AI Agents for Lead Generation: What to Automate First

AI agents for lead generation take over the repetitive pipeline work: finding prospects, researching them, qualifying them against your criteria, drafting outreach and following up. Your team keeps the conversations and the decisions. Build in this order: research, qualification, drafted outreach, follow-ups, and only then automatic sending.

This is a build guide, not an overview. By the end you will have the inputs to write, the exact prompts to paste, the qualification rubric, the message sequences, the KPI table and a day-by-day plan for your first week. You can run all of it yourself. If you are new to the term, read what an AI agent is first, then come back.

How AI agents for lead generation work

A lead generation agent is a chain of small jobs. Each job has one input, one output and one check. Anthropic's engineering team describes this kind of design as a workflow built from patterns such as prompt chaining (each step feeds the next, with checks in between) and routing (classify an input, then send it down the right path), and separates it from fully open-ended agents that decide their own steps. For lead generation the chained, checked approach is the right one, because you want predictable output you can review. (Source: Anthropic, "Building effective agents", accessed 6 November 2026.)

Before you build: the four inputs

Agents amplify whatever you give them, so fix the inputs first. Write these four things in a document before you touch any tool.

1. Your ICP in plain language. Industry, company size range, country or region, buyer job titles, and the two or three signs that someone needs what you sell.
2. A definition of "qualified". A checklist a human could apply in 30 seconds. Section 3 gives you a rubric to copy.
3. Your offer in one paragraph. The problem, the outcome, who it is for, and what a first conversation covers.
4. An exclusion list. Existing clients, competitors, partners and anyone who has asked not to be contacted. The agent must check against this before anything is drafted.

Step 1: Define your ICP and signals (copy-paste prompt)

Start by turning your offer into a structured ICP. This prompt is adapted from the prompt library we use at SFC for new engagements. Paste it into the AI tool you use, fill in the angle-bracket fields, and keep the output as your ICP document.

Asset 1: ICP and signals prompt

You are a B2B go-to-market strategist.

Identify the ideal customer profile for the company below.

Company: <COMPANY_NAME>
Website: <WEBSITE_URL>
Offer: <ONE_PARAGRAPH_OFFER>
Target regions: <REGIONS>
Typical deal value: <DEAL_VALUE_RANGE>

For each ICP segment (maximum 3) define:
1. Decision-maker job titles (primary and secondary)
2. Industries and company size range
3. Buying signals that suggest they need this now (for example hiring for a relevant role, a leadership change, a new location, a product launch, a funding round)
4. The problem they are trying to solve right now, in their words
5. One qualifying question that separates a good fit from a poor one

Return a table with these columns:
Segment | Job titles | Industries and size | Signals | Core problem | Qualifying question

Then list 10 exclusion rules (types of company we should never contact).

How to check it worked: a colleague who has never heard of your business should be able to read the table and say yes or no to any company you name. If they cannot, tighten the signals and the qualifying question.

Step 2: Build the list and pilot on 10 to 20 leads

Pull a first list from the sources you already trust (a sales navigator export, a database you pay for, a client list, event attendees). Resist the urge to start big. In our setup the rule is to build the whole table end to end on a sample of 10 to 20 leads first, then estimate the cost of running it on the next batch of 100 to 300, and only then scale to the next few hundred. We do this because enrichment and AI steps cost money per row, and a flawed prompt multiplied across thousands of rows is an expensive mistake.

Your pilot table needs these columns, in this order:

1. Company, website, contact name, job title, profile link
2. Normalised company name (drop "LLC", "Ltd", "Inc" and legal suffixes so messages read naturally)
3. Company summary (two sentences from the website)
4. Signal (one recent, verifiable fact, with the source link)
5. Qualified (yes, no or maybe), Score, Reason
6. First line, Email draft, Status

Where your tooling lets you, use your own API key for the AI model rather than a platform's bundled credits, and check usage after the pilot. Re-run steps only on rows that failed or came back empty, never on the whole table.

Step 3: Qualify with a rubric the agent must justify

The most common failure in AI lead generation is a vague "is this a good fit?" prompt. Replace it with a rubric that forces evidence. Score each lead out of 10 using criteria you can verify from public information.

Asset 2: Qualification rubric

Decision rule: 8 to 10 is qualified, 5 to 7 is maybe (a human reviews it), 0 to 4 is not qualified. Now give the agent the rubric.

Asset 3: Qualification prompt

You are a careful B2B lead qualifier. Use ONLY the information provided. If a fact is not in the information, write "unknown" and score that criterion 0. Never guess.

ICP:
<PASTE_ICP_TABLE>

Exclusion rules:
<PASTE_EXCLUSIONS>

Lead:
Company: <COMPANY>
Website text: <WEBSITE_TEXT>
Contact title: <TITLE>
Region: <REGION>

Score using this rubric: industry match 0-2, size 0-2, region 0-1, decision maker 0-2, buying signal 0-3.

Return JSON only:
{
  "excluded": true or false,
  "scores": {"industry": n, "size": n, "region": n, "decision_maker": n, "signal": n},
  "total": n,
  "verdict": "qualified" | "maybe" | "not_qualified",
  "reason": "one sentence",
  "evidence": ["short quote or fact for each point scored above 0"],
  "signal": "the single best buying signal, with date, or none"
}

How to check it worked: pick five rows you know well. The verdict should match your own judgement on at least four. If it does not, the rubric or the input text is the problem, not the model.

Step 4: Write personalised first lines that are not generic

Personalisation that only inserts a name and a company is easy to spot. A good first line shows you read something specific and says nothing about you. Use the signal from Step 3.

Asset 4: First-line prompt

Write ONE opening line for a cold message.

Rules:
- Under 18 words.
- Start from a specific, verifiable fact about the company: <SIGNAL>
- Positive and factual in tone. No flattery words such as "impressive", "amazing" or "love".
- Do not mention our company, our offer or any pitch.
- Do not use exclamation marks or em dashes.

Company: <COMPANY_NORMALISED>
Contact first name: <FIRST_NAME>
Summary: <SUMMARY>

Output only the line.

The names above are fictional, used only to show the pattern.

Step 5: Build the outreach sequence

Use short messages. For email, keep subject lines under 7 words and avoid images and tracking-heavy formatting. For LinkedIn, keep the connection note to one sentence under 300 characters and DMs under 80 words. The cadence below follows the one we run in our own LinkedIn sequences: a connection request, then follow-ups roughly 2 to 3 days, 4 to 5 days and 7 to 10 days after the previous step.

Asset 5: Four-step email sequence

EMAIL 1 (day 1)
Subject: <COMPANY> and <PROBLEM_AREA>

<FIRST_LINE>

We help <ICP_TYPE> with <PROBLEM>. Teams usually come to us when <TRIGGER>.

Worth a short conversation, or is someone else the right person?

<SENDER_NAME>

EMAIL 2 (day 4) - proof or insight
Subject: Re: <COMPANY> and <PROBLEM_AREA>

One pattern we see with <ICP_TYPE>: <SPECIFIC_INSIGHT_IN_ONE_SENTENCE>.
If useful, I can send the two-page checklist we use for it.

EMAIL 3 (day 9) - what is working
Subject: <PROBLEM_AREA> checklist

<ONE_PRACTICAL_TIP_THEY_CAN_USE_WITHOUT_YOU>
Happy to share the full version. Want it?

EMAIL 4 (day 16) - close the loop
Subject: Should I close this out?

I will assume this is not a priority right now. If <TRIGGER> comes up later, reply here and I will pick it up. If you would rather not hear from me, tell me and I will remove you.

Asset 6: LinkedIn sequence

CONNECTION NOTE (under 300 characters)
Hi <FIRST_NAME>, I work with <ICP_TYPE> on <PROBLEM_AREA> and saw <SIGNAL>. Thought it would be useful to connect.

FOLLOW-UP 1 (2 to 3 days after they accept)
Thanks for connecting. Are you currently doing anything about <PROBLEM_AREA>? A lot of <ICP_TYPE> we speak to find it harder than expected.

FOLLOW-UP 2 (4 to 5 days later)
<FIRST_NAME>, most <ICP_TYPE> do not need more of <THING>, they need a better <PROCESS>. We built a simple way to do that. Want a quick walkthrough?

FOLLOW-UP 3 (7 to 10 days later)
No problem if now is not the right time. I will leave it here. If <TRIGGER> becomes a priority, message me.

Have the agent draft each message from these templates, but lock the structure: it may change the wording of the first line and the specific insight, never the ask or the opt-out language.

Step 6: Classify replies and set handover rules

Replies are where agents earn their keep and where they do the most damage if left unchecked. Classify first, then decide who acts.

Asset 7: Reply classification prompt

Classify the reply below into exactly one label:
interested | not_now | not_a_fit | wrong_person | referral | unsubscribe | out_of_office | question | unclear

Reply: <REPLY_TEXT>

Return JSON only:
{"label": "...", "confidence": 0-1, "one_line_summary": "...", "suggested_next_step": "..."}

Rules:
- Any request to stop contact is "unsubscribe", whatever else the reply says.
- If confidence is below 0.7, use "unclear".

A worked example (illustrative)

This example is illustrative. The company and all names are fictional, and the numbers are invented to show how the steps connect. They are not a client result.

The seller: a four-person HR advisory firm that helps growing companies set up hiring processes. The ICP: logistics and warehousing firms with 50 to 200 staff in the GCC, buyer is the founder or head of operations, signal is hiring for several operational roles at once.

Of four leads, two go to drafted outreach, one is held for the founder to judge, and one is dropped with a reason on file. The founder spends about ten minutes reviewing instead of an afternoon copying names. When Harbor Freight Co replies "can you send details?", the classifier labels it question, a draft with a booking link is prepared and the founder sends it. The agent never answers pricing or scope questions on its own.

Deliverability and compliance: the non-negotiables

Strong prompts do not matter if your messages land in spam or your domain gets blocked. Google publishes sender requirements for mail to Gmail accounts. As of 6 November 2026 its guidance says:

• All senders should set up SPF or DKIM authentication for their sending domains, use a TLS connection, and keep spam rates reported in Postmaster Tools below 0.30 percent.
• Senders who send more than 5,000 messages a day to Gmail accounts must set up SPF, DKIM and DMARC (the DMARC policy can be set to none), keep the sending domain aligned with SPF or DKIM, and support one-click unsubscribe for marketing and subscribed messages.
• Google advises keeping spam rates below 0.10 percent and never reaching 0.30 percent or higher.

(Source: Google, "Email sender guidelines", accessed 6 November 2026.)

Practical rules we apply regardless of volume: send from a secondary domain, never your main one; warm up new mailboxes gradually; include a plain opt-out line in every email; stop all sequences the moment someone replies or unsubscribes; and run a spam-check on each email draft before launch. Rules on unsolicited messages differ by country and channel. We are not giving legal advice, so check the requirements for each market you contact.

Decision rules and common mistakes

What to measure and how to read it

Asset 8: KPI table

Review these weekly on a sample of at least a few hundred sends before drawing conclusions. Change one thing at a time (the first line, the subject, the ask) so you know what moved the result.

Pre-launch QA checklist

• The pilot ran end to end on 10 to 20 leads and I read every output.
• Five spot-checked verdicts matched my own judgement on at least four.
• Every signal has a source link and a date.
• The exclusion list is applied before drafting.
• Every first line passes the rule: specific fact, no flattery, no pitch.
• Sending domain is authenticated (SPF, DKIM, DMARC) and is not my main domain.
• Every email has an opt-out line; the suppression list is connected.
• Reply classification routes unsubscribes to stop sequences automatically.
• A named person owns every "interested" reply and has a response-time target.
• Costs for the next batch are estimated from the pilot.

Your first week, day by day

1. Day 1: Write the four inputs. Run Asset 1 and edit the ICP table until a colleague can use it.
2. Day 2: Pull 20 leads. Build the pilot table columns. Run Assets 2 and 3 on all 20.
3. Day 3: Review verdicts, fix the rubric, re-run only the rows that disagree with you.
4. Day 4: Generate first lines with Asset 4 for qualified rows. Edit by hand until you trust the pattern.
5. Day 5: Load Assets 5 and 6 as drafts. Set up authentication and the suppression list. Send a first small batch (10 to 20) after approval.
6. Day 6: Add Asset 7 for reply handling and write the handover rules.
7. Day 7: Review the KPI table, note what to change, and decide the size of next week's batch from the cost you saw.

Frequently asked questions

Sources and further reading

• Anthropic, "Building effective agents" (workflows versus agents, prompt chaining, routing). Accessed 6 November 2026.
• Google, "Email sender guidelines" (authentication, spam rate and one-click unsubscribe requirements). Accessed 6 November 2026.
• SFC internal prompt library and enrichment SOP (methods adapted and anonymised).

What to do next

Do Day 1 today: write the four inputs and run the ICP prompt. If you would rather have the system built and running in your own accounts, we build AI outbound agents that source, qualify and reach out to prospects and push qualified leads into your CRM, and you can book a short call to talk through your pipeline. Once leads reply, our guide to the AI appointment setter covers turning conversations into booked calls.