal substrate + 1 pilot | WD itself | Internal |
| Advisory · paid route-economics | Airlines, boards, governments, investors | R100–400k/mo or R250k–R2m+ project |
In any buyer conversation, name the SKU. "Carmen's stuff" or "our AI" is too vague. The buyer can't price it, compare it, or justify it internally if it's unnamed.
The four canonical doctrine sentences
- Ella positioning — "She is not trying to be good at everything; she is trying to be the very best in the whole world at this one thing."
- AVN governance — "Suppliers don't pay to join the network. Buyers pay to use it."
- Supplier pitch (save + earn) — "AVN suppliers pay us nothing. We 10x them anyway — by replacing what they overspend on (OTA commission, broad marketing, unqualified-lead time) with what they can't buy at any price (direct access to verified African enterprise buyers, at full rate, on their slow days)."
- Data-rights TOS — "WilliamDavid retains the right to use anonymised transaction-level data — itinerary patterns, supplier rates, conversion timings, route demand signals — for tourism intelligence, route-economics analysis, and aggregate reporting to third parties including but not limited to tourism boards, airlines, and investors."
If a piece of copy violates any of these, the copy is wrong.
The Ella Embedded sales line (30h → 10min)
Most travel operators spend 20–30 hours on the front end of a complex itinerary. We get version one in front of you in 10 minutes — verified African supply, real availability, three options you can click through and pressure-test. You still get a business day to refine. The 30 hours go away.
The Day-91 gate (17 Aug 2026)
One paying Ella Embedded customer at R15–25k/mo by 17 Aug. Lane C is parallel-but-secondary. One Ella Embedded customer is worth more than three Lane C operators.
| Condition | Met | Marginal | Missed |
|---|---|---|---|
| Paying customer | R15k+ contracted + invoiced + collected | Signed, payment pending | Discovery only |
| Buy-side pipeline | 2+ named buyers active | 1 named buyer active | Categories only, no humans |
| Public voice artefact | Intelligence report published + 1 appearance booked | Report OR appearance | Neither shipped |
| Data substrate | 20+ verified suppliers, queryable, consent-compliant | 15+ verified, partial structure | <10 verified or unstructured |
Passes if 3 of 4 = Met and the 4th is Marginal at worst.
The execution plan — six phases, sixty working days
| Phase | Anchor milestone |
|---|---|
| 1 · 26 May – 6 Jun | AVN pack rewritten (buyer-funded). 10-named-buyer list exists. First 2 outreaches sent. |
| 2 · 9 – 22 Jun | Blueprint launches 22 Jun. 5 verified suppliers. Ella Embedded prototype runs on Carmen's own ops. 1 buyer discovery call. |
| 3 · 23 Jun – 6 Jul | 1 Ella Embedded scoped pilot in active discovery. 1 Lane C trial. 10 verified suppliers. Q3 intelligence report scoped with 2 pre-committed buyers. |
| 4 · 7 – 20 Jul | One contract signed at R15–25k/mo. Intelligence report 50% drafted. AviaDev/Routes Africa appearance booked. 15 verified suppliers. |
| 5 · 21 Jul – 3 Aug | Pilot embedded. First invoice raised. Intelligence Report v1 published under Carmen's byline. 20 verified suppliers. One follow-on prospect. |
| 6 · 4 – 17 Aug | Day-91 gate. First paid month delivered + collected. Renewal note signed. Second pilot in scoping. |
The daily three-action rhythm
- One Reasoning Extraction paragraph written into the corpus.
- One named outreach action — message sent, intro requested, follow-up booked.
- One supplier-side action — verification touch, data capture, AVN record updated.
Three × 60 working days = 180 small actions. Most days, 45 minutes total. Architect and agents handle the rest.
The workload — by the hour
Added 2026-05-28. The time-cost of executing the six phases above. Estimates are Carmen-hands-on hours; the Architect agent absorbs most engineering wall-clock, shown separately. The 90-Day Plan remains the north star and bible — this is simply what executing the Core costs in hours.
A · The recurring floor — the daily three-action rhythm
| Action | Per day |
|---|---|
| Reasoning Extraction paragraph | 15 min |
| One named buyer outreach | 15 min |
| One supplier-side action | 15 min |
| Daily total | ~45 min |
28 May → 17 Aug ≈ 57 working days (weekdays minus 16 Jun Youth Day). Floor total: ~43 hours across the quarter (~3.5–4 hrs/week). This absorbs the first two outreaches, ongoing buyer contact, and most supplier verification — not double-counted below.
B · The discrete builds — additional work on top of the floor
| Phase | Task | Carmen hrs | External queue |
|---|---|---|---|
| 1 ✅ | AVN buyer-funded pack + data-rights TOS | shipped | — |
| 1 | 10-named-buyer list | 2–3 h | — |
| 2 | Ella Embedded prototype on Carmen's own ops | 10–15 h | Architect builds plumbing in parallel |
| 2 | Blueprint launch (22 Jun) — state unconfirmed | 6–10 h | payment / contractor |
| 2 | First buyer discovery call (prep + call + follow-up) | ~2 h | scheduling |
| 3 | Ella Embedded pilot proposal (Loom + Doc) | 3–5 h | — |
| 3 | Lane C trial setup (secondary) | 4–8 h | operator availability |
| 3 | Q3 report scoping + 2 pre-commits | 2–3 h | buyer replies |
| 4 | Close 1 contract @ R15–25k/mo | 4–8 h | legal review (days–2 wks) |
| 4–5 | Q3 intelligence report (full draft → publish) | 25–45 h | — |
| 5 | Pilot embedded into customer ops | 8–15 h | Architect parallel; integration depth |
| 5 | Report distribution (50 named) + publish mechanics | 5–8 h | — |
| 5 | First invoice raised | 1–2 h | collection timing |
| 6 | Renewal note + second-pilot scoping | 4–6 h | — |
| — | Supplier verification to 20 (active time beyond floor) | 6–10 h | supplier response latency |
Discrete total: ~90–155 Carmen-hours across ~12 weeks.
C · The weekly reality-check
- Floor (~4 h/wk) + discrete (~8–13 h/wk avg) ≈ 12–17 hrs/week.
- Against a ~20 hr/week working block, that's 60–85% loaded — feasible, almost no slack in July.
- Two rocks bust the 45-min daily rhythm and need dedicated multi-hour blocks, not daily slivers:
- The Ella Embedded build (prototype Phase 2 + embed Phase 5): ~18–30 h combined.
- The Q3 intelligence report: ~25–45 h — the single largest line, and the credentialling artefact the AviaDev invite and buy-side credibility hang on.
The binding constraint is July.
Phases 4–5 stack the contract close, the report draft, and the pilot embed into the same fortnights. If anything slips, the report is the swing item — protect its hours first.
Caveats not papered over.
- Blueprint launch hours are an estimate — its current build state is unconfirmed. Verify before trusting that line.
- External queues are real and outside Carmen's control: legal review on the first contract, payment collection timing, supplier response latency, and the AviaDev/Routes invite (which depends on the report landing first). These can move the Day-91 gate even if the hours are on track.
This week (Tue 26 May – Sun 31 May)
- Tue 26 May: kill the R499 supplier pack as written.
- Wed 27 May: rewrite AVN one-pager + Typeform consent field + pitch script (Architect drafts, Carmen sign-off).
- Thu 28 May: lock the data-rights TOS clause in
strategy/contracts/data-rights-clause-v1.md. - Fri 29 May: 10-named-buyers list written and committed.
- Mon 2 Jun (Off-Sprint Block): Ella Embedded thin-layer scoping with Architect — system prompt, MCP integrations, scoped data view. End the block with it scoped, not built.
- Tue–Fri 3–6 Jun: first 2 named-buyer outreaches sent. Daily three-action rhythm running.
What we are NOT doing
- Travel Desk as wide-sale SaaS. Internal + max 1 pilot.
- R499/mo as a supplier entry fee. Free entry only.
- Bidding as primary on SAT TRDMP / Brand Tracker tenders in 2026.
- Wide supplier-acquisition campaigns. 20–25 in 90 days; the rest in 2027.
- Standalone advisory page / product page in 2026.
- LinkedIn content factories. The marketing is the work itself.
- Pre-deciding Year-3 customer counts. Projection ≠ strategy.
- Comparing ourselves to multi-decade global firms when we should be sending outreach messages.
Already benchmarked (don't redo)
| Benchmark | The take-home |
|---|---|
| Surge AI | Closest structural analogue. Lessons banked: rigorous gate as marketing, publish selectively, concentrate revenue, founder-as-named-voice. Trap to avoid: their stealth posture only works because their buyers were inbound; Ella must do outbound. |
| Bloomberg | Aspirational ceiling, not realistic peer. Different market size (~$30tn vs ~$200bn). Per-seat pricing already comparable ($16k USD/yr Ella vs $25-30k USD/yr Bloomberg). The gap is seat count + time, not price. |
| Phocuswright + global research firms | Complementary, not competitors. They cannot defend the Africa-resident, operator-grade square. Future partner / referral candidates, not threats. |
| Plus94 | Adjacent SA institutional research peer. Possible future partnership for first tier-1 procurement engagements. |
| Travel Desk SaaS to agents (initial thesis) | Killed by Council Round 4. Travel Desk = internal substrate + 1 pilot. Door B (data substrate) is the strategy. |
| Supplier-paid AVN entry (initial thesis) | Killed 2026-05-26. Inverts the moat. Suppliers free at entry; opt-in amplification only. |
What makes the product formidable
- Solve a real problem for one named person first. The first Ella Embedded build is for one named corporate travel manager, by name. Generality after specificity.
- Refuse the obvious version. Not a chatbot. An embed that anticipates — "I notice three delegates land Friday with no SAA-delay buffer; here are alternatives." The buyer didn't ask; the system gave them an answer they needed and didn't know to look for.
- Be opinionated. Have taste. The product should sound like Carmen, not like a polite assistant.
- Treat the error rate as the product. Zero hallucinations on dates, names, prices, availability — even if it costs speed. Buyers forgive slow, not wrong.
- Make the buyer feel smarter for using it. Not just faster. Smarter in the next board meeting.
The three credentialling artefacts for 2026 (the only outbound)
- One intelligence report under Carmen's byline by Q3 2026 — 50-named-contact distribution, 2 paid pre-orders.
- One AviaDev or Routes Africa appearance by Q4 2026 — invited, not pitched.
- One published anchor-customer case study by early 2027 — quoted, named, verifiable result.
End the day on a name and a sent message.
Not another saved artefact. The doctrine is finished. The strategy is locked. From here, the work is the work. One named human reached every working day until 17 Aug. That is the entire plan.