AI Lead Research and Sales Agent

What does the agent actually do with a new lead?

It does the reading that a salesperson would do if they had twenty spare minutes per lead, and then stops. Here is the sequence we normally build, in order.

1. Receive and clean. A lead arrives from a form, an inbox, a list or a CRM view. The agent standardises the company name, domain and contact fields and checks whether the person or company is already in your CRM.
2. Check history first. Past emails, open deals, earlier notes and any previous "not now" are pulled before anything outside is read. An existing customer or a lead someone is already working should never be researched as if new.
3. Read public sources. The company site, its careers and product pages and, where permitted, other public pages are read and summarised, with the source address kept next to each point.
4. Score against your rubric. Each rule is applied separately and returns met, not met or unknown, with its evidence.
5. Draft a note. Two or three sentences that refer to something real from the brief, in your team's tone, saved as a draft.
6. Queue and record. The brief, score and draft go to a review queue, and the CRM is updated with everything marked as agent-written.

How is the fit score built, and who controls it?

You do. The score is not a model's gut feeling; it is a rubric your team signs off, applied one rule at a time. The table shows the kinds of rules sales teams typically write. These are illustrative examples, not defaults.

The "unknown" row matters most. A model that treats missing information as a negative will quietly throw away good leads from companies with thin websites.

What can it read, and what can it not?

Be sceptical of any agent that promises to "know everything" about a prospect. Realistic sources and their limits:

• Company websites: useful but often out of date, and some sites block automated readers or load content in ways a simple fetch cannot see.
• Your CRM and mailbox: the most reliable source, provided your records are tidy. An agent cannot fix years of duplicate contacts by itself, though it can flag them.
• Professional-network sites: most prohibit scraping in their terms, so we do not build scrapers against them. If you need that data, use a licensed provider whose terms allow your use.
• Paid data providers: reliable only to the extent the provider is, and the licence usually limits how the data may be stored and reused. Read it before you plan around it.
• Personal data about individuals: we keep this to what is needed for a business conversation, and we avoid collecting anything about private life.

Why should the agent never send cold email on its own?

There are three reasons, and the first is not legal.

Quality. A language model can state a wrong fact in a confident sentence. In a private brief that is a nuisance; in an email to a stranger it is a first impression you cannot take back. A person reading two lines catches most of these.

Reputation. Mailbox providers watch complaint and bounce rates. A tool that sends at machine speed can damage the domain your normal business email also uses.

Consent rules. Marketing rules differ by country and by recipient. In the UK, the Privacy and Electronic Communications Regulations treat email to a limited company differently from email to a sole trader or some partnerships, and UK GDPR still applies to the personal data of a named person at a business. Other countries have their own regimes. The ICO publishes B2B marketing guidance. This is general information, not legal advice, so confirm your position with a lawyer or data-protection adviser.

Because a human approves each message, there is a point at which someone checks the recipient type, the contact basis and the suppression list. Our agent supports that by showing the recorded basis next to each draft and blocking drafts for anyone on a do-not-contact list.

Build, or use a sales-intelligence SaaS?

Products from data vendors and sales-engagement platforms already cover contact data, sequences and basic scoring. Those are often the quicker choice, particularly if your rubric is simple and your CRM is on their supported list.

A custom agent earns its cost when your criteria are specific to you, when the research must combine your CRM history with outside pages, or when you want drafts held for review rather than released into a sequence. If you are weighing it up, our n8n workflow service covers the plumbing that sits underneath, and this explainer on agents, chatbots and automations sets out where an agent is genuinely needed.

How do you know it is working?

Before launch we replay a batch of past leads whose outcomes you know and compare the agent's scores with your team's judgement. After launch we track the edits people make to drafts and the leads they reject; both are measurements of agent quality that cost nothing extra to collect. Our guide on testing and evaluating AI agents covers the method in more detail. Approvals and review queues are explained in the human-in-the-loop article.

When you should not buy this

• You receive a handful of leads a week. A person with a checklist is cheaper and better.
• You have no written idea of what a good lead looks like. Write that down first; no software can supply it.
• Your plan is to send high volumes of automated cold email. We will not build that.
• Your CRM data is badly broken. A clean-up should come before any agent.

For inbound enquiries that need an immediate answer rather than research, an AI chatbot or an AI customer support agent is a better fit. If you would like to talk it through, tell us about your lead flow and we will say honestly whether an agent is worth building.