Find leads with AI

Most people do not have a lead problem. They have a 'writing forty specific emails takes a whole day' problem. The list has been available this whole time. What was missing was the hour per person it takes to know enough about them to say something worth reading.

I work with international property buyers at a Dubai brokerage, and I build automations for other people's businesses. What I have seen in both: the outreach that gets answered is the outreach where the sender clearly knew something specific before they wrote. That knowledge is research. Research is what a model is genuinely good at compressing, as long as you keep it on a leash.

What does AI lead generation actually change?

The easy answer sounds great. AI finds your ideal customers at scale, personalises every message, and books meetings while you sleep. Fine. Which person do I write to first, and what do I put in the first line?

Here is the honest version. A model does not know who your buyers are. It has no access to intent you have not given it. What it will do, reliably, is read twelve public pages about a company and hand you back the four facts that matter, in a format you can act on, in about two minutes.

• What it changes: the cost of going from a name to a useful picture of a person and their situation.
• What it changes: the cost of keeping that picture organised across two hundred people instead of ten.
• What it changes: the blank-page problem when you sit down to write the first sentence.
• What it does not change: whether the person needs what you sell, whether they are allowed to hear from you, and whether your offer is any good.

So there is real time to be saved, and it sits earlier in the process than the pitch suggests. It is preparation, not discovery.

Why buying a list and blasting a template does not work

This is the thing that made everyone's inbox worse, so it is worth saying plainly. You can buy ten thousand addresses, generate ten thousand variations of one email, and send them all before lunch. The result is a spam complaint rate, a burned sending domain, and a reputation with the exact people you wanted to reach.

It also runs straight into the rules. The UAE regulator's policy on unsolicited electronic communications defines spam as marketing sent to a recipient without that recipient's consent, and it separately tells licensed operators not to supply or use tools that facilitate harvesting electronic addresses. In the UK, the ICO's position on electronic mail marketing is that consent has to be specific to you, and that the soft opt-in for existing customers explicitly does not cover bought-in lists. Two different regulators, same shape of answer.

The generated template has a second problem that has nothing to do with law. It is obvious. Everyone has now read the line 'I loved your recent post about [topic]'. The compliment is generated, the post is real, and the two have never met.

Who are you actually looking for?

Before any tool, write the definition of a good-fit lead. Not a persona. A definition another person could hold against a real name and answer yes or no.

Three parts make it checkable: the situation (what is true about them), the role (who specifically, by job, not by company), and the trigger (what happened recently that makes this a live conversation rather than a cold one). The trigger is the part most people skip, and it is the part that does all the work.

The sharpened rows are narrower on purpose. A definition that returns forty names you can actually research beats one that returns four thousand you will not.

How do you research a lead with AI without inventing facts?

Give the model the sources. Do not ask it who someone is.

The failure mode is not that the model refuses. It is that it answers beautifully. Ask a model to tell you about a mid-sized firm it has never seen a page about, and you will get a confident paragraph with a plausible founding year, a plausible headcount, and a plausible expansion into a market they have never entered. It reads exactly like the true version.

Anthropic's own guidance on reducing hallucinations says the useful things out loud: give the model explicit permission to say it does not know, make it pull word-for-word quotes out of the source before it does anything with them, and have it attach a supporting quote to every claim so you can audit the output. That is not a prompt trick. It is the whole method.

1. 1Collect the sources yourself: the company site, their careers page, a public post, a news item, a public register or filing if there is one.
2. 2Paste the source text and the URL into the prompt. Both, every time.
3. 3Ask for quotes first, conclusions second.
4. 4Require a 'could not find' list. An empty one is a warning sign, not a win.

The prospect table: the boring output that makes outreach possible

This is the artifact. Six columns, in a spreadsheet, one row per person. It is not impressive. It is inspectable, which is better.

Notice what the opening sentence does not do. It does not compliment them, it does not mention their impressive growth, and it does not say I have been following their journey. It repeats back a real thing, in my words, and offers a guess they can correct. A guess someone can correct is an invitation to reply.

Why I would skip the AI lead score

Every tool in this category wants to hand you a number. This lead is an 87. That one is a 34.

I would not build my process on it, for one plain reason: when the 87 is wrong, you cannot tell what went wrong. A score is a conclusion with the reasoning removed. The Evidence column is the same judgement with the reasoning still attached, which means you can read ten rows on a Friday, notice that your trigger is picking up companies at the wrong stage, and fix the definition instead of the model.

If a score helps you sort a long sheet, use it to sort. Do not use it to decide, and do not let it replace the column that tells you why.

The research prompt I would run

This fills one row. Run it per prospect, with the sources you collected pasted in. The placeholders in square brackets are the only parts you change.

You are helping me research one prospect before I write to them.
Use ONLY the sources pasted below. Do not use anything you already know
about this person or company. If the sources do not say it, it does not exist.

PROSPECT: [name], [role], [company]
WHAT I SELL: [one sentence]
MY GOOD-FIT DEFINITION: [situation + role + trigger]

SOURCES (url, then the text I copied from it):
[paste source 1]
[paste source 2]
[paste source 3]

Step 1. For each source, pull out up to three word-for-word quotes that say
something about what this person or company is doing right now. Keep the URL
next to each quote. If a source has nothing useful, write "nothing useful".

Step 2. Fill exactly one row of this table:
Who | Trigger (with date) | Evidence (quote + URL) | My read | Opening sentence | Status

Step 3. Rules for the row:
- Evidence holds only quoted or copied material with its URL. No paraphrase.
- My read holds your interpretation, in one sentence, starting with "Looks like".
- Never move anything from My read into Evidence.
- Trigger must have a date from the sources. No date, write "no dated trigger".

Step 4. Write the Opening sentence so I could send it with no edits. It must
refer to the Evidence, not to their job title. No compliments. Under 30 words.
End it on a guess they can correct.

Step 5. List what you could not find, including anything you are unsure about,
such as whether [name] still holds that role.

Set Status to "ready" only if Evidence has at least one quote with a URL.
Otherwise set it to "thin" and tell me which source I should go get.

Run it on five people you already know well. If the rows tell you things you already knew, and the 'could not find' list is honest, it is working. If it produces a confident row for a person you know it has no information about, your sources are too thin and the model is filling the gap for you.

How do you turn research into outreach you would actually send?

Personalisation is not a variable. A merged first line on top of an unchanged pitch is still a template, and it reads like one.

The test I use: could this message be sent to the person in the next row without changing anything except the name? If yes, it is not personalised, it is just addressed. The Evidence column exists so the answer is no.

• Open on the trigger, in your own words, not theirs.
• Say the guess you made from it, and leave room for them to correct it.
• Make exactly one ask, and make it small enough to answer in one line.
• Cut every sentence that would survive being copied into someone else's email.

Outbound is also not the only road. Research this specific tends to surface the same three questions across your whole list, and those questions are content. If you would rather the next forty of these people arrive on their own, turning one idea into a week of content is the other half of the same work.

What are the rules on consent and cold outreach?

I am not a lawyer and this is not legal advice. What I can tell you is that the rules differ by market, that they apply based on where you and your recipient are, and that reading the actual regulator takes about twenty minutes. Do that once, properly, for the markets you sell into.

Here is the shape of it in three places I looked at directly, so you know what kind of thing you are looking for.

The practical read, and this is my read rather than anybody's rule: if your process depends on the recipient never noticing they did not ask for this, the process is the problem. Send fewer, to people you can name a reason for, with your real identity and a real way to tell you to stop.

Where is the line on scraping?

This is where a lot of lead generation advice goes quiet, so let me be direct.

LinkedIn's user agreement prohibits using software, scripts, bots, crawlers, browser plugins or extensions to scrape or copy the service, including profiles and other data. It also prohibits automated methods to add or download contacts or send messages. Their own help page on prohibited software says accounts using those tools risk being restricted or shut down. So a scraper, a connection-request bot, or the extension someone sold you as a growth hack is against the terms you agreed to, whatever the seller told you.

Reading a public page yourself, in a browser, like a person, and copying what it says into your own notes with the link, is a different activity. That is what the prompt above is built for. I keep the volume at human scale on purpose, because at human scale the question never comes up.

Where a platform publishes an official API or an export, use that, and read what it permits before you build on it. The line is not 'public data is free'. The line is the agreement you accepted to get access.

When should you hand this to an agent?

Once the table works by hand, and only then.

By 'works' I mean something specific: you have run twenty rows yourself, you know which sources are worth opening for your market, you know what a thin row looks like, and you have a definition of 'ready' you would defend. All of that is the spec. Handing it over before you have it means you have automated a process you cannot judge, at a speed that makes the mistakes bigger.

That handover is its own build, and it has its own failure modes. Building an agent that does real work covers scoping the job and keeping it on rails. If you have not yet connected a model to the tools you already use, setting up your first AI workflow is the step before that one.

My honest position on the agent version: it is worth it when you are researching enough people per week that the manual pass is the bottleneck. Below that, the agent is a project you are doing instead of outreach. I have caught myself on the wrong side of that line more than once.

What to do this week

Write the good-fit definition, with a trigger and a date range. Find ten names that match it. Fill the six columns by hand, no model, and time yourself. Then run the prompt on ten more and compare the rows.

The rows will tell you something either way. If the model's rows are thinner, your sources need work. If they are the same and faster, you have your process. If both sets are thin, the problem was never the research, and the definition needs another pass.

Then the question worth taking to your own desk: when you filled those first ten rows, which column were you tempted to leave empty? That column is where your outreach is actually failing.