ChatGPT for Lead Generation: The Jobs It Does and the One It Must Not

Ask ChatGPT for the work email address of a named person at a named company and it will give you one. It will be correctly formatted, it will follow a plausible pattern for that domain, and it will be delivered with exactly the same confidence as a fact the model actually knows. Some of the time it will be right, which is the part that causes the damage, because a method that works often enough to seem reliable is harder to abandon than one that fails visibly.

That single behaviour draws the line through everything else on this page. There are jobs in lead generation that ChatGPT does genuinely well, and there is one category it must never be asked to do, and the difference is whether the answer can be checked.

A disclosure before going further. RevenueFlow runs cold outbound as a service and uses language models daily in that work, on the boundary described below. Everything here is either our own documented practice or comes from OpenAI's own published material as it rendered on 15 August 2026.

What the vendor itself says it is for

OpenAI publishes its own guidance for using ChatGPT in sales, first posted in July 2025 and updated in May 2026. The use cases it names are drafting outreach messages, preparing for meetings, handling objections, summarising calls and updating CRM notes. Its published prompts are shaped accordingly: write a short cold email to a job title at a company, using background you supply.

That list is worth reading closely for what is not on it. Every use case involves the model working on material the user brings to it. None of them asks the model to supply a fact from its own memory, and none of them asks it to find a person.

The vendor's own framing is therefore the same as the honest one: this is a tool for transforming information you already have, not for producing information you do not.

The four jobs it does well

Four jobs come back reliably, and they share one property that is easier to state than the jobs themselves: in every case the answer can be checked against something already in front of you.

Jobs, sorted by whether the answer is checkable

• ✓Yes: Turning a vague market description into a testable ICP definition
• ✓Yes: Summarising a page, a filing or a transcript you have supplied
• ✓Yes: First-draft copy from evidence you provide, for a human to edit
• ✓Yes: Classifying a reply into interested, referral, not now or wrong person
• ✕No: Supplying an email address, a phone number or a headcount from memory
• ✕No: Confirming whether a company uses a named piece of software
• ✕No: Producing a list of companies matching criteria, without a data source

Turning a description into a definition. The most common reason outbound fails is a target definition that was never written down properly, and a conversation with ChatGPT is a good way to force one. Describe your best customers, ask it to name the attributes they share, and argue with the result. The output is a hypothesis you then check against a data source, which is exactly the right relationship. What that definition needs to contain, including the market-size arithmetic that makes it real, is in ideal customer profile.

Summarising material you supply. Paste a company's own page, a job posting, an annual filing or a call transcript and ask what matters in it. This is retrieval rather than recall, the source is in front of both of you, and every claim in the output can be checked against the text you provided in seconds.

First drafts from evidence. Give the model a real observation about a prospect and ask for three ways to open a message with it. The quality tracks the evidence you supplied far more than the model you chose, which is the finding that matters most in practice and the one that gets ignored most often.

Sorting replies. Interested, referral, not now, wrong person, unsubscribe. This is classification against a fixed taxonomy on text the model can see, it is fast, and a wrong answer is caught by the human who reads the reply anyway.

Using ChatGPT for sales prospecting specifically

Sales prospecting is where the temptation to use ChatGPT badly is strongest, because the request feels so close to something the tool obviously can do. Asking for a summary of a company works. Asking for a list of companies that match your criteria produces a list that looks identical and is not the same object at all.

What comes back is a plausible set of names assembled from training data of uncertain age, with no coverage guarantee, no recency, and no way to distinguish a company that fits from a company the model has simply heard of. There is no source to check because there was no source. Two runs of the same prompt will produce different lists, and neither run can tell you what it missed.

The workflow that does work puts the model on either side of a real data source rather than in place of one.

1. Step 1Define, with the model

   Argue your way to a written target definition: the attributes, the size band, the disqualifiers. The output is a hypothesis, not a list.

2. Step 2Source, without the model

   Query an actual data provider against that definition. This step returns records with provenance, which is the thing a model cannot produce.

3. Step 3Verify, without the model

   Run the addresses through verification before anything sends. A found address and a deliverable address are different facts.

4. Step 4Research, with the model

   Give it each prospect's own pages and ask what is genuinely notable. Grounded in a source you fetched, so every claim is checkable.

5. Step 5Draft, with the model

   One message per prospect from that evidence, then a human edits. The model writes from what it was given rather than from what it recalls.

The second and third steps are the ones people try to skip, and they are the two ChatGPT cannot do. The mechanics of getting real contact data, and why the sequence of providers matters, are in waterfall enrichment. What separates a durable targeting attribute from a dated event worth acting on is in b2b prospecting.

The economics that changed, and the one that did not

Writing a personalised message used to cost a person several minutes. With ChatGPT it costs a fraction of a cent. That is a real change and it is the reason this category exists.

The recipient's willingness to read did not change at all. Their inbox now receives the output of everybody else's cost reduction as well as yours, so the same effort buys less attention than it did, and volume stopped being an advantage the moment it became available to everyone. The full argument is in AI lead generation.

What follows from that is not a reason to avoid the tool. It is a reason to spend the saving on the part that did not get cheaper. Knowing which specific fact about a prospect makes your message worth reading is still expensive, still human, and now the only differentiator left, because everybody has the same models.

Where we run this differently

Two of our own standing positions bear directly on how a model gets used here, and both cost us something.

We ground every generated message in the prospect's own material rather than letting a model write from what it recalls. In practice that means fetching real pages before generating anything, and running a check for invented claims before a message is eligible to send. It is slower and it costs more per message. It is also the only version of this that survives contact with a reader who knows their own business.

And we run one message per campaign for cold outbound, with no thread replies and no bumps. That is a harder discipline when generation is cheap, because the obvious use of a cheap writer is to write more messages to the same people. Every message after the first reaches only the people who saw the previous one and declined to answer, which is the population most likely to file a spam complaint, and the reputation cost lands on the sending domain across everything else it sends. What replaces the follow-up is a new campaign with a genuinely different premise. The full argument, including what the position costs us, is in email sequence software.

Prompts are the least interesting part

Most of what is written about ChatGPT for lead generation is a list of prompts, and prompts are the component that matters least. A well-phrased request to a model with nothing useful in front of it produces a well-phrased guess.

What actually determines the output is what you put in the context: the prospect's own words, a real signal with a date on it, your actual offer, and a target definition specific enough to exclude somebody. Teams that get value here are not the ones with better prompts. They are the ones who did the unglamorous work of assembling evidence first, and who kept a person accountable for the result rather than assuming the software owns it.

The short version

ChatGPT is good at four jobs in lead generation: turning a description into a testable target definition, summarising material you supply, drafting from evidence you provide, and classifying replies. All four share the property that the answer can be checked against something in front of you.

ChatGPT must not be asked to supply facts from memory. Contact details, headcounts, technology usage and lists of matching companies come back correctly formatted and confidently wrong often enough to be dangerous, and there is no source to check because there was no source.

The workflow that works for both lead generation and sales prospecting puts ChatGPT on both sides of a real data step and never inside it: define with it, source and verify without it, then research and draft with it against pages you actually fetched.

The cost of writing collapsed. The cost of knowing something worth writing did not, and that is now the whole difference. If you would rather have that part run for you, you can see what a campaign would look like for your market.

Vendor guidance verified against OpenAI's own published sales resource as of August 2026. Verify current product behaviour with the vendor before relying on it.

Sources: ChatGPT for sales, OpenAI Academy