AI Agents for Lead Generation: A Full Worked Example

AI agents for lead generation watch for buying signals, enrich what they find, score fit with auditable math, and hand sales a complete record with evidence attached. This post shows exactly how that works with one end-to-end build: 4,000 raw signals captured in a week, 312 ICP candidates, 214 verified records, and 68 sales-ready leads delivered through an approval gate. Every number is shown so you can substitute your own inputs and estimate the run rate before you commit.

What Do AI Agents for Lead Generation Actually Do?

Strip away the vocabulary and the category is simple: software that notices a company showing buying intent, fills in everything you need to know about it, decides whether it fits your ICP, and hands your sales team a finished record instead of a raw list. The interesting word in that sentence is "decides." Agents, as AWS describes them, pursue goals across multiple steps and adapt as they go, and that is precisely what separates them from the automation that came before.

That earlier generation ran on scrapers and blast sequences: a static list, a five-step cadence, and hope. Here is the kind of catch that model always missed. A mid-market logistics company quietly posts a director of revenue operations role on a Tuesday. No funding announcement, no press, nothing a list vendor would surface for months. An agent watching job boards flags it within hours, because a company hiring someone to fix its revenue plumbing is telling you in public that the budget and the pain both exist right now. We will follow that single posting through every stage of this post.

On AstroFabric, this workflow spans three of the eight specialist agents: market intelligence watches the signals, pipeline handles enrichment and qualification, and content drafts the first touch. The promise here is numbers at every stage, so you can judge the approach on evidence rather than vibes.

The four stages every ai lead generation workflow shares

Capture, enrichment, qualification, handoff. Vendors dress these stages up under different names, but the spine never changes, and each stage exists to make the next one cheaper and more accurate.

Where agents beat scripts and where they should not act alone

Agents win wherever judgment meets volume - deciding which of 4,000 events deserves a second look, for instance. They should never act alone at the boundary where your company touches a prospect. That boundary is where approval gates live, and we will keep returning to it.

The Build: Our Scenario, Stack, and Starting Assumptions

Our scenario: a mid-market B2B software company selling revenue operations tooling, targeting companies that just posted rev-ops job listings. One thing stated plainly before any number appears: everything here is a worked example. The ratios are illustrative and chosen to be realistic, and the whole point is that you can swap in your own volumes and recompute.

The workflow runs inside AI agents for outbound pipeline, one instance of the broader catalog of AI agent solutions. Three operating principles shape everything downstream:

1. Metered tool calls. Every data pull costs credits, so the economics of each stage stay visible instead of disappearing into a subscription.
2. A code sandbox for exact computation. Scoring math runs as actual arithmetic, which means every number can be audited line by line.
3. Approval-gated writes. Nothing touches the CRM and no email leaves without a human saying yes.

The ICP and the signal thesis

The ICP: 100 to 1,000 employees, B2B, a modern CRM in the stack, US or European headquarters. The thesis: a company hiring for revenue operations has diagnosed its own problem and funded the fix. That signal runs warmer than any intent-data topic cluster, because it arrives with a job description that tells you precisely what hurts.

Guardrails before speed: approval gates and metered spend

Gates come first in the build order, before the first signal is ever captured. A team that bolts on review after an agent has embarrassed it in a prospect's inbox never fully recovers trust internally. Start gated, then earn autonomy stage by stage.

Stage One: Signal Capture - From 4,000 Raw Events to 312 Candidates

The market intelligence agent watches four streams: job postings, funding events, tech-stack changes, and hiring velocity. In a typical week for this scenario, those streams produce about 4,000 raw events. That sounds like a flood, and it is meant to be. Capture runs deliberately wide and cheap; the expensive judgment happens later, on a much smaller set.

The funnel math goes like this. Deduplication collapses 4,000 events into roughly 900 distinct companies, because a company on a hiring spree throws off a dozen signals in a week. A first-pass ICP filter on cheap, already-available data - headcount band, region, business model - cuts 900 down to 312 candidates worth spending enrichment credits on.

Our logistics company survives both cuts easily: one posting, one company, squarely inside the ICP band. It is now one of 312.

Which signals earn their keep and which just make noise

Job postings carry this build almost single-handedly, because a posting packs budget, timing, and the problem statement into one document. Funding events add context but fire too broadly to carry a pipeline on their own. Hiring velocity is the quiet workhorse, separating companies growing into a problem from companies churning through staff. Generic "web activity" signals mostly generate noise, so this build ignores them.

Deduplication and ICP filtering before a single enrichment credit is spent

The ordering here is the entire economic argument. Every company filtered out at this stage costs a fraction of a credit; every company filtered out after enrichment has already cost you the full enrichment spend. Metered pricing puts that tradeoff on the bill where nobody can ignore it, and that is healthy.

Stage Two: Enrichment - Turning 312 Names into 214 Complete Records

Now the pipeline agent goes deep on each of the 312: firmographics, tech stack confirmation, buying committee mapping, and, critically, email verification for every contact it proposes to include. This is where credits concentrate, and they should, because this is where raw names become usable records.

The attrition is honest and worth staring at. Of 312 candidates, 61 fail email verification or lack any reachable contact on the buying committee. Another 37 turn out to sit outside the ICP on closer inspection - the headcount was stale, or the "B2B software company" was actually a consultancy. That leaves 214 complete records. A 31 percent drop feels painful until you remember that every one of those 98 companies would have wasted a rep's time or burned sender reputation.

Completeness itself is computed rather than eyeballed. The code sandbox scores each record against a required-fields rubric with exact arithmetic: a record holding a verified contact, confirmed firmographics, and a mapped committee scores complete, while anything missing a required field gets flagged partial with the specific gap named. Our logistics company comes through clean - verified email for the VP of Sales, committee mapped, posting attached as evidence.

Verification before personalization: why bounce risk gets settled here

Personalization spent on an unverifiable contact is pure waste, so verification runs first and everything downstream runs conditionally on it passing. Simple ordering, large savings.

What a complete record contains and what gets flagged as partial

Complete means the original signal, verified contact details, a firmographic profile, a buying committee map, and a completeness score with its inputs shown. Partial records are held with a named gap instead of being discarded, because next week's data often fills it.

Stage Three: Qualification - Scoring 214 Records Down to 68 Sales-Ready Leads

Here is the rubric, weights and all, because automated lead generation with AI only earns trust when a rep can audit the math:

• Signal strength, 35 percent. How directly does the signal point at buying intent?
• ICP fit, 30 percent. How closely does the firmographic profile match?
• Contactability, 20 percent. Verified contact at the right seniority?
• Timing, 15 percent. How fresh is the signal?

Every score is computed in the sandbox as exact arithmetic on those weights - no model vibes, no black box. When a rep asks why a lead scored 84, the answer is four numbers and a multiplication.

The distribution across the 214 records: 68 clear the sales-ready threshold, 97 route to nurture, and 49 are archived with a reason code attached so the decision can be revisited later. Our running example scores 87. A director-level posting sits near maximum signal strength, the ICP fit is clean, the contact is verified, and the posting is four days old. It clears the bar with room to spare.

The scoring rubric, weights and all

Treat the weights above as a starting point and let them drift. Handoff feedback - which leads reps accept, which they bounce back - is exactly the evidence you need to retune them weekly.

Nurture is a destination, and it deserves a real path

Those 97 nurture-routed companies are real future pipeline. They matched the ICP, but timing or signal strength fell short, so they stay on the watch list and re-enter scoring the moment a fresh signal fires. Archiving them would throw away capture spend you already paid for.

Stage Four: How Does the Handoff Work Without Breaking Sales Trust?

Every CRM write and every outbound draft passes through an approval gate. The agent proposes; a human disposes. Most automated pipelines die at exactly this moment, because reps have been burned by tools that dumped junk into their queue. The handoff artifact has to prove itself on sight.

For our logistics company, the packet a rep sees contains the original job posting, the evidence trail from every enrichment pull, the score broken into its four weighted components, and a suggested first touch drafted by the content agent. That draft references the posting directly: the rev-ops mandate, the tooling implied by the job description, and a reason to talk this quarter.

The evidence-first handoff packet

The design principle is simple: a rep should be able to verify the lead without leaving the packet. Every claim links back to its source, which is what turns "the AI said so" into "here is the posting, judge for yourself."

Approval surfaces that meet reps where they already work

Approval happens wherever the rep already lives: the console, a Slack message, a Telegram tap, or an email reply. Review takes seconds instead of a login. Accepted and rejected leads both feed back into next week's scoring weights, and that is how the funnel sharpens every cycle.

What Does This Agentic Lead Gen Stack Cost to Run?

The credit picture follows the funnel's shape. Capture is cheap per event but runs at high volume. Enrichment is the mirror image: low volume, high cost per record, carrying the bulk of weekly spend, which is precisely why the capture stage filtered so aggressively first. Qualification is nearly free because it is sandbox computation on data already paid for, and handoff costs a little for drafting.

The comparison with seat pricing matters more than any single number. When cost tracks work performed, trimming spend means trimming waste: tighten the pre-enrichment filter, drop a noisy signal source, raise the completeness bar. Every optimization shows up on the bill.

Where credits concentrate and how to trim them

Enrichment. Always enrichment. The lever sits upstream: every candidate the ICP filter removes before enrichment saves the full per-record cost. A stricter filter that cuts 312 to 250 with minimal lost fit is the highest-ROI change available in this build.

A back-of-envelope sizing formula for your own volume

Take your weekly signal volume, apply your expected dedup ratio, then your ICP pass rate, and multiply the survivors by enrichment cost. That product will be most of your run rate. Then treat it as a hypothesis: pilot with your own inputs for two weeks before committing, because your ratios will differ from this worked model and the pilot tells you exactly how.

AI Agents for Lead Generation Beyond This One Build

Zoom out and the capture-enrich-qualify-handoff spine turns into a general pattern. An agency can point it at brands showing website-refresh signals; an ecommerce operator can point it at wholesale buyers. The stages hold while only the signals and the ICP change. The same operating model powers adjacent work across the stack too - content briefs built from demand data, competitive teardowns, AI visibility tracking. Lead gen is one instance of a much broader approach to agentic work.

Adapting the spine to other ai agents for business use cases

The transferable skill is funnel thinking: wide cheap capture, narrow expensive enrichment, computed qualification, gated handoff. Once your team has run the loop for lead gen, applying it elsewhere feels obvious. For the mechanics one level deeper, read how AI prospecting works from signal to sequence, and when you are ready to think about the outreach side, see how AI SDR agents run autonomous outbound.

Your first-week checklist

Run the Numbers on Your Own Pipeline

Every figure in this post was a worked model, and the honest next step is to replace it with your own. Point the pipeline and market intelligence agents at one signal source, keep the approval gates on, and watch your own funnel math take shape over two weeks. Start with AstroFabric and let the first handoff packet make the argument for you.

Frequently asked questions

Sources

• What are AI agents? - AWS
• AI agents explained - TechTarget