I've spoken to about 100 companies about their GTM, just looking at how they're set up and how they're using agents. A lot of B2B companies, some B2C too. And along the way I've noted a few key things that separate the companies who are just throwing tokens down the drain from the ones actually getting 10x ROI. this blog (or video if you prefer below) is a my guide on what you need to know to build good b2b lead gen agents.

So here's what we're going to cover:

as you can see I'm not going to list every lead gen tactic that works with agents. That could be a whole video on its own. This is about the setup and the mindset, the part that transfers to any channel as you scale.

It's the sentiment I see all over X where people selling the idea of fully automated lead gen machines. And this isn't to say agents aren't good at lead gen. They are really good, and you can see the uptake from 2025 to 2026 is massive.

The data also says the same thing. 74% of GTM teams now use AI for lead generation (ICONIQ, State of GTM 2026).

And yet 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before (S&P Global). Everyone's buying the tools and most are getting nothing back.

If I share some definitions and concepts up front, the clarity of execution later will land a lot better.

Before we get into the actual setup, let’s level-set on vocabulary so everyone is working from the same page and nobody is left behind.

You need to understand that speed is very important. Every action you can take faster helps the whole process move faster, so a lot of the time with agents I am optimising for more speed, more experiments and more numbers, because at the end of the day it’s a numbers game.

We are all leads in someone’s pipeline and no one likes being spammed or harassed. With AI, yes, you can spam everyone and you might get one or two leads, but I think it’s better to just be a good actor and good things will happen to you.

quick disclaimer: I'll be talking about a mix of tools throughout, some I built myself for clients, some from other companies but the same practices and the same way I use these tools can be applied to pretty much any harness.

Before we find a single lead, two things have to exist somewhere for the agent to think and a clear definition of who's worth chasing.

The AI is only as smart as the context you give it. There are a few ways to build a brain for these systems. Agents need somewhere to keep what they know. Minimum viable version like a CRM. I run that plus a context graph, which stores the decisions my agents make so the system understands what I'm doing over time.

And keep it clean, because B2B contact data decays around 2% a month once you count job changes, title changes and dead emails (industry estimates). An agent pointed at a rotten CRM doesn't fix the rot. It automates it.

Once your data is set up, you can define what a high-intent lead actually is for you. Maybe they visited your pricing page, hired for a relevant role, and made a post on social and that's a tier 1 lead. A tier 2 lead is a mix of that, but the signal was weaker. Then tier 3, then unqualified.

That tier system is the main thing we give the agent, so it knows exactly how to react to the priority of each lead. Here's how I build the rubric:

Now we know how to score a lead. So how do we find the high-intent ones? Quick definition, because "lead gen" and "prospecting" get used interchangeably.

I'll do a bit of prospecting, but I'm focused on lead gen and attracting people who might be interested. And the key term is signals having data that tells you who's in the market and ready to be marketed to.

The best leads usually already exist in your data: newsletter sign-ups, webinar attendees, website visitors, closed-lost deals that went quiet. Point your agents inward before you buy lists. Tools like RB2B do person-level identification of US website visitors, which is about as first-party as an intent signal gets.

My favourite signal is social where someone makes a post that links to your offering. Here's a custom system doing exactly that where I am monitoring subreddits for posts relevant to my offering, weighting the ICPs, scraping Reddit, and scoring buying intent against the rubric:

Hiring spikes, funding rounds, tech installs, job changes. The triggers that happen outside your data and tell you a prospect just went in-market.

Clay's Custom Signals lets you watch nearly any market event and trigger a workflow off it. And Workflows.io's signal architecture guide is the clearest blueprint I've seen for wiring signals into if-then plays.

This is the map I work off. First-party is what happens inside my own systems, second-party is what happens around my network, third-party is what's out in the market. I trust them in that order, because the closer a signal sits to my pipeline, the stronger it usually is.

Inbound is about earning attention rather than chasing it and really, earned attention is earned trust. You give something genuinely useful away free, and the right people come to you. That's a lead magnet. Technically, you're in one right now.

I'll show how I use AI to generate content from blogs I've already made by turning one blog into LinkedIn posts as part of a general strategy. Ideally you want to funnel readers in some kind of order toward lead management. The point is consistency because one viral post does less for pipeline than a boring engine that publishes and captures every week.

There are basically 3 types of lead magnet, from that famous Alex Hormozi video, which I agree with in theory:

The reason to tie the magnet back to your rubric is because a magnet that only reveals the problem pulls in early-stage readers, while a free trial or free step 1 tends to pull people who are closer to buying.

Personalised follow-up on those leads matters too, since personalised emails run about 6x higher transaction rates than generic ones. So the format you pick should map to the tier you want to fill, not just to what gets the most sign-ups.

What we can do is use agents to work out which of these formats fits your audience, then build the lead magnet and tie it back to your high-intent rubric so inbound gets scored the moment it comes in.

Outbound is where personalisation at scale really pays off. Instead of spraying templates, we let agents research each prospect and trigger outreach off real signals, so you reach out when someone is actually in-market.

This is where agents should really shine with personalisation at scale. Agents researches each prospect and writes 1:1 outreach, and it triggers off the signals in your rubric so you only reach out when someone is actually in-market.

Two rules straight from the data.

First is that don't pitch in the first email as it cuts replies by up to 57% (Gong). Lead with relevance, pitch later. Second, from Kyle Poyar, before you automate a single message, answer one question. What market event makes this person a relevant target right now? No signal hypothesis, no automation.

McKinsey pegs personalisation's typical lift at 10-15% of revenue, and Unify's CEO claims their signal-timed, human-approved outbound runs at 160% of industry standard. His number, but the architecture matches everything above.

Don’t forget your existing customers. Agents can watch your current base for upsell and expansion signals, like heavier usage or new hires, and flag the accounts worth a conversation. This is the cheapest pipeline you have, so it is often the fastest win.

So now we need to look at how we scale this whole process up and build in some understanding of housekeeping. Scaling multiplies everything, the wins and the mistakes at the same rate. Two habits keep the engine from rotting.

The system is only as good as what you feed it. For me that means dropping updates into Slack as I go, and the system prints them into its context. That's how it knows the things I'm doing that it can't scrape on its own.

11x raised from a16z, promised autonomous AI SDRs, and per TechCrunch ended up with customers claimed that weren't customers, churn reportedly at 70-80%, and AI emails landing in spam. That's what ungoverned looks like.

The guardrails aren't optional anymore. Gmail and Yahoo cap spam complaints at 0.3% for bulk senders. An ungoverned agent can burn your domain in days.

Jason Lemkin's take, after actually running agent-heavy GTM the AI SDRs work if you QA them daily and manage them like reps. Almost nobody does that but it will only help you keep going with more speed and better brand.

The end state is agents that hold a working model of your market. Instead of reacting to signals one at a time, they keep a picture of how your buyers, competitors and channels behave, so they can plan ahead.

Just like chess you play out the likely sequence of moves before you commit, so you can test how a campaign lands before you spend real pipeline on it.

My big conclusion on using AI agents for B2B lead gen comes down to four things.

If you’ve got this far you should DEFINITELY subscribe to my newsletter here.

thank you.

roman

(If you want systems like this for your team, i build custom GTM engines at Scale Intelligence, feel free to book a call here: https://www.scaleintelligence.dev/)