Articulate · Build Scope

The build: tools, prototypes, delivery

What Anthony actually builds as the build partner — the engine architecture, the five prototypes that prove it, the tools already in hand vs to acquire, and the week-by-week plan to a measured pilot.

Scopes Phase 1 (energy re-sign) with Phase 2 (Dynamo PCP) portability designed in · companion: opportunity · dossier

AThe engine — one architecture, swappable brains

Channel layer
WhatsApp Business API · web chat on landers · voice later (ElevenLabs)→ Conversation brain
Claude API + vertical RAG ("knowledge brain") + persona→ Guardrail layer
CMCOB scripts · free-route disclosure · opt-out · full audit log→ Intake & e-sign
plate/postcode qualifier · agreement finder · LOA e-sign→ Delivery & ops
case handoff to CMC/law firm · funnel telemetry · alerts

The design rule from the opportunity analysis: the engine is permanent, the vertical is a plug-in. A vertical = knowledge brain (RAG corpus) + persona (Dynamo Dave / energy expert) + intake config + commercial agreement. Porting a new vertical must take days, not weeks — that's the moat, because Fintan's pipeline is databases, plural.

A2How we actually contact them — and how everyone else does

Our route: the database owner sends, we converse

The legal architecture is the channel. The first touch always comes from the data owner, under their brand, on channels their consent covers. We never cold-message their list ourselves — that's what makes it lawful and what nobody else can replicate.
Email — Dynamo/introducer's own ESP, their opt-in. For claims content the consent wording needs checking (Phase 0 DPO question); "soft opt-in" covers similar own services, claims marketing is a stretch on cold slices.
Letter post — the sleeper channel: no PECR consent needed, GDPR legitimate-interest only. A Dynamo-branded letter with a QR straight into the conversation covers the un-consented slice lawfully. Slow, ~60p/unit, but it rescues volume the email lawyers will strike out.
SMS — PECR consent required (B2C). Use only on the consented slice; highest open rates, highest complaint risk.
WhatsApp — NOT a cold channel. Meta requires user opt-in; cold blasts get the number banned. Correct use: capture the WhatsApp opt-in inside the first web conversation, then Emily continues there. Anyone promising "WhatsApp the whole database" is describing a banned pattern.
Phone — live calls under legitimate interest but TPS-screened (B2C) / CTPS (B2B). This is the £25/contact cost the engine exists to delete; voice-AI outbound to consumers is a regulatory minefield — keep voice inbound-only for now.
B2B (energy) — corporate subscribers sit outside the PECR consent rules that bind B2C: email and calls to businesses run on legitimate interest + CTPS screening. This is the real reason the energy re-sign is the right pilot — the channel law is genuinely lighter, and the 20k are existing clients besides.

How the rest of the market does it (rented attention)

Paid social — Meta/TikTok ads → plate-first landers. The dominant CMC channel; CPLs bid up by 78 near-identical funnels (database).
Paid search + SEO — Google ads on lender-name terms; programmatic SEO (pcpclaims.info's FOS archive is this play, not yet ranking).
Bought leads — CPL/CPA from lead-gens (Lead Pronto, MCB, Impact) at £3–£30+ raw, multiples for signed LOAs.
Editorial rental — Locksley Law's Top Gear ad-feature: borrow a top-15k domain instead of building one.
TV/radio — legacy CMC brands; expensive, dying.
Lender letters — the scheme itself: lenders must write to eligible customers. The state-mandated channel that eventually erodes everyone else's.
The asymmetry: all of it is rented attention at rising auction prices. An owned, consented database with a conversational layer is the only acquisition channel in this market with a marginal cost near zero — that's the entire venture in one line.

BThe five prototypes

Each is independently demoable, ordered so the most persuasive (and riskiest) is first. P1+P3 together are the demo that closes Fintan's partners.

P1"Emily-Energy" — conversational re-sign demo2–3 days · week 1

WhatsApp (sandbox number) + Claude + an energy mis-selling knowledge brain built from the introducer's claim pack + public sources. Persona: empathetic, expert, never pushy. Handles: what happened in the scandal, no-win-no-fee mechanics, "what's the catch", "how much could I get", and routes to re-sign.

Done when: a 10+ turn natural conversation survives hostile questions, quotes fee terms accurately, discloses the free route unprompted, and issues an e-sign link. demo to Fintan = the negotiating card

P2Reg-plate-first web intake1–2 days · week 2

Lander with a single plate field (the market's converged pattern — 7 of 10 top funnels lead with it, and Dynamo's own phone traffic proves the behaviour) → vehicle confirm → chat takeover by the brain. Built on our standard Vercel stack.

Done when: plate → vehicle lookup → qualified handoff works end-to-end with zero PII before the qualifier step, matching the PII-gate-at-step-2 market standard.

P3Compliance guardrail harness1–2 days · week 1, parallel

Automated eval suite that attacks the bot with CMCOB failure probes: does it disclose the free route, does it invent urgency, does it overstate compensation, does it respect "stop messaging me", does it claim to be human. Runs on every brain/persona change, results logged. This is the productised version of the regulatory analysis in the dossier §3 — and the single biggest differentiator vs the 50/70 non-compliant competitor herd.

Done when: 100% pass on the probe suite + every conversation stored with full audit trail. this artefact is what the law firm signs off

P4LOA e-sign loop1 day · week 2

Conversation → pre-filled LOA → e-sign (DocuSign already in the stack, or the partner firm's pack) → signed doc delivered to the law firm + logged. The £15/£60 only exists when this loop closes cleanly.

Done when: signed test LOA lands in a partner inbox with case metadata attached, round-trip under 3 minutes from chat.

P5Funnel telemetry dashboard1 day · week 3

Contacted → responded → engaged → qualified → LOA signed → clean case, per vertical per day, with cost per stage and anomaly alerts (ntfy). The pilot's success criteria (§5) are read off this dashboard, not asserted.

Done when: the 60-day pilot gate (conversion ≥20%, cost/re-sign ≤£1.50, zero flags) is measurable live by Anthony and Fintan without asking anyone.

CTools — in hand vs to acquire

Already in the stack (£0)

LLM + RAG: Claude API (keyed), embedding/RAG pattern proven in-house; knowledge-brain build is content work, not engineering.
E-sign: DocuSign connector live in the toolset.
Voice/transcription: ElevenLabs (keyed) for voice-note handling and later voice persona.
Hosting/deploy: Vercel pipeline (this site is on it), cwdeploy chain, landers in hours.
Telemetry plumbing: ntfy self-hosted, dashboard patterns from MediaServer/VillaManagement builds.
Competitive intel: the 78-site probe tool re-runs on demand; funnel patterns already mapped.

To acquire (small)

WhatsApp Business API number: via BSP (360dialog/Twilio) — days to provision, ~£50–100/mo + per-conversation fees. UK number, business-verified.
Vehicle lookup: DVLA VES API (free tier) for plate → vehicle confirm; the partner chain's agreement-finder (soft credit search) for finance history — partner API, ask Fintan.
Energy knowledge pack: the introducer's claim pack + original LOA — Phase 0 document ask.
Script sign-off: solicitor review via Blue Lion/partner firm, £2–5k one-off.
CRM handoff spec: how the law firm wants cases delivered (webhook/CSV/portal) — unknown until asked.

DDelivery plan — six weeks to a measured pilot

Week 0–1 · Phase 0 diligence + P1 + P3
The three documents (introducer agreement, FRN/SRA verification, energy-list consent basis) run in parallel with the P1 demo build. If diligence fails, sunk cost is a demo that's reusable anyway.
Week 2 · P2 + P4
Intake lander + e-sign loop. End of week 2: full chain demoable — plate/contact → conversation → LOA signed.
Week 3 · P5 + integration + sign-off
Dashboard live, law-firm script sign-off in, WhatsApp number verified, 500-contact soft launch list agreed.
Week 4 · Soft launch — 500 contacts
Measure, fix, re-run guardrail suite on every change. Gate: response + zero complaints before scaling.
Week 5–6 · Scale to the 20k
Batched sends, telemetry-paced. Pilot verdict read off P5 against the §5 criteria at day 60.
Sept–Oct · Port to Dynamo (Phase 2)
New brain (PS26/3 rules from the dossier), Dynamo Dave persona, reg-plate intake from P2, consent-checked list from the DPO. Launch the week the Tribunal rules (~Nov).

EEffort & cost roll-up

Item	Anthony-time	Cash
P1–P5 prototypes	~7–9 days	£0 (existing stack)
Integration, sign-off, soft launch	~4–5 days	£2–5k legal review
20k run (LLM + WhatsApp + e-sign)	~1–2 days oversight	≈£3–5k
WhatsApp BSP + misc SaaS	—	~£150/mo
Pilot total	12–15 days over 6 weeks	≈£5–10k

Capacity check: 12–15 days over 6 weeks ≈ 2–2.5 days/week. Stacked against GIG (≤2 days/wk), Roche-Debbie (HOT) and Articulate Phase 1, this only fits if the pilot displaces something or weeks 4–6 drop to oversight-only (realistic — the build front-loads). Phase 0 + P1 alone = ~4 days and produces the demo that decides everything else.

The strategic point: every prototype here is a reusable Articulate asset. P3 (compliance harness) and P5 (telemetry) are productisable for any regulated-conversation client; P1's pattern is the Connie/Emily lineage. Worst case — Fintan's paperwork never materialises — the build still ships as Articulate IP. That asymmetry is what makes the seat worth taking, gated.

Articulate internal · not legal advice · vault: Clients/Fintan/PCP/build.html · voice notes in Clients/Fintan/whatsapp/