Kaizan Demand Engine
Taha Zaheer · July 2026
This plan is deployed, not proposed. The engine, inbound ladder, and outbound radar are live; MVT-1 sends are wired but intentionally not executed in this artefact. Unit costs below are measured from the July build; the live ledger refreshes on every page load.
1·Lead magnet ladder
each rung earns its place with a mechanism
The Agency Client-Health Index (quarterly)Live →
Not searching yet — scrolling: “are my peers losing accounts, and am I in this?”
Aggregated from the engine's exhaust: CS-leadership moves and named account losses across trade press. Owned data no one else compiles; every agency MD reads it asking 'am I in it?'. Marginal cost ~ £0.
CARE Client-Health Self-AuditLive →
11pm search: “how do I know which client accounts are actually at risk?”
Score your top five accounts by hand on C·A·R·E. The buyer discovers mid-worksheet they can't answer without data — the gap is the demo. Deliberately manual, so it funnels into the product, never replaces it.
Signal Teardown → Paid PilotLive →
11pm search: “what does the market already see about my agency — and what would a tool show me?”
'What the market can see about your agency': a personalised scan of their public footprint, bridging straight into Kaizan's existing paid-pilot motion (£2,995 / 90 days) — the ask sales already makes.
Rung four (paid-pilot conversion) is at /inbound/4 — deliberately not built until rungs 1–3 produce qualified inbound.
2·Channel mix
one hypothesis each, tested not assumed
Signal-led outbound (email + LinkedIn DM)
Low CAC · high intentTiming beats persona: a first touch triggered by a verified event converts ≥3× a cold list. The engine's core motion.
Signal-led content — one authored company voice
Zero media spend · compoundsThe radar doesn't just find prospects; it feeds a single company channel weekly angles from live market events, so the brand reads as the category's insider. RevOps-published — no dependence on reps posting.
Client-Health Index as PR / inbound
Earned reach · owned dataA quarterly data release is inherently newsworthy to the same trade press the engine reads. The magnet and the radar are one machine.
Paid — deliberately not yet
Amplify only a proven messagePaid amplifies a proven message. Enters only after an organic message clears its MVT. Seed money never amplifies a guess.
3·AI architecture
Kaizan's product, pointed outward (the part most people skip)
The outcome, made cheap and repeatable: a qualified agency lead with a verified, nameable reason to reach out — approved at the RevOps gate and assigned in HubSpot — for under £0.15. Everything below is only how that stays true at volume.
Kaizan watches client relationships for churn risk and expansion white space (CARE); this engine runs the exact same discipline on the market — where's the revenue risk, where's the expansion — pointed at which agencies to pursue.
Live web, two parallel angles (people-moves + account moves), citations on.
Maps events to the signal taxonomy. Every source_url must exist verbatim in the citations — no invented URLs.
ICP blocklist · homepage URLs rejected as evidence · signals older than 12 months dropped.
Firmographic + people waterfall. Returns verified LinkedIn profiles and business emails for the named contact.
ICP fit 1–5, portfolio-count-first (Kaizan's pricing axis) × signal recency × severity × buyer accessibility.
Dedupe by domain · suppression (22 existing clients never re-surface) · running cost ledger.
Qualify · Claude drafts the first touch · RevOps approves and assigns — the engine never sends.
Nothing sends or pushes without explicit operator approval.
The outcome routes centrally — nothing depends on a rep opening a tool. Adoption risk is designed out.
The seam is deliberate. Discovery is custom because Clay assumes you already have a list — here there is no list until the engine manufactures one from live signal. Everything downstream of discovery (enrich → score → dedupe/suppress → CRM) is exactly what Clay orchestrates; at n=1 rep on a seed budget a few hundred lines beat a Clay seat, and Clay earns its place the moment a team shares the board.
What the scan searches
Neither a fixed company list nor the open web. Perplexity searches the live web, but every hunt is steered through an ICP-shaped lens: firm type, size band, geography, and verifiable trade-press coverage. The engine does not crawl arbitrary URLs or check off a pre-built database — it finds firms where public coverage matches the universe below, then scores how well they actually fit.
- Firm types: media agency; creative agency; digital / performance marketing agency; PR / communications agency; consultancy; professional-services firm
- Size: 50–500 staff
- Geography: UK (primary), US
- Pricing axis: retained-client portfolio count outweighs raw headcount
- · Campaign (campaignlive.co.uk / campaignlive.com)
- · PRWeek (prweek.com)
- · PRmoment (prmoment.com)
- · The Drum (thedrum.com)
- · Prolific North (prolificnorth.co.uk)
- · More About Advertising (moreaboutadvertising.com)
- · Ad Age / Adweek (US agency press)
- · Consultancy.uk / Consulting.us (consultancy trade press)
Not a hard domain whitelist — Perplexity can cite adjacent sources. The hard guarantee is citation integrity: every signal URL must appear in the search results, not be invented.
Runs itself: a daily cron (07:00 UTC) re-runs the live segment — today UK Agencies: UK marketing, PR, creative and media agencies, 50–500 staff, showing client-services and account-movement signals in the last 3 months
The engine grades its own data before anything reaches the board, so what RevOps assigns is already trustworthy — reliable fields personalise, noisy fields only route, and gaps never ship as half-leads. The hard guarantee is citation integrity: every reason-to-reach-out traces to a real source, never an invented one.
| Verdict | Fields | Rule |
|---|---|---|
| Trust | Domain · firmographics · the named event + its citation URL | Machine-verifiable against a live source — safe to put in front of a rep. |
| Treat as noisy | Inferred seniority · role-guessed titles · generic role inboxes (info@, hello@) | Never the reason-to-reach-out; used only to route, never to personalise. |
| Gap | Missing contact · unverifiable event · stale coverage | Routes back to enrichment or gets dropped — it does not reach the queue as a half-lead. |
Control sits with RevOps, not with rep habits. The engine automates hunting, citation-verified structuring, enrichment, scoring, drafting, CRM hygiene and the cost ledger. The human is RevOps: approve at the gate, assign in HubSpot, own the relationship. Because the outcome routes centrally, nothing depends on a rep choosing to open a tool — so whether reps adopt anything or not, the pipeline still lands and still gets worked.
Deferred, with triggers: HeyReach sequencing (after reply-rate proof) · Clay (at team scale, as the orchestration surface) · paid (after message proof). Built one thing that generates signal; consciously refused the other six until they earn their place.
4·Minimal viable tests
hypotheses with kill criteria, not campaigns
| MVT | Hypothesis | Mechanic | Success metric | Kill criteria | Cost |
|---|---|---|---|---|---|
| MVT-1 Outbound | Signal-triggered first touch converts ≥3× a cold list — timing beats persona | Radar sweep → score → human gate → 60–90-word first touch (wired; sends await domain warm-up) | ≥8% reply in 4 weeks; ≥2 demos per 50 sends | <3% reply after 100 sends across two signal types → re-cut the ICP. The engine survives; the hypothesis dies. | ~£0.15/lead + review minutes |
| MVT-2 Content | One authored company voice, fed by live market signal, reads as the category's insider — no rep posting required | RevOps publishes 2 signal-led posts/week from the company channel; engine supplies the angles | ≥5 ICP-title conversations per month from content | <2 after 6 weeks → change format, keep cadence | ~£0.01/post (grounding-checked) |
| MVT-3 Client-Health Index | Owned churn data pulls the exact buyer inbound | Quarterly release, gated download, PR to the same trade press the engine reads | Download → demo ≥3% | One cycle without a sourced demo → demote to blog content | ~1 day/quarter |
5·Seed-stage economics
representative unit costs from the July build · live ledger below
£2.40 ($3.04 USD)
13 scan runs · 19 companies surfaced · 19 signals added · 0 CARE reports · 0 pushed this month
Board total: 146 leads · 165 signals
Breakdown: scans £2.40 · content £0.000 · voice £0.000 · reports £0.000 · pages £0.000
Representative per-unit costs measured during the build (3–4 July 2026):
score a lead (ICP + signals) — API
draft email + LinkedIn note — API
full radar sweep (two search angles) — API
marginal API cost per scored, drafted lead
API is the cheap layer. The real unit cost of a lead is the reviewer's minutes at the gate — so the number that matters is the blended one, not the token bill.
| Cost layer | Per lead | Note |
|---|---|---|
| API — hunt + score + draft | <£0.15 / lead | Measured. Refreshes on the live ledger above. |
| Human review at the gate | ~£2–3 / lead | ≈5 min of a RevOps reviewer at a fully-loaded ~£30/hr — the real unit cost. |
| Tooling seats — Apollo, Perplexity, Anthropic | ~£0.50 / lead | Amortised at low volume. HeyReach / Clay added only once a gated test unlocks them. |
| Fully loaded | ~£3 / lead | Not the <£0.15 API figure — that is the cheapest layer, not the CAC. |
At a target 0.5% lead→customer, one customer costs ~200 reviewed leads ≈ £600 blended variable CAC against a £59k–£108k ACV — so the variable CAC is trivial and the binding constraint is conversion and sales capacity, not media. The real seed-stage bet is the fixed cost of this hire: it pays back the moment the engine keeps AEs fed with qualified pipeline — which is exactly what MVT-1 tests.
Measured vs target: every cost figure here is measured from the live ledger. The 0.5% conversion and the reply-rate targets are hypotheses — not yet proven. No demo has been booked from this engine: sends are wired and deliberately held for domain warm-up, so cycle-1 numbers are the first thing the job produces, not something this artefact claims to have already.
6·If hired — the first fourteen days
Audit what exists: HubSpot state, Apollo plan, domain health. Repoint the engine at production; warm a sending domain.
First 50 signal-led sends through the human gate once the sending domain is warm. Baseline: reply rate by signal type, cost per lead.
MVT-2 live: content cycle from market signals, with Leah on format. Client-Health Index scaffold begun. Kill/scale call on MVT-1.