Lead Research Agent for B2B Agencies: Sourced Briefs, Human Outreach

What does the agent hand a rep at the end?

A one-page account brief, created as a note on the company record. It is meant to take the place of twenty minutes of tab-hopping, not to replace the rep's judgment. The fields are fixed so briefs can be compared:

• What the company sells and to whom, in two sentences, with the page the summary came from.
• Signals relevant to your offer, such as a redesign mentioned in a blog post or roles advertised on a careers page, each with a link and the date it was seen.
• A fit score against your rubric, with the reasons listed criterion by criterion.
• Two or three possible contacts from sources you approved, with titles as stated there.
• Two suggested openers, each tied to a cited fact, and a line stating what the agent could not find.

The brief is the only output. Sending, sequencing and follow-up belong to people, or to tools you already run under your own consent rules.

Which rules come before any automation?

Lead research is where an agent can create legal and reputational trouble fastest, so the rules come first. This is general information, not legal advice.

• Site terms. LinkedIn's User Agreement lists, under its prohibited conduct in section 8.2, using bots or scripts to scrape its services and using unauthorized automated methods to access them. We would not build an agent that scrapes LinkedIn, and we would read the terms of any other site before letting an agent visit it.
• Who counts as a business. The ICO's guidance on business-to-business marketing separates corporate subscribers from individual subscribers under the UK's PECR rules, and sole traders and some partnerships are treated as individuals. It also says that data protection law still applies when you hold a named person's details. An agent that cannot tell which type of company it is looking at should flag the record, not guess.
• Your own lists. Customers, opted-out contacts and anyone on a do-not-contact list are removed before research starts.
• Other countries. Rules for cold email and calls differ by country and US state. Ask your lawyer before any automated outreach, and treat the agent as research support only.

Our post on AI personalization and compliance covers the wider privacy picture.

Illustrative scenario: one target account, traced stage by stage

A worked example with an invented company, not a client result. Suppose an agency sells conversion audits to software firms. A rep adds Example Analytics Ltd, a made-up company, to the target list and sets its status to "to research".

Stage	What happens	Tool called	Output
1. Trigger	The status change fires a workflow with the company record id	CRM webhook	A queued job
2. Screen	Checks the domain against customers and the do-not-contact list	CRM lookup	Proceed or stop with a reason
3. Read	Fetches the home, about, pricing, careers and blog pages, honouring robots.txt	HTTP fetch	Page text with URL and fetch date
4. Extract	Pulls facts into the brief fields, each with its source link	Language model, structured output	Fact list; unknowns marked
5. Score	Applies the rubric and writes the reason for each criterion	Rubric prompt	Score with reasons
6. Draft	Writes two openers, deleting any line it cannot tie to a source	Language model	Draft text
7. Approval	A task appears for the rep, who edits, accepts or rejects the brief	CRM task	Reviewed brief
8. Write-back	Saves the brief as a note and stores the email draft unsent	CRM API	Updated record
9. Audit	Logs pages read, facts used, score, rep decision	Log store	One row per run

The rep, not the agent, decides whether Example Analytics gets a message at all. Our AI lead research and sales agent page describes the service in general terms.

Who decides what: alone, with approval, never

The agent does alone

• Reads approved sources and records what it found with links.
• Scores accounts against the rubric and explains the score.
• Writes briefs and draft openers into the CRM.
• Flags duplicates, missing data and uncertain company types.

A person approves first

• Whether an account enters outreach at all.
• Every email, connection request or call script, before it is sent by a person or by a tool under your consent rules.
• Any change to the rubric or the source list.
• Adding a new data provider.

Never automated

• Scraping sites whose terms prohibit it, or getting around a login or rate limit.
• Guessing or constructing personal email addresses.
• Sending messages, or enrolling people in sequences, without a human decision.
• Claims about a prospect that have no source, or that pretend a relationship exists.

How can it go wrong?

What if the website is out of date?

Each fact carries the fetch date and the page, so the rep can see how old it is. Facts older than a limit you set are shown but marked stale rather than used in an opener.

What if it confuses two companies with similar names?

The agent matches on domain and the registered name together. When they disagree it stops and flags the record for a human to confirm.

What if the website text contains instructions?

Pages the agent reads are untrusted. Text on them is treated as content to summarize and never as a command, and the agent holds no tool that can send anything. See our post on prompt injection in AI agents.

What if the personalization sounds creepy?

Openers may use only company-level facts the business itself published. Personal social profiles, family details and anything that suggests monitoring are excluded by the source policy.

What if the score is simply wrong?

That is why the first stage is calibration against accounts your team has already judged. Disagreements are reviewed, and the rubric wording is fixed, not the score argued with.

How would success be measured?

We do not quote an uplift. We would agree on metrics before the first brief and compare against your own earlier numbers:

• Research time per account, sampled before and after.
• Brief acceptance rate: briefs accepted as written, edited, or rejected.
• Score agreement between the agent and your reps on a held-back set of accounts.
• Source error rate: cited facts a rep found to be wrong or missing.
• Reply rate and meeting rate for outreach that used briefs, against comparable outreach that did not, with the caveat that small samples mislead.
• Complaints and opt-outs per batch, which act as a brake on volume.

If this fits how your agency sells, talk to us, or see how it connects with n8n workflow automation.