AI Lead Generation: What the AI Does and Where You Step In
What the AI does at each step of lead generation, where a person is still needed, and what our test of 3,570 emails showed.
Robbert Veer, Founder
8 min read
AI lead generation is the use of AI models to find, research, score and write to potential customers. The AI earns its place in the middle of that chain, where it reads thousands of company websites and ranks them. A person still picks the market, decides how much to review before it goes out and holds every conversation.
This guide goes step by step for an agency or freelancer that sells websites, SEO or marketing to local businesses. The evidence is our own test of 3,570 cold emails, where the choice of who to write to mattered more than the wording.
What is AI lead generation?
AI lead generation is lead generation where software does the reading, ranking and drafting that a person used to do by hand. Most uses of AI for lead generation come down to three jobs.
- Reading. A model opens a company's website, listing or profile and records what it finds.
- Ranking. A scoring model turns those findings into a number, so the best prospects sit at the top of the list.
- Drafting. A language model writes a first line, or a whole message, from what the reading found.
The guides that rank for this search are written for software companies with a CRM, a sales team and a stream of inbound leads. They cover chatbots, intent data and lead routing.
An agency that sells to dentists and roofers has a different problem. Nobody is filling in a form. Say your list holds 5,000 local businesses. The question is which 200 deserve a letter, and what that letter should say.
What does the AI do at each step?
The AI does most of the work in steps 3 to 5 and little of it anywhere else. The table shows the eight steps of an outbound campaign to local businesses.
| # | Step | What the AI does | What a person does |
|---|---|---|---|
| 1 | Choose the market | Nothing you should rely on | Picks the trade, the area and the offer |
| 2 | Build the list | Little. This step is a database search | Sets the filters and removes current clients |
| 3 | Research each business | Reads every website and records the problems | Spot-checks a sample of the findings |
| 4 | Score and shortlist | Scores each site and sorts the list | Sets the cut-off |
| 5 | Draft the first line | Turns one finding into one sentence | Chooses how much to review, and rewrites what is wrong |
| 6 | Send | Little. Plain automation keeps the schedule | Sends under their own name |
| 7 | Answer replies | Can suggest a draft | Writes the answer |
| 8 | Hold the call | Nothing | Runs the meeting and closes |
Steps 1, 7 and 8 belong to a person. Steps 3 to 5 are where the hours go, and those are the hours AI gives back.
Where does AI help the most?
AI helps the most at research and scoring, because no person can do that step at volume. Two sets of numbers show the size of it.
In an earlier campaign, the agency It'sJustContent emailed 1,382 Dutch companies. Reviewing every one of those websites by hand would have taken about 184 hours, at about 8 minutes a site. The campaign brought 5 positive replies, 3 sales calls and no clients.
We then scored 20,000 Dutch company websites with a model that uses 206 indicators. At 8 minutes a site, the same review by hand would take about 2,667 hours, which is 20,000 times 8 minutes. The 20% of websites with the most problems became the prospect list, and 3,570 emails went out to three groups over 5 working days.
| Group | Who received the email | Positive replies |
|---|---|---|
| A | A random sample from the other 80% of websites | 0.50% |
| B | The weakest 20%, with a generic email | 1.60% |
| C | The weakest 20%, with an email that names the problem | 2.35% |
Businesses with the weakest websites were about 3.2 times more likely to reply positively. Group B points to selection as the cause, since its email was generic. The whole experiment cost €242.25, and the cost per positive reply in the selected groups was €5.15. The method is in our write-up of the test, published in October 2026.
Three limits apply. The companies were Dutch, the test covered email only, and it left out businesses with no website.
Scoring is also the part you can copy without any AI on a short list. Our guide to lead scoring for agencies shows how to use the website as the score.
Does an AI-written first line get more replies?
Possibly, and our test could not prove it. Group C received an email that named the problem on the company's own website. It drew 2.35% positive replies against 1.60% for the generic email. That gap was not statistically significant, with p = 0.238, so read it as a hint.
A larger and older study points the same way. Backlinko and Pitchbox analyzed 12 million outreach emails in 2019 and found that personalized message bodies got 32.7% more replies (Backlinko). That study covered outreach of every kind, link requests included. The figure will differ for an agency that writes to local businesses.
Two things follow for your own campaigns.
Selection is the larger lever. It was worth about 3.2 times in our test, against a wording gap too small to call. A sharp first line sent to the wrong business is still a wasted message.
The writing is the cheap part. Generating 1,190 problem descriptions cost about €5 in that test, which is under half a euro cent each. The words cost almost nothing, so the list deserves your attention. For the line itself, see our guide on how to write a cold email.
What should the AI not do?
The AI should not make a decision that carries your name. Four jobs stay with a person.
- Choosing the market and the offer. A model can rank 900 dentists by website quality. It cannot know which trades you serve well, or what a new patient is worth to a practice in your town.
- Stating a fact nobody checked. A model can misread a page. If your first line tells an owner the site has no booking button and it has one, the message is finished. Read every line against the website before it goes.
- Being the sender. The agency is the sender, whoever wrote the text. The FTC's guide to US email law sets eight requirements, among them a valid postal address, a working opt-out and a duty to monitor what others send for you (FTC). This is not legal advice.
- Answering a reply. An owner who writes back has a question for you. Answer it yourself, and answer it the same day.
How do you start with AI prospecting as a small agency?
Start with one trade in one area, and let the AI do only the reading and the ranking at first. Keep the first list short enough that you can check the output yourself. AI prospecting goes wrong when nobody looks at what the model found.
- Pick one trade and one area, such as dentists in one county.
- Build the full list for that trade and area. Our overview of lead generation software explains what each kind of tool returns.
- Score every website on the list. Do 50 by hand, and use software for 5,000.
- Keep the weakest fifth as your shortlist. In our test, the prospects were the 20% of websites with the most problems.
- Draft one first line per business from its biggest finding. Read each line before it leaves.
- Send under your own name, and answer every reply yourself.
Leadodo is in this category. Twip, the AI inside Leadodo, scores every website in a search out of 100 across 10 domains. Sites that score below 40 form the shortlist, and Twip's findings become the opening line of each letter and email. Your Outbound Machine sends under your agency's name. You choose how much to review. You can review every letter and message before it goes, review a sample such as the first 100, or let your Outbound Machine send without review. See how Twip scores a website.
Scoring is one part of a larger routine. The full routine for an agency is in our guide to agency lead generation.
Questions
Can AI generate leads on its own?
No. AI can build a list, rank it and draft a message. A lead exists once an owner replies, and what follows is a conversation between two people. The AI shortens the road to that reply. It does not hold the conversation for you.
Is AI lead generation worth it for a small agency?
It is worth it once your list is longer than you can review by hand. At about 8 minutes a site, 50 websites take under 7 hours and 1,000 take about 133 hours. Below 50, a spreadsheet and one working day will do.
Does AI personalization increase reply rates?
The evidence is thin. In our test of 3,570 emails, an email that named the website's problem got 2.35% positive replies against 1.60% for a generic one. That gap was not statistically significant, with p = 0.238. Choosing the weakest websites had the clear effect, at about 3.2 times.
Is it legal to send AI-written cold emails?
In the United States, commercial email falls under the CAN-SPAM Act, and the FTC's guide says the law has no exception for B2B email. The agency that sends is responsible for each message, however the text was written. Other countries set their own rules. This is not legal advice.
What is the difference between AI lead generation and an AI SDR?
AI lead generation is the whole practice of using AI to find and reach prospects. An AI SDR is one kind of product inside it, built to run the outreach steps of a sales development representative. Our guide to the AI SDR covers what it does well and where a person is needed.
Next step. See how Twip scores a website across 10 domains and builds your shortlist. See AI analysis.
