AI Lead Generation: The Funnel I Actually Run, With Receipts

Search for ai lead generation and you will find the same article eight times. A numbered list of tools. A paragraph about personalization at scale. A call to action for whichever tool the publisher happens to sell. The number one result on Google right now is a Reddit thread complaining that there are too many sales and lead gen AI tools. When a complaint thread outranks Salesforce and IBM, the market is talking. It has plenty of tools and no method.

I am a marketer, not a vendor. I have worked with AI every day for two and a half years, in my day job at SSI Schäfer, a Swiss intralogistics company, and in my own consulting practice. I wrote the full account of those years here. This article is the lead generation slice, and it contains no tool list. It contains the funnel I actually run, with the receipts behind it.

One premise before those receipts. A lead is a person deciding whether to trust you. AI does not create that trust. It speeds it up. The machine sorts, classifies, drafts, and removes waiting time. The promise, the price, and the responsibility stay with you. Every failed AI lead generation project I have watched up close confused those two halves, so this article keeps them apart on purpose.

Key Takeaways

• AI lead generation is a funnel job, not a tool job. The listicles rank tools. The work happens at five stations: qualification, the form, attribution, follow-up, and channel decisions.
• The best conversion win I shipped was a shorter form. Five fields, Stripe Checkout, and a webhook that seeds the customer’s draft listing: 9 files, +164 and -506 lines, 97 unit tests. The new funnel is literally smaller than the old one.
• First-touch attribution is the least glamorous, highest leverage piece. My kit tracks 44 attribution vectors and classifies every lead source identically on client and server.
• AI does not create trust. It speeds it up. Sorting, drafting, and classifying drop from hours to minutes. The promise and the price stay human.
• A new channel is not a free lead machine. I evaluated ChatGPT Ads in August 2026 and said no at a 600 CHF budget: contextual and geo targeting only, and a meaningful B2B test starts at 5-10k USD per month.
• My cheapest lead source compounds. This blog and its content built for AI search bring consulting leads while I sleep. Ads stop the moment you stop paying.

Why Every AI Lead Generation Listicle Sounds the Same

I pulled the top results before writing this, because I wanted to know what a searcher actually gets. The answer: a Reddit thread about tool overload, two enterprise explainers from Salesforce and IBM, and five pages from companies that sell lead generation software. Six of the eight results rank a category they profit from. Nobody shows a live form. Nobody shows an attribution table. Nobody shows a decision not to spend money.

The structural flaw is tool-first thinking. If you start from which tool to buy, every problem looks like a missing subscription. Start from the funnel instead and the question changes: where, exactly, does a machine help turn a stranger into a conversation? I count five places, and only one of them involves buying anything. The strategy layer above all of this lives in my AI marketing playbook.

The Five Places AI Touches a Lead

Here is the map I work from, top to bottom.

Qualification. A lead arrives as an email, a form submission, a message. The first job is sorting: does this fit, how urgent is it, who should see it. AI drafts the sort. I confirm it.

The form. The conversion moment itself. Mine got rebuilt around a simple idea: ask less, automate the handoff. The result converts better and contains less code.

Attribution. Where did this person come from. Forty-four signals captured at first touch, classified the same way in the browser and on the server.

Follow-up. The reply that decides whether an inquiry becomes a conversation. AI drafts it in minutes. I review and send.

Channel calls. Where to spend and, more often, where to refuse. AI informs the analysis. The refusal stays mine, and you will read about one below.

Four of the five stations run on tools I already owned. That is the quiet truth of ai lead generation: the compounding wins are structural, not purchased.

Lead Qualification: Let AI Sort, and Let a Human Judge

Inbound leads arrive messy. A one-line email. A form with three fields filled and two abandoned. A message that says only “is this still available”. Before AI, sorting a batch of those meant an afternoon of opening tabs and pasting into a spreadsheet.

Now the sorting pass takes minutes. The model reads the raw inquiry and drafts a classification: source, stated need, language, rough fit against my ideal client profile, urgency signals. Every draft lands in a queue. Nothing sends itself. I review, correct maybe one label in ten, and decide who gets a reply today.

The line I hold is this: AI prepares the decision, it never makes it. A model can estimate fit, but it cannot carry the consequences of a wrong estimate. If it auto-rejects a lead that looked small and was actually a company’s first step toward a bigger project, no model apologizes for that. I do, or I make sure it never needs saying.

There is a second reason the human stays in the loop. Prospects can smell an automated rejection. Their inquiry took effort. A reply that is obviously templated tells them their effort was worth a script. The sorting can be instant. The judgment has to answer back.

The Form Is the Funnel: Five Fields, a Checkout, and a Webhook

Receipt time, because everyone claims conversion improvements and nobody shows the diff. I run a project called aiexhibitions.net, a directory of immersive exhibitions and planetarium shows with more than 900 pages generated statically in Astro. Venues can subscribe to be listed. That subscription funnel is where I rebuilt the conversion path.

The rebuild went pay-first. The new flow asks for almost nothing up front: venue name, email, chosen plan, billing interval, and a consent checkbox. Five fields. Then straight into Stripe Checkout. When payment clears, a webhook seeds a draft listing for the new subscriber automatically. The customer sees the beginning of their value before I have even read the welcome email.

The numbers, because that is what a receipt is. The rebuild touched 9 files, added 164 lines, deleted 506, and shipped with 97 unit tests. I built it by steering AI the way I describe in my guide to shipping tools without a dev team, and that is the other half of the ai lead generation story nobody tells: the model can help you build the infrastructure, not only fill it.

The lesson generalizes beyond subscription sites. Conversion is not about asking more questions to qualify harder. It is about asking fewer, then letting automation do the qualification after the commitment, when the prospect is warm instead of wary.

First-Touch Attribution: 44 Vectors and One Version of the Truth

Ask a marketing team where their leads come from and you often get a shrug pointing at a dashboard bucket called Other. My answer is a small kit I built myself, and it is the most useful piece of plumbing in my funnel.

It captures 44 attribution vectors at first touch and classifies every lead into clean buckets. It runs in Node and in PHP, because my sites are not all on the same stack. And it classifies identically on the client and on the server, which sounds like a detail until you have debugged why the two disagree.

The taxonomy is where the boring becomes valuable. A lead from youtube, whatsapp, telegram, or tiktok is Social. A lead from duckduckgo, ecosia, or brave is Organic, not Direct. Without that mapping, the privacy-friendly search engines and the messaging apps all collapse into a bucket called none, and your channel decisions get made on fiction.

Attribution is also the precondition for everything else on this page. Qualification without source is half a judgment. Follow-up without source is generic. Channel budget without source is a donation. If you fix one thing in your funnel this quarter, fix this, and fix it before you buy a single new tool.

Follow-Up: Speed to Lead With a Human Signature

The lead that waits cools. Every hour between an inquiry and a reply costs intent. My policy is a draft-first reply: every inbound lead gets a drafted answer within minutes, and I review and send it myself.

The draft is where AI earns its keep, because it is not generic. Attribution tells it where the lead came from, so the draft can open with context: the article the person read, the channel they arrived from, the plan they picked. A reply that starts from the reader’s context reads like attention, and attention is most of what a first reply sells.

The line, again: draft, not send. Automated replies are recognizable within a sentence, and they convert trust into annoyance at machine speed. The broadcast side of staying in touch runs on the pipeline I documented in my newsletter automation piece. The one-to-one side stays manual on purpose, because the one-to-one moment is where trust is actually made.

The 600 CHF Channel Test I Refused

In August 2026 I sat down to evaluate ChatGPT Ads as a lead channel for my consulting. The test budget on the table was 600 CHF. Reasonable, I thought, for a first signal.

The evaluation killed the idea before the budget moved. The targeting is contextual and geographic. There is no targeting by role, no targeting by company, nothing that lets me reach the people who actually buy what I sell. For a B2B offer, 600 CHF under those rules buys noise. A test big enough to punch through that limitation starts at 5-10k USD per month, which is a different decision entirely.

So I said no, and I wrote the reason down, because the lesson is bigger than the channel. A new channel is not a free lead machine. The hype says early adopters get cheap attention. Maybe, but only if the targeting can express who your customer is. Run that check against your own customer acquisition cost math before the budget leaves the account, not after.

While I was refusing that channel, my actual lead source kept working. This blog is where my consulting leads come from: articles that answer real questions, structured so that people and AI engines can cite them. That compounding is the whole point. A piece that gets cited keeps producing inquiries for years at zero marginal spend, while ads produce exactly as long as you pay and not one day longer. The article you are reading is part of that machine.

What I Never Delegate to AI in Lead Generation

There is a list I keep about AI in my funnel, and it is a list of refusals. The machine never makes the promise. It never sets or defends a price. It never apologizes for a failure, because an apology generated from a template is not an apology. And it never says no to a prospect, because rejecting someone is a judgment with consequences.

The practical guardrails follow from those. Nothing sends without my review. No client data goes into a public model, a rule I hold for my employer’s data with the same force as for my own. And every AI-drafted reply ships under my name, which means I read all of it, every time.

If you are starting from zero, the order matters. Fix attribution first, so you can see. Shorten the form second, so more people finish. Adopt the draft-first reply third, so speed stops depending on your mood. Buy tools last, and only where a station is still slow after the first three fixes. That sequence costs almost nothing, and it is the whole method.

Frequently Asked Questions

Can AI generate leads on its own?

No, and the framing is the problem. AI does not generate trust, and a lead is trust with a form attached. What the machine does is remove the waiting and the guesswork at every step after someone notices you: it sorts, classifies, drafts, and feeds the analysis. The noticing still comes from your content, your offer, or your channel spend. Teams that expect the model to conjure demand are the ones whose projects end in the failure pile.

What should I automate first in my lead funnel?

Attribution, without hesitation. It is invisible, unglamorous, and it makes every later decision honest. The second automation is the follow-up draft, with a strict review before anything sends. The last thing to automate is judgment, which is to say never.

Does ai lead generation work on a small budget?

It works best on a small budget, because the structural wins cost almost nothing. Shortening a form costs an afternoon. A follow-up draft costs the subscription you probably already pay. The honest answer about ai lead generation on a tight budget is that the wins come from structure, not spend: see where leads come from, ask less of them, reply faster, and refuse the channels whose targeting cannot even describe your customer. Save the 5-10k monthly experiments for when the funnel underneath them deserves the traffic.