Overview

End-to-end B2B sales-acquisition automation for a German company. Built to research, score, email, call, and close prospect companies from a single data model and a single benchmark score, running one funnel end-to-end: Research → Campaign assignment → Outreach → Connect → Closing.

Guiding principle

Build the glue. Buy the commodities.

Dialers, email transport and inbox-warmup infrastructure, Maps data, crawling, e-signature and payment rails are solved, mature products. Rebuilding them is a multi-year arms race with no payoff. What doesn't exist as a single product — and what makes this platform worth building — is the connective tissue:

1. Funnel orchestration — one automated pipeline from research to close, with humans only where quality matters.
2. Research + benchmark engine — target discovery (Serper.dev Maps by postal code + industry, enriched from several sources) plus a single 0–100 benchmark computed once and flowing through all stages.
3. Sender management — the VirtualCompany / VirtualEmployee layer that owns synthetic sender identities, warmup orchestration, and the bandit that picks the best content variant.
4. Call-center workforce orchestration (shifts, gatekeeper KPIs, per-agent commissions, benchmark-driven call packages).
5. A unified target lifecycle that ties Outreach → Connect → Closing into one state machine and one data model, plus cross-stage controlling.

Everything else, we integrate. This decision is the difference between a short MVP and a 12+ month rebuild.

The funnel

Research  →  Campaign assignment  →  Outreach (email)  →  Connect (calls)  →  Closing
   │                │                     │                    │                 │
   └────────────────┴──────── one benchmark score ────────────┴─────────────────┘
   └──────────────────── one target lifecycle state machine ────────────────────┘

Who it's for (personas)

• Supervisor / admin (the product owner): defines campaigns (postal codes + industry + budget), sets assignment rules, schedules call-center shifts and releases, watches cross-team controlling.
• Connect-Agent: works a focused cockpit — the current target's full context (research data + benchmark + email history), the script, a disposition form; CloudTalk's power dialer auto-advances to the next call. Sees only released campaigns.
• Closing-Agent: works a drag-drop pipeline of interested targets → appointments → contracts. Sees only their round-robin-assigned targets.
• Owner / domain experts: want the cross-stage overview — cost per stage, yield per source, per-agent performance, budget health.

Scope

Core system — what we build

1. Target research pipeline — queue-driven, budget-governed discovery (Serper.dev Maps) + enrichment (Firecrawl website crawl, Meta ads, Wayback web-archive age, job-portal signal), with per-fact provenance.
2. Benchmark scoring engine — versioned, pluggable, one 0–100 score across 5 weighted dimensions, with a stored breakdown.
3. Campaigns & target lifecycle — campaign CRUD, targets belonging to 0..1 campaign (re-assignable), per-target lifecycle state machine.
4. Outreach — VirtualCompany/VirtualEmployee sender identities, Microsoft Graph / Exchange Online transport, Trulyinbox warmup, a 3-mail sequence, bandit variant selection, reply processing.
5. Connect / call center — Connect-Agents, shifts, benchmark-driven call packages, call cockpit, gatekeeper/interest KPIs, per-agent commissions. The biggest market gap — worth extra polish.
6. Closing — drag-drop pipeline, appointments, notes, contracts (SignRequest), payment collection (GoCardless), no-show tracking, closing commissions.
7. Controlling — cross-team read-model dashboards + budget allocation optimization.

Non-goals (explicitly out, at least for now)

• Building our own email transport or telephony stack (Exchange Online / Microsoft Graph and CloudTalk are fixed).
• Running our own crawler against Google Maps (use Serper.dev; Firecrawl for site crawl).
• Reputation monitoring for virtual employees.
• A full GDPR / spam / UWG compliance layer (design consideration only in v1).
• General-purpose CRM for arbitrary industries; AI presentation generation; closing QA.

Build vs. integrate — component matrix

Locked product decisions

The layer that used to be a scope risk is now core: rather than integrating a third-party deliverability/persona product, this platform owns the sender layer — VirtualCompany / VirtualEmployee identities, Trulyinbox warmup, vacation/sick-day randomization, and the bandit that optimizes content. Transport is delegated to Exchange Online / Microsoft Graph, so the heavy lifting is mail content and orchestration, which is exactly the differentiator.

Glossary

• Target — a prospect company (the word "Lead" is never used). A Target carries the lifecycle state, the benchmark score, and its research data. A Target belongs to 0..1 Campaign (it may be "unassigned"; re-assignment is supported). There is no separate deduplicated "Company" master reused across many campaigns — the Target is the company.
• Contact — a contact person at a Target (1:n).
• Benchmark — a normalized, versioned 0–100 score (with a stored breakdown) computed once per target from research signals across 5 weighted dimensions; the single sort key for outreach ordering, call prioritization, and shift assignment.
• Research data point — a fact gathered about a Target (paid Meta ads, website age, hiring activity, industry, size proxy), stored in research_data_points with provenance (source, retrieved_at, confidence) and feeding the benchmark.
• Campaign — a run definition: postal codes + industry + budget + assignment rule + product, bundling OutreachVariants and ConnectVariants.
• Assignment rule — the per-campaign rule that decides which researched targets attach to the campaign: a threshold rule (numeric benchmark cutoff) or an LLM prompt rule (natural-language matching).
• VirtualCompany / VirtualEmployee — synthetic sender identities (1 company → 1..3 employees) the platform manages to send Outreach mail via Exchange Online.
• Bandit — a multi-armed bandit (Thompson Sampling default, epsilon-greedy alternative) that selects the best StepVersion at send time and learns from reply outcomes.
• Budget guard — a counter checked before every paid external call; pre-checks against the campaign budget, warns at 80%, and auto-pauses the campaign at 100%.
• Shift — an assignment of agent × campaign × datetime × target-count that falls inside a ConnectVariant's weekday/time window; on activation the system locks the top-N unworked targets by benchmark into call Packages for the shift.
• Gatekeeper quote — call-center KPI: rate at which agents get past the gatekeeper to a decision-maker.
• Provision — per-agent commission derived from call/close outcomes.
• Warm handoff — a non-cold follow-up handed from the call team back for closing; a distinct job type from cold-call shifts.

Legal & compliance

Not optional throat-clearing — it shapes the architecture. Raise before Phase 2 (Outreach). Note: a full GDPR/spam/telemarketing compliance layer is explicitly out of scope for v1 — the items below are design considerations, and the deliverable is a written legal opinion, not a built compliance module.

• GDPR (EU/German). Researching business contact data, storing named contact persons, and cold-emailing them all touch GDPR. Need a lawful basis, retention limits, deletion handling, and a record of processing. Bake right to erasure into the data model from day one.
• UWG (German unfair-competition law). Unsolicited B2B cold email generally requires prior consent; unsolicited cold calling of businesses has its own constraints. Directly affects how Outreach and Connect flows may legally operate.
• Sender-identity design. Fabricated VirtualEmployee personas carry legal exposure — keep this in view when designing the sender layer.
• Provider ToS. Google Maps and Meta both restrict scraping; use the official/licensed APIs (Serper.dev, Meta Ad Library) rather than raw crawling.

Recommendation: get a short written opinion from a German data-protection/UWG lawyer on the Outreach + Connect design before building those modules. Cheaper than rebuilding after launch.

Open questions for the client / domain experts

1. Sender layer scope: confirm the depth of the VirtualCompany/VirtualEmployee simulation (vacation/sick-day randomization, per-employee send caps) for v1. ⚠️ Needs client sign-off.
2. Maps access route: Serper.dev as the primary ID source is locked — confirm budget ceilings for paid, on-demand sources (Firecrawl, TheirStack, handelsregister.ai).
3. Closing CRM: confirm we build it (recommended) vs. bolt on an external pipeline tool.
4. Multi-tenancy: one internal tool, or will it serve multiple client organizations later? Current design: fork the instance per tenant. ⚠️ Needs client sign-off.
5. Legal sign-off: who owns getting the GDPR/UWG opinion, and by when?
6. Benchmark definition: which signals matter most? Domain experts' 15+ years of experience should seed the initial 5-dimension scoring weights.

See also: 01-architecture · 02-roadmap · 03-data-model · modules/