AI Lead Generation: What We Built, and What Actually Works

Most agencies selling AI lead generation are reselling someone else's tool with a markup. We built ours — Agentic Leads — because the off-the-shelf options each did one part of the job well and the rest badly.

This is how the pipeline actually works, including where it goes wrong.

The five stages

Lead generation is not one problem. It is five, and most tools solve two of them.

1. Finding

Sourcing businesses that match a defined profile — industry, location, size, and signals suggesting they might need what you sell.

Our platform pulls from Seek job listings, Google Maps exports and Yellow Pages. Job listings are the underrated source: a business advertising for a receptionist is telling you something useful about its current constraint.

Where this fails: scraping without a defined profile. Volume is easy and worthless. A list of ten thousand businesses with no qualification is worse than two hundred good ones, because it trains your team to ignore the list.

2. Enriching

A business name and phone number is not a lead. You need a contact, a verified email, and enough context to say something relevant.

Enrichment adds email addresses, direct contacts, ABN verification and technology stack detection. That last one matters more than it sounds — knowing a business already runs a particular scheduling tool tells you whether your integration story is easy or hard before you ever call.

Where this fails: stale data. Contact details decay at roughly two to three percent a month. An enrichment run from six months ago is materially wrong today.

3. Scoring

Not every lead is worth the same effort. Scoring sorts them into hot, warm and cold against criteria you configure — and the configuration is the important word. Scoring rules that work for a plumbing supplier do not work for an accounting firm.

Where this fails: scoring on the wrong signals. Company size is the default and it is usually a poor predictor. Recency of a relevant event — a job ad, a new location, a funding round — predicts far better.

4. Outreach

Multi-channel: cold email, AI voice calls, LinkedIn. The platform runs all three because different industries respond to different channels, and assuming everyone responds to email is how campaigns quietly underperform.

Where this fails, and this is the big one: volume without relevance. Sending three thousand generic emails will get you a spam complaint rate that damages your sending domain for months. Australian Spam Act obligations also apply — consent, identification, and a functioning unsubscribe are not optional, and the penalties are real.

5. Reply handling

The stage almost every tool ignores. Replies arrive and someone has to read them, sort them, and act.

Ours classifies replies using Claude — positive, negative, question, objection, unsubscribe — and routes accordingly. Unsubscribes are processed immediately and automatically, because doing that manually is how businesses end up in breach.

Where this fails: treating a question as a rejection. A meaningful share of "not right now" replies are genuine interest with bad timing, and they are the cheapest pipeline you will ever have if you follow up in three months rather than never.

What it is honestly good at

Consistency. A human doing outreach has good weeks and bad weeks. A pipeline does not.

The unglamorous middle. Enrichment and reply sorting are the parts people skip when busy, and skipping them is why most outreach underperforms.

Learning what works. Because everything is logged, you can see which industries reply, which openings work, and which channel suits which segment. Most businesses running manual outreach have no idea, because nobody recorded it.

What it is not good at

Relationship selling. If your sales cycle involves multiple meetings and a relationship built over months, automation gets you the first conversation and nothing more.

Very small addressable markets. If there are 200 businesses in Australia who could buy from you, do not automate. Research each one properly and approach them individually. Automation is for when the list is too long to handle personally.

Fixing a weak offer. Outreach amplifies. If the proposition does not land, sending more of it produces more rejection, faster.

The compliance part people skip

Australian Spam Act 2003 requirements apply to commercial electronic messages:

• Consent — express or inferred. Inferred consent from a publicly listed business address is narrower than most people assume.
• Identification — who you are and how to contact you, clearly.
• Unsubscribe — functional, honoured within five working days, and it must not require the recipient to log in anywhere.

Privacy obligations apply to the data you collect and hold on those contacts as well.

A vendor who does not raise this is either unaware of it or unbothered by it, and both are a problem when the penalties land on you rather than them.

A realistic expectation

For an Australian B2B business with a defined target market and a proposition that already converts in person, expect reply rates in the low single digits on cold email and meaningfully higher on well-targeted voice.

Anyone quoting twenty percent reply rates is describing a warm list, a very narrow niche, or fiction.

The businesses that do well with this treat it as a way to have more of the conversations they are already good at. The ones that do badly treat it as a substitute for having a proposition.