Automate Your Own Outbound: Lead Generation on Companies House Data

An AI automation agency that prospects by hand is a bit like a builder living in a tent. If you sell automation, your own outbound is the obvious first workflow to build, and once it works it doubles as a demo you can sell to clients.

This guide sets out automated lead generation on Companies House data. You pull new companies every day, enrich them with contact details, let an LLM draft only the first line of each email and have a person approve every message before it goes. The parts that need judgement stay with a human, and the machine does the copying and drafting.

It is written for AI automation agencies and technical founders. The steps work with whatever email, CRM and orchestration tools you already use, so we don't name any.

Why Companies House data suits automated lead generation

Most lead sources are snapshots, out of date the day you buy them. The Companies House register changes every working day, and every new limited company has to appear on it with a name, a registered address, its directors and up to four SIC codes describing what it does. That gives you a fresh, structured feed with a built-in timing signal, since a company registered yesterday is choosing its suppliers now.

The volume is steady and predictable. Here is what a week of incorporations looked like in Company Explore's registration data:

Across the 30 days to 18 October 2026 there were 62,561 incorporations, about 2,800 on an average weekday and around 600 on a weekend day. Nobody should email all of them. Filter by SIC code and region and a daily pull shrinks to something a person can review over coffee. Management consultancy (SIC 70229) is one of the largest codes, with 3,740 new companies in those 30 days, or about 125 a day on average. Most narrower niches give you a handful a day, which is the right size for a pipeline with a human checkpoint.

The AI lead generation workflow, step by step

Each step below hands a clean record to the next. Use the company number as the unique ID throughout, because company names change and two companies can have very similar names.

1. Pull yesterday's new companies. Either run a search in the B2B List Builder with your SIC codes, regions and an incorporation date of yesterday, then export it, or connect Claude or another MCP client to the Company Explore MCP server (available on Entry, Pro and admin accounts) and have it build the list for you each morning.
2. Filter and deduplicate. Drop anything that isn't active, drop holding company codes such as 70100 and 64209, and drop any company number already in your CRM or suppression list. If one director has formed three companies this week, keep one record and attach the others to it.
3. Enrich. An enriched export adds each company's published website, email and phone number, plus scored suggestions for the director's LinkedIn, Facebook or Instagram profile. Files come as CSV, JSON and XLSX, so JSON can go straight into your pipeline. Over MCP, the same enriched export can be started from your client.
4. Qualify. Apply simple rules first: no published email, skip it; a holding page instead of a website, tag it for a later follow-up. Then send what's left to a person for a quick look at the company profile.
5. Draft the first line. Give the LLM three fields only: the company name, the SIC code description and the town. Ask for one sentence that connects the company's industry and location to the problem you solve. The rest of the email is your own template, written once and reused.
6. Review and approve. A person reads every draft, checks the first line against the facts and either approves, edits or rejects it. Nothing sends without a click from a human.
7. Send at a controlled rate. Send from a real mailbox with a real signature, a few at a time across the day, with a working opt-out in every message.
8. Log and learn. Record the SIC code, region, company age and outcome for every send. After a few weeks, cut the segments that only produce unsubscribes.

Keep the LLM on a short lead

The first line is the only part that needs to be personal, and it is also where a model can do the most damage. Give it more data and it will happily use it, including guessing at things it can't know, like the founder's background or the company's plans.

Three inputs keep it honest. "A new property management firm in Leeds" is enough to write a relevant opening, and every fact in that sentence comes from the register. Tell the model not to invent clients, results, compliments or anything else that isn't in the input, and to return nothing if it can't write a line from what it was given. Then check the output in code before the human sees it: if the line mentions a number, a name or a place that wasn't in the input, flag it.

Avoid lines that pretend to know more than you do. "Congratulations on your exciting launch" reads as automated to anyone who got ten of them that week. A plain observation about their industry and town reads as if someone looked.

Guardrails that keep the pipeline safe

An automated pipeline will repeat a mistake as many times as you let it, so build the limits in from the start:

• Human review: a person approves every message. This is the main defence against wrong facts, odd tone and the occasional company name that turns out to be a joke.
• Opt-outs: every email identifies you and includes a simple way to opt out. Add every opt-out to a suppression list of company numbers and email addresses, and check that list at step 2 so nobody is ever re-added.
• PECR: limited companies and LLPs are corporate subscribers, so you can usually email them at corporate addresses without prior consent if you identify yourself and offer an opt-out. Sole traders and some partnerships count as individuals and need consent first.
• UK GDPR: a director's name and work email are still personal data. Legitimate interest usually covers relevant B2B prospecting, but record your assessment, keep only the data you use and tell people where you got their details. Our GDPR guide covers the detail.
• Profile matches are suggestions: the scored LinkedIn, Facebook and Instagram matches in an enriched export are likely candidates that nobody has confirmed yet. A person should check a match before anything uses it.
• Deduplication: match records on the company number, and limit each director to one message even when they run several new companies.
• Rate limits: send slowly from a mailbox with a sending history, raise volume gradually and stop automatically if bounces or complaints jump. On the data side, page through results instead of re-running large queries.

Turn your pipeline into a client offer

Once the workflow runs for you, it is a product. Most of your clients have the same problem you had, and the pipeline barely changes between them: swap the SIC codes, the regions and the email template, keep the guardrails. A web designer might target new companies with no website, as in our guide to finding web design clients. A logo designer would target the same new companies for a different reason, which we covered in how logo designers win new business clients.

Price the setup as a fixed project and the daily running as a monthly fee, and tell the client plainly that a person on their side approves every send. It tends to reassure clients who worry about compliance. For help choosing which industries to target first, see which new UK companies need AI automation.

Frequently asked questions

What is automated lead generation?

It is a workflow that finds, enriches and prepares prospects with little manual work, usually ending in a drafted message. The safest versions automate the data handling and drafting, and keep a person in charge of who gets contacted and what they receive.

Can I use Companies House data for cold email?

Yes, for limited companies, if your message is relevant to their business, you identify yourself and you include an opt-out. The register gives you company and director details; the published email and phone come from enrichment.

Should an LLM write the whole email?

We'd keep it to the first line. A short template you wrote yourself does the explaining, and limiting the model to one sentence built from three register fields makes errors easy to spot in review.

How do I connect Claude to Company Explore?

Entry, Pro and admin accounts can connect Claude or another MCP client to the Company Explore MCP server, then ask it to build lists, search companies and directors and run enriched exports. Plans and access are listed on the pricing page.

To start your daily feed, open the B2B List Builder, pick your SIC codes and regions and set the incorporation date to yesterday.