AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
The usual mistake is to point an AI at a list of companies and let it write and send a thousand "personalised" emails. It feels efficient because the first line mentions the prospect's company. It fails because the model fills gaps with plausible guesses, the list was never checked for who you may lawfully email, and nobody reads what goes out under your name.
This guide builds a smaller, safer system: an AI that researches and drafts, a table that holds the evidence, and a human gate before anything is sent.
The rule: let AI do the reading and the drafting, but never let it send a claim you cannot trace to a source, to a person you were not allowed to email.
You need four things. None of them requires a custom app.
| Piece | What it does | Cost at time of writing |
|---|---|---|
| A list of target companies | Limited companies in a niche you understand | Free if you build it from public sources |
| A research agent | Claude (or another assistant) that reads a company's own pages and fills one row per company | Needs a paid plan for sustained use; check the current £ price at checkout on the vendor's page |
| A database | A spreadsheet or Airtable-style table with one row per prospect | Free tiers exist; check the current plan |
| A sending tool | A mailbox or outreach platform on a domain you control | Varies by tool; check the £ price at checkout |
Claude's custom connectors (remote MCP servers) let it reach external tools. Anthropic's help centre says they are available across Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector. Confirm that on the current page before relying on it.
You also need to know the law. In the UK, the Privacy and Electronic Communications Regulations (PECR) treat businesses differently. Per the ICO, the consent rule for email does not apply to "corporate subscribers" (companies, LLPs, Scottish partnerships and some public bodies), but sole traders and some partnerships count as individual subscribers and need consent or the soft opt-in. Fines under PECR can now reach the greater of £17.5 million or 4% of global turnover, in force from 5 February 2026. You are very unlikely to face that for a careful small campaign, but it is the reason the steps below exist.
subscriber_type with one of: corporate, individual, unsure.individual or unsure rows unless you have consent or a valid soft opt-in.Use a role address or a named business contact at the company, and record where you found it. Do not guess address formats.
Create a table with these columns: company, website, subscriber_type, contact name and role, contact source URL, evidence 1 (quote plus URL), evidence 2 (quote plus URL), pain-point hypothesis, confidence (high, medium, low), draft email, status (new, researched, approved, sent, replied, opted out), approved by, date sent.
Add a second tab called suppression with every address or domain that has opted out. This is the most important tab in the file. Anything on it is never emailed again.
Give the assistant one company at a time (or a small batch), with access to its public website. Do not let it browse the open web for "insights" about individuals.
You are a careful B2B research analyst. I run [YOUR_BUSINESS] and I sell [YOUR_OFFER] to [TARGET_NICHE].
Goal: for the company below, find at most two specific, publicly stated facts that suggest a need for my offer, and say whether that evidence is strong enough to justify an email.
Company: [COMPANY_NAME]
Website: [WEBSITE_URL]
My offer in one sentence: [OFFER]
What a real fit looks like: [FIT_CRITERIA]
Steps:
1. Read only the company's own public pages (about, services, news, careers).
2. List up to two facts. For each, quote the exact words and give the page URL.
3. Write one pain-point hypothesis, labelled clearly as a hypothesis, not a fact.
4. Rate confidence: high (two quoted facts support it), medium (one), low (none).
5. If confidence is low, say "do not email" and stop.
Output as a table with columns: fact, exact quote, URL, hypothesis, confidence.
Constraints: do not infer anything about named individuals. Do not invent figures, clients, or events. If a page cannot be read, say so rather than guessing. If any input above is missing, ask me before starting.
Before answering, check that every quote appears word for word on the page you cite.
Fill in the square brackets with your business, offer, niche and fit criteria. Paste the result into the evidence columns.
You are a UK B2B copywriter. Write a first-contact email using only the evidence I give you.
Sender: [YOUR_NAME], [YOUR_BUSINESS], [POSTAL_OR_REGISTERED_ADDRESS]
Recipient: [CONTACT_NAME], [ROLE] at [COMPANY_NAME]
Evidence (quote and URL): [EVIDENCE_ROWS]
Offer: [OFFER]
Single ask: [ONE_SMALL_ASK]
Requirements: British English; under 120 words; opening line refers to one piece of the evidence; plain statement of who I am and why I am writing; one ask; no flattery, no fake urgency, no claims about results I have not supplied; end with a one-line way to opt out ("reply 'no' and I will not email again").
If the evidence is weak or missing, reply "insufficient evidence" instead of drafting. Before finishing, check that every factual claim in the email maps to a quoted source above.
Save the draft in the draft email column and set status to researched.
approved by with your initials. Nothing is sent without that.suppression the same day. The ICO notes that individuals have an absolute right to object to direct marketing, and recommends honouring objections from corporate contacts too.corporate row has a legal-form check you can point to.