Using Claude or ChatGPT for lead generation: what works, and where it stops.
You can do a good part of what we do yourself in Claude or ChatGPT, and for one company at a time you should. The problem is volume. A chat that's sharp for 20 companies has drifted by the 30th, and nobody can tell which answers went bad.
This isn't about Claude or ChatGPT being bad at this. They're very good at it. They just weren't built to do it 500 times a month without a person watching.
Six things an assistant does well, one company at a time.
If you do nothing else from this page, do the first one.
An afternoon, onceConnect it to real data
Connect a contact-data provider, your CRM and your inbox, so it works from real records instead of guessing.
Sign up to a contact-data provider such as Apollo or Lusha and connect it to Claude or ChatGPT through the provider's connector. Then connect your CRM and your inbox the same way. From then on it works from real contact details and your real records instead of guessing.
5 to 10 minutes a companyResearch one company before a proposal
Save a research brief with your criteria and the facts you want, and run it on one company at a time.
Save a research brief – your criteria for a good customer, the facts you want (what they make, who they sell to, sites, size), the sources to prefer, and an instruction to mark anything unconfirmed as unconfirmed rather than filling it in. Run it on one company at a time.
A few minutes a companyFind the right people at one company
Give it your personas and ask who holds each role at that company. Check before you write to them.
Give it your personas (the roles that matter to your sale) and ask for the people who hold them at that company, with the provider's verified email where it has one. Check the role against the person's own profile before you write to them.
10 minutes a companyDraft the first email and two follow-ups
From the research, in your own voice. Put the follow-up dates in your calendar yourself.
From the research, in your own voice, with one product or angle picked from what the company actually does. Send from your own mailbox. Put the follow-up dates in your calendar yourself – nothing will send them for you.
Minutes, and a happy customerFill in their paperwork for them
New account forms and NDAs filled in from the customer's details, so they only have to check and sign.
New account forms and NDAs, filled in from the customer's details so they only have to check and sign. Customers like this one more than anything else on the list.
A minute a replyRead a reply and tell you what it is
Paste a reply in and ask what it is and what to do. It's good at this.
Paste a reply in and ask: is this a yes, a not now, a referral, an out of office or a request to stop, and what should I do? It's good at this. It just can't read the mailbox for you.
This is where the assistant earns its keep.
What we wouldn't do is hand it a list of 500 companies and ask it to work through them.

Sharp for 20 companies. Drifted by the 30th.
Three problems turn up as soon as the job is a list rather than a company. We hit all three ourselves before we built anything, and most people who've tried it will recognise them.

Context fatigue
A chat is sharp for the first 20 or so companies. Then the instructions you gave at the start are the oldest thing in the conversation, and the model stops following them. It skips rows, redoes rows, and lets details from one company land on another.
You can't tell which answers went bad
The 30th answer looks exactly as confident as the third. Finding the invented revenue figure or the wrong-company contact means checking every row, which takes nearly as long as the research did.
It isn't automatic
Nothing runs while you sleep. No follow-up goes out on day four. No reply gets read unless you paste it in. Nothing gets written to the CRM unless you build that. Every batch needs a person to start it, watch it and finish it.
Why a fresh chat doesn't fix it
The first two problems are really the same problem seen twice – the more the assistant has to hold in one conversation, the less reliable each answer gets, and the harder the failures are to spot. Starting a fresh chat every 20 companies helps with the first and does nothing for the third.
The same 170 companies, four ways.
Take a modest volume – 500 new contacts a month, which is roughly 170 companies at three people each. The minutes are from our own experience and yours will differ a bit. The shape won't.
The sums in a table, and what they leave out
| Way of working | Per company | 170 companies a month | Who's doing it |
|---|---|---|---|
| By hand | 25 to 30 minutes of searching, reading and typing it up. | Around 75 hours | Two full weeks of one person, every month. |
| An assistant, one company at a time | 6 to 10 minutes including reading and checking the answer. | Around 22 hours | Half a week of one person, every month, prompting and checking. |
| An assistant, in batches of 20 | 3 to 4 minutes of running, plus 2 to 3 minutes checking every row, because you can't tell which went bad. | Around 17 hours, and nine fresh chats | Most of a week, and the checking is the job. |
| A system built around the assistant | About 2 minutes of machine time, with the checks built in and a person only on the exceptions. | Around 6 machine hours, spread across the month | An owner's hour or two a week on the exceptions and the sample. |
Then add the follow-ups (two per company, on days four and eleven), reading every reply, and writing what you learned into the CRM. None of that is in the minutes above, and all of it is where a sequence falls over when a person is doing it alongside a real job.
The assistant takes the hand time down by two-thirds or more. It doesn't take the person out. For a handful of companies a week that's a very good deal. At 500 contacts a month it's already most of a week of someone's time, every month, before the follow-ups and the replies – and nobody was hired to do it.
The same assistant, asked much smaller questions.
Not a cleverer model. Code around it does the parts an assistant can't.

Hundreds of small jobs, each with one question
One big job becomes hundreds of small ones. Each call gets one clear question and no memory of the other 499.
One big job – refresh 500 records – becomes hundreds of small ones: is this company still trading, does it match the profile, who holds this role, find this one person's email, is this address deliverable. Each call gets one small, clear question and no memory of the other 499. The code looks after the order, the checking and writing the result back. In our experience the assistant is very good at small, well-defined jobs and fairly unreliable at big, vague ones, so the system simply never gives it a big, vague one.

A gate before anything reaches the CRM
Every company is scored against the profile and every person checked against a persona before a record gets written.
Every company gets scored against the profile and every person checked against a persona before a record gets written. Anything that doesn't pass doesn't go in, and anything incomplete goes to a person first. That's what stops a CRM filling up with people who look about right and are the wrong companies or the wrong roles.

A schedule
Follow-ups go out four days after the first email and a week after that, from the rep's own mailbox, in the same thread.
The first follow-up goes out four days after the first email, the second a week after that, from the rep's own mailbox, in the same thread, with the same colleagues copied in. In our own outreach roughly six replies in ten arrive after the first email, so stopping after one send misses most of them.

Every reply read
Each reply gets sorted into what it is, and the facts in it get acted on.
Each reply gets sorted into what it is, and the facts in it get acted on – a referral becomes a new, checked contact, a leaver's record gets updated so they stop getting email, an out-of-office gets recognised as automatic and not counted as a response, a yes becomes a deal and a task. In our outreach about one reply in three comes from someone other than the person we wrote to, and that's exactly the work nobody gets round to by hand.
Yes. The code is the easy part now. The design is the hard part.
With AI coding tools, anyone with an account and a clear description can have a working agent, a schedule and a CRM connection by the end of the week.
- 3 years
- of decisions that were wrong the first time and right by the fifth
- 2.7 million
- company and contact records processed through it, across many customer engagements
- 607
- unique guard rails built in, around small AI steps that each ask one precise question
The long answer
The code is no longer the hard part. With AI coding tools such as Claude Code and Codex, anyone with an account and a clear description can have a working agent, a schedule and a CRM connection by the end of the week. A year ago the build was the barrier. It isn't now, and we wouldn't pretend otherwise.
The hard part is the design. What makes a system like this work at volume isn't the code, it's the thinking the code expresses – which jobs are cut small enough to be reliable, which checks run before anything gets spent or written, where a person has to sit, how a reply gets read, what happens when a company has four sites or a contact has left, and how all of it is tuned so the output is as high in volume and quality as the sources allow. Each of those is a decision that was wrong the first time and right by the fifth, and there are a lot of them.
Ours has 3 years of exactly that in it. More than 2.7 million company and contact records processed through it, across many customer engagements, and 607 guard rails built in around small, carefully placed AI steps, each asking one precise question at the moment it's needed – is this company still trading, is this the same site or a second one, is this address deliverable, is this reply a referral or a no. None of them is clever on its own. Together they're why it runs every day without a person watching, and why the records it writes can be trusted.
If you build your own, the parts are the same whoever writes the code:
- connections to the data provider, the CRM and the mailboxes, with permissions granted in one place
- somewhere to keep the state of every company and every email that's due
- the profile and personas written as rules a program can apply, with saved test cases to prove it does
- email verification, and an unsubscribe process that works when colleagues are on the same thread
- a scheduler for the follow-ups and a reader for the replies
- a review queue for the exceptions, and a person who clears it
- cost control, because every small job is metered
The coding tools will give you all of that quickly. They won't tell you which 607 things need a guard rail, because nobody knows that until they've run a few million records through it.
The build is the smaller bill. These six lines arrive monthly.
All of them sit inside our monthly price. If you build your own they land on your own invoices.
Contact data
Your subscription and your credits, rising with volume.
Web search and page reading
Metered per query and per page.
Email verification
Per address, every time.
AI usage
Metered per call, and it climbs with the volume you process.
Hosting and the plumbing
Monthly, and someone has to keep it patched.
A person to run and maintain it
The biggest line, and the one that stops the moment that person is busy or leaves.
What drives each line
| Running cost | What drives it | If you build your own |
|---|---|---|
| Contact data | A subscription to a provider such as Apollo or Lusha, plus credits for every person looked up and every email or phone number revealed. | Your subscription and your credits, rising with volume. |
| Web search and page reading | Every company researched means dozens of searches and page fetches, through paid search and crawling services. | Metered per query and per page. |
| Email verification | Every address checked for deliverability before it's used, so your domain's reputation isn't spent on bounces. | Per address, every time. |
| AI usage | Each of the hundreds of small jobs is a paid model call – the fit check, the research summary, the reply reading, the draft. | Metered per call, and it climbs with the volume you process. |
| Hosting and the plumbing | Somewhere for the automation to run on a schedule, keep its state, log what it did and connect to your CRM and mailboxes. | Monthly, and someone has to keep it patched. |
| A person to run and maintain it | Checking the samples, clearing the exceptions, changing the rules when the process changes, fixing it when a provider changes its API. Part of a week, every week, for as long as it runs. | The biggest line, and the one that stops the moment that person is busy or leaves. |
The build is the smaller bill. A system like this runs on paid services and every one of them is metered. These are the lines – all of them sit inside our monthly price, and if you build your own they land on your own invoices, plus the one line nobody budgets for.
We don't publish the provider rates here because they change and depend on your volume – the proposal shows what's included for yours. The point is that the fee you compare against isn't the licence for an assistant. It's all six lines.
So build it if you want to, and use the coding tools – they're good. Just go in knowing that the build is a week and the design is a year, that the running costs above arrive monthly whoever built it, and that the year is mostly spent finding out which guard rails you were missing. A managed service or your own tool stack goes through the trade-off, and the pages on what we add to Salesforce and to HubSpot show the same work costed side by side.
Two things come first, whichever way you go.
Skipping them is the most expensive mistake in this whole subject.

Define what good looks like
A written profile of the companies worth going after, with the exclusions, and the personas of the people you want to reach.
The ICP workshops
Clean what you already have
One manufacturer we worked with had records that looked about three-quarters complete on paper. Once each was checked against the profile, nearer one in eight was a usable target.
Sort out the records you holdAI multiplies whatever you give it – good target accounts and you get more of them, rubbish and you get a lot more rubbish, very quickly. The data is everything.
More on both
Define what good looks like. A written profile of the companies worth going after, with the exclusions, and the personas of the people you want to reach. Without it, every tool, including an assistant, will produce contacts that look about right and aren't. Don't ask an assistant to write the profile for you – it doesn't know your customers, and it'll give you everyone's profile. The ICP workshops are how we get it out of your team's heads and tested against your own customers, with the research done for you, and how we score fit shows what the written profile turns into.
Clean what you already have. One manufacturer we worked with had records that looked about three-quarters complete on paper. Once each one was checked against the profile, nearer one in eight was a usable target, and the usual data providers had very little on the rest because they were small firms whose people aren't on the professional networks. Automation pointed at that CRM would have multiplied the problem.
Around 10 companies a week.
Up to 10 or so a week: do it yourself
With the setup at the top of this page. You'll be faster than anyone without an assistant and you'll know every account personally.
Beyond that: you need a system
Or as soon as the follow-ups, the replies and the CRM start to matter more than the first email. Ours is one option. Building your own is another. Doing it by hand in a chat window is the one that doesn't scale – and the fatigue is the model's, not yours.

It starts with a discovery meeting.
A conversation about your market, what's in your CRM today and what a good customer looks like for you. If there's a fit, you get a written proposal.