Live in production
A B2G outbound pipeline where automation does the work and judgment does the deciding.
Built for long-cycle armoured vehicle sales into government and defense-adjacent markets — an environment where one careless email costs more than a hundred good ones earn. Every design decision below follows from that constraint.
What changed
- Before
- Hours of manual research per batch, and no clean handoff from research to outreach.
- Now
- Sourcing, AI qualification and human approval in one event-driven pipeline. One opportunity from the first 21 emails.
Production telemetry
Queried from live tooling, July 20, 2026
- Workflow executions
- 79
- Success rate, last 50 runs
- 100%
- First emails → positive replies
- 21 → 1
- Opportunity created
- 1
These are real figures from a deliberately low-volume launch phase: 15/day cap, 1 bounce, open tracking disabled by design (see Decision 02). Small numbers, honestly reported — the system is built to scale volume only after deliverability is proven.
The problem
Sourcing B2G prospects meant hours of manual research per batch, inconsistent record quality, and no clean handoff between research, qualification, and outreach. Sales attention was going into copy-paste work instead of conversations.
The constraint
Defense-adjacent outreach cannot tolerate spray-and-pray. Wrong targets, burned domains, or sloppy personalization carry reputational cost far beyond marketing. The system had to be fast and conservative — which is exactly the tension the architecture resolves.
The pipeline.
- Campaign config submitted from ops console
- Prospect search via data provider API
- Field validation + combined mapping
- Extraction guardrails (rate limits, retries)
- AI qualification against ICP rules
- Airtable review queue — human approval
- Event-driven push to sending platform
- Sequenced outreach, stop-on-reply
- Merged live reporting endpoint
Six decisions that define the system.
Anyone can wire APIs together. The system is the set of choices about what not to automate, what not to track, and what not to send.
- 01
A human approval console sits between AI and send
Qualified leads land in an Airtable review queue. Nothing is contacted until a person clicks approve; the click fires an event-driven POST that pushes the lead to the sending platform.
Why not the obvious alternative
Full automation was the obvious build. In defense-adjacent B2G, one badly-targeted email to a government contact costs more than a week of manual review saves. Judgment stays where it matters; automation handles everything around it.
- 02
Open tracking is switched off — on purpose
The live campaign config has open_tracking and link_tracking disabled. Reported opens on this system are zero, and that is correct.
Why not the obvious alternative
Tracking pixels and rewritten links raise spam-filter risk and inflate dashboards with unreliable data. For a low-volume, high-consideration pipeline, the only metrics that matter are replies and opportunities — so those are the only ones measured.
- 03
Volume is capped, plain-text first, stop-on-reply
Daily send cap of 10–15, first email is text-only, sending restricted to UAE business hours (Sun–Thu), and the sequence halts the moment a lead replies.
Why not the obvious alternative
A 100/day blast would produce bigger vanity numbers and a burned domain. Sender reputation is the pipeline's scarcest asset; every config choice protects it.
- 04
Event-driven, not polling
The approval console POSTs directly to a webhook on click. No scheduler loops asking "anything new?" every minute.
Why not the obvious alternative
Polling wastes executions, adds latency, and creates race conditions on double-approval. The event model means the recent execution log reads as a clean list of sub-second successful runs.
- 05
Guardrails before AI, validation before guardrails
Extraction runs through field validation and rate-limit-aware guardrails before any AI qualification step. Batch continuation uses a search-signature so re-runs never re-process the same segment.
Why not the obvious alternative
LLMs applied to dirty data produce confident garbage. Cleaning first means the qualification layer scores real records — and the dedup signature keeps API spend flat as the pipeline scales.
- 06
Two sending platforms, one reporting endpoint
A read-only n8n workflow fetches live campaign analytics and mailbox health from both sending platforms and merges them into a single JSON response for the ops console.
Why not the obvious alternative
Two dashboards means two versions of the truth. One merged, read-only endpoint (no AI calls, no write access) gives management a single view without another SaaS subscription.
Stack
- n8n
- Apollo API
- Apify
- Airtable
- Instantly
- Smartlead
- Claude / LLM APIs
- Webhooks / REST
Want the deeper walkthrough?
I can walk through the workflow architecture, the qualification prompt design, and the deliverability setup live — with the actual system on screen.