AI Outbound Sales Pipeline for a B2B Technology Consultancy
How a guarded AI outbound system cut founder prospecting time from 10-15 hours to 1-2 hours a week
Executive Summary
Founder outbound time cut: 10-15 hours/week down to 1-2 hours
Researched, personalized sends per week
New leads at steady-state campaign intake
Per quarter avoided versus an SDR-plus-agency alternative
The Challenge
A founder-led B2B technology consultancy needed a consistent flow of qualified conversations with mid-market US prospects, but the entire outbound process lived inside the founder's calendar. Building lists, researching accounts, judging fit, writing openers, loading campaigns, watching replies, and updating spreadsheets all competed directly with delivery work. The work was essentially an SDR's job description, but the company was too lean to justify an SDR hire or a traditional outbound agency retainer. The first attempt at AI personalization also created a credibility problem: unguarded AI fabricated social proof and used hedged language such as 'probably' and 'likely,' making outreach sound guessed rather than researched. The business needed an outbound system that could scale volume without creating reputational risk — one that produced genuinely grounded personalization, failed loudly when quality dropped, and gave the founder clean go/no-go evidence for new market experiments.
Key Pain Points
- ✕Every outbound cycle assembled manually inside the founder's calendar
- ✕Unguarded AI fabricated social proof and hedged language, risking credibility
- ✕Prospecting volume capped by founder hours, warm replies missed in the inbox
- ✕Too lean to justify an SDR hire or agency retainer
Our Solution
Amasa built a scheduled AI outbound machine covering search, qualification, personalization, sending, reply classification, Slack notification, and campaign audit — a daily pipeline with preflight checks, deterministic gates, model-assisted classification, and a human handoff only where founder judgment matters. The pipeline starts with API-credit and deduplication checks, then pulls company candidates from a B2B contact data API. Free rule gates filter on headcount, industry tags, red-flag keywords, and email-gateway reachability before any paid export step. GPT-4o mini classifies company and contact fit under a hard 'in doubt, say no' rule. Claude Haiku generates personalized openers, but quality does not rely on prompt obedience: deterministic code guards verify that every number, magnitude word, and proper noun in the opener appears in the verified input data, and banned phrases or hedge words trigger a regenerate-then-fallback loop, so fabricated claims cannot ship. Leads are logged to Google Sheets and dispatched to SmartLead through API merge fields. Webhooks classify every reply into seven categories — interested, question, not now, referral, out-of-office, unsubscribe, and hostile — then alert Slack with a suggested response. A weekly audit job compares cumulative stats against pre-committed go/kill criteria, and campaigns are defined in YAML so new verticals launch as configs, not build projects. The whole system is backed by 1,100+ automated tests.
Implementation Approach
Pipeline & Gating Design
Designed the daily pipeline with preflight checks, free rule gates before paid enrichment, and 'in doubt, say no' fit classification
Guarded Personalization
Built deterministic anti-fabrication guards verifying every number, name, and claim against source data, with regenerate-then-fallback loops
Sending & Reply Automation
Integrated SmartLead sequenced sending, seven-category reply classification webhooks, and Slack alerts with suggested responses
Audit & Experimentation Layer
Shipped weekly go/kill campaign audits, YAML campaign configs, and 1,100+ automated tests with verify-by-readback discipline
The Results
The implementation delivered transformative results across all key metrics, with immediate impact on operational efficiency, accuracy, and customer satisfaction.
Impact Metrics
| Metric | Before | After | Improvement |
|---|---|---|---|
| Founder time | 10-15 hours/week | 1-2 hours/week | 80-90% reduction |
| Emails per week | A few dozen manual sends | 400-500 researched, personalized sends | ~10x volume |
| Reply handling | Inbox watching, manual triage | Auto-classified into 7 categories with Slack alerts | Fully automated |
| Quality control | Manual proofing or none | Deterministic guards plus 1,100+ tests | Fabrication blocked by code |
Key Takeaways
- Anti-fabrication belongs in deterministic code, not prompts — every number and name is verified against source data before an email can ship
- Free rule gates before paid enrichment steps keep API costs proportional to genuinely qualified volume
- Seven-category reply classification with Slack alerts means warm opportunities surface instantly instead of sitting in an inbox
- Pre-committed go/kill criteria turn market experiments into two-week decisions instead of months of drift
Inside the Engagement
The Workflow Before
- Build target lists from ad-hoc sources.
- Research each company and contact manually.
- Judge ICP fit from public signals.
- Write a personalized opening line for each prospect.
- Load contacts into a sending tool by hand.
- Monitor the inbox for replies and categorize them manually.
- Update campaign tracking sheets after the fact.
This capped prospecting volume at the number of hours the founder could spare. Manual outreach was trustworthy but slow. Unguarded AI outreach was fast but unsafe. Reply handling depended on someone watching the inbox, which meant warm opportunities could be missed or delayed.
The Workflow After
- A daily cron runs preflight checks for credits, deduplication, and campaign readiness.
- Company search pulls candidates matching the target profile.
- Rule gates remove poor-fit companies before paid enrichment.
- GPT-4o mini classifies surviving companies and contacts.
- Verified emails are exported, fail-closed deduplicated, and capped at one contact per firm.
- Claude Haiku writes the opener while deterministic guards block hallucinated numbers, names, or social proof.
- Leads are logged to Google Sheets and pushed to SmartLead.
- Reply webhooks classify every response and notify Slack with a suggested next step.
- Weekly audits judge campaigns against pre-committed kill or extend criteria.
Full Results
| Metric | Before | After |
|---|---|---|
| Prospect research and qualification | Manual, founder hours per prospect | Fully automated daily screening, gating, and classification |
| Personalization | Manual writing or unsafe unguarded AI | Seconds per prospect with deterministic anti-fabrication guards |
| Sending volume | Capped by founder hours | 25 new leads/day and ~75 sequenced sends/day per campaign |
| Emails per week | A few dozen manual sends in a good week | 400–500 researched, personalized sends |
| Reply handling | Inbox watching and manual triage | Auto-classified into seven categories with Slack alerts |
| Founder time | 10–15 hours/week | 1–2 hours/week for reports, spot-checks, and replies |
| Quality control | Manual proofing or none | Deterministic guards plus 1,100+ automated tests |
| Cost avoided | SDR hire plus agency retainer | $10K+ per quarter avoided |
The Impact
The pipeline moved the founder from producing outbound to judging outbound. Research, qualification, personalization, sequencing, and reply triage run automatically, while the founder focuses on final replies and campaign decisions.
The system also protects brand credibility. Because fabricated claims are blocked by code rather than merely discouraged by prompts, prospects receive grounded, researched outreach instead of generic AI copy. A zero-lead day cannot be accidentally reported as healthy – the system fails loudly in Slack.
The strategic benefit is experimentation speed. New verticals launch as YAML campaign configs rather than custom build projects, and market wedges are judged with clean evidence in roughly two weeks instead of drifting for months.
The Takeaway
Amasa turned outbound from a founder-time bottleneck into a daily AI-operated sales pipeline – reducing weekly founder effort by roughly 80–90% while preserving factual integrity and campaign-level decision discipline.
Quick Facts
Industry
Professional Services
Solution Type
Process Automation
Published
August 24, 2026
Technologies Used
Related Resources
Ready to Achieve Similar Results?
Let's discuss how we can transform your business with AI solutions tailored to your specific needs