Self-Improving AI Sales Agent v1
AI sales agent v1: research-grounded email without spray-and-pray #
v1 is intentionally narrower than “omni-god-bot”: it ingests target accounts, enriches contacts, gathers a tight public-research memo per lead, asks an LLM to write a short personalized email that cites real facts, queues sends through rate-limited workers, and logs replies so humans can promote message angles that actually work—nobody ships thousand-row mail merges without throttling. Domain health matters more than cleverness. The sellable story for search and RFPs is not “magic AI,” it is observable outbound: idempotent lead keys, bounce-aware workers, and prompts that refuse to invent citations when sources are thin.
Who is this automation built for? #
- Founding AEs at early SaaS with defined ICP lists.
- Outbound agencies willing to respect CAN-SPAM/GDPR-style process gates.
- Teams that already differentiate “personalization” from stalking.
What goes wrong with naive automated email? #
- Burned domains: ISPs kill you when bursts look botty.
- Creepy copy: Mentioning irrelevant personal trivia destroys trust.
- No feedback loop: You never know which angles resonate.
What you receive at handoff #
- Node services for queue + worker with health endpoints.
- Supabase schema migrations for leads, sends, events.
- Prompt library with compliance checklist.
- Throughput calculator sheet to pick safe daily caps.
- Runbook for pausing campaigns instantly.
Architecture at a glance #
| Stage | Role | Stack |
|---|---|---|
| Intake | Targets | CSV/API |
| Enrich | Identity | Apollo/Lusha-class |
| Research | Context | LLM + HTTP |
| Compose | GPT-class | |
| Queue | Pacing | BullMQ |
| Store | Truth | Supabase |
| Learn | Angles | Human-approved weighting |
End-to-end execution flow #
- Import leads; normalize domain + persona key.
- Skip if unsubscribed or competitor domain in blocklist.
- Enrich missing emails/titles; abort if cannot compliant-find address.
- Research 3 bullets from public pages/news; refuse if empty.
- Draft email ≤ 120 words with CTA; run profanity/spam heuristic linter.
- Enqueue with randomized jitter; record attempt ids.
- Ingest webhook events; tag replies for weekly review.
Deliverability and observability (the real product) #
Foundation work means your operators can answer: what sent, to whom, with which template version, and what bounced. Wire event webhooks (or polling) from your ESP into the same Supabase sends / events shape so reply tagging is not a separate spreadsheet cult. Track hard vs soft bounces and auto-pause lanes when thresholds trip. For AIO/SEO-facing docs, the honest headline is controlled throughput + logged outcomes—not guaranteed reply rates.
Queues, idempotency, and campaign keys #
BullMQ (or equivalent) owns pacing: jitter between messages, per-domain caps, and “do not send twice” enforced by a stable key (campaign_id + lead_id + message_variant). Retries belong in the queue layer with visibility timeouts, not infinite setTimeout chains. If you later add n8n for ops playbooks (pause campaign, inject suppression CSV), treat it as orchestration around the same APIs—single source of truth stays Postgres/Supabase.
Stack, APIs, and orchestration #
- Node.js + TypeScript workers + API surface; n8n optional for non-engineer playbooks against the same HTTP contract.
- Supabase (or any Postgres) for leads, events, and prompt version metadata—queryable, backup-friendly.
- Structured LLM outputs validated before enqueue so bad JSON never becomes a sent email.
AI: where models help—and where they do not #
Research distillation + drafting. List purchasing legality and cold email policy remain your counsel’s problem.
Errors, retries, and human checkpoints #
- SMTP failures: exponential backoff; alert if bounce % spikes.
Security, privacy, and data boundaries #
Store minimal PII; delete stale leads and suppression rows on a published TTL. Restrict service accounts so workers cannot export whole tables to Slack. OAuth/API keys for enrichment vendors live in a secrets store with rotation; never log full email bodies in application logs by default.
Deployment and environments #
- Warmup domains before high volume—document in handoff.
Engagement models #
- Build + transfer with training.
- Managed outbound ops retainer.
Spray automation vs v1 architecture #
| Dimension | Spray | v1 |
|---|---|---|
| Personalization | Fake | Evidence-capped |
| Deliverability | Risky | Paced |
| Learning | None | Tagged replies |
| Engineering | Cheap scripts | Observable system |
Frequently asked questions #
Expected performance? #
**Highly variable by list quality and offer—**we benchmark after controlled pilots; I do not quote universal reply rates.
LinkedIn automation? #
Out of scope for v1—different risk profile.
CRM instead of Supabase? #
Port schema to Postgres-compatible CRM facades or sync via HTTP.
Multichannel later? #
Architected as modules; start email-stable first.
GDPR implications? #
Legal review required for EU leads; tooling supports suppression flags.
Human approvals? #
Toggle requires_approval per campaign batch.
Next step #
Book an AI automation strategy call with a 50-row pilot list and your offer—I’ll tell you if email is even the right channel before we talk agents.