Skip to content
← All posts

AI Agents

AI for sales prospecting: 10 practical use cases that work

10 practical ways to use AI for B2B sales prospecting: qualification, research, signals, personalization, replies and call prep, with prompts.

Victor Mendonça
Victor Mendonça

Founder, OretonGTM · ·6 min read

Most discussions about AI in sales swing between hype (“AI SDRs will replace your team”) and dismissal (“AI emails are all garbage”). The useful middle ground is less exciting: AI is extremely good at specific prospecting tasks, and mediocre at others.

The tasks it’s good at share a pattern: reading and classifying unstructured information at scale. Websites, job posts, news, LinkedIn profiles, email replies. Humans do these slowly; AI does them in seconds, consistently enough to be useful when properly guided.

Here are 10 practical use cases, how to set them up and the guardrails each one needs.

The principles first

Before the use cases, four principles that apply to all of them:

  1. Ask specific questions. “Research this company” produces fluff. “Does this company sell to enterprise customers? Answer yes, no or unknown, and quote the evidence” produces data.
  2. Give the model the source. Pass the website text, job post or article. Don’t rely on what the model “knows”.
  3. Use structured outputs. Ask for JSON or fixed fields so results can be filtered and used downstream.
  4. Keep a human in the loop where mistakes are expensive: messages to important accounts, replies, anything public.

My take: The quality of AI prospecting depends far more on the question you ask and the data you give it than on which model you use.

1. Account qualification

What it does: checks accounts against ICP criteria that databases can’t filter on.

Examples:

  • Is this a B2B company?
  • Do they sell a product or a service?
  • Do they appear to have a sales team?
  • Do they serve a specific market (e.g., healthcare)?

How: scrape the homepage and a few key pages, pass the text to the model with a list of yes/no/unknown questions, and require a short evidence quote for each answer.

Guardrail: sample and review results manually at first; refine the questions where the model gets confused.

See AI lead qualification.

2. Account research

What it does: builds a concise research brief on an account: what they do, who they sell to, recent news, likely priorities.

How: combine website content, recent news, job posts and LinkedIn company data. Ask for a structured brief with specific fields, not a free-form essay.

Guardrail: each claim should come with a source. If the model can’t find something, it should say “unknown”.

See building an AI research agent.

3. Signal detection in job posts

What it does: reads job descriptions to detect signals: a team being built, tools in use, problems they’re trying to solve.

Example questions:

  • Is this role related to [function we support]?
  • Which tools are mentioned?
  • What problems or goals does the description mention?

Why it’s powerful: job posts are public statements of priority. A post for “SDR — help us build our outbound motion from scratch” tells you a lot. See hiring signals.

4. News and event detection

What it does: scans news and company announcements for relevant events: expansion, launches, partnerships, leadership changes.

How: pull recent news for target accounts and ask the model to classify each item by relevance to your offer, with a short explanation.

Guardrail: filter out irrelevant or old items; check dates.

See signal-based outbound.

5. Persona matching

What it does: decides whether a person is the right contact, based on title, profile summary and company context.

Example: “Given this title and profile, does this person likely own outbound sales? Answer yes, no or unclear, and explain in one sentence.”

Why: titles are messy. “Growth Lead” at one company owns outbound; at another, it’s paid ads.

6. Personalized opening lines

What it does: writes one or two specific, relevant lines per account based on research.

How: pass the research brief, your segment’s core message and strict rules:

  • Max 25 words
  • Must reference a specific, verifiable fact from the research
  • Must connect to the problem in the email
  • No flattery, no “I noticed”, no exclamation marks
  • If no relevant fact exists, return “NONE”

Guardrail: human review for Tier A accounts; spot checks for the rest. See AI personalization and personalization prompts.

7. Segment-level email drafts

What it does: helps write and iterate the core message for a segment.

How: give the model your ICP, segment problem, proof points and examples of good emails. Ask for several angles, not one “best” email. Then edit by hand.

Why: AI is useful for generating options and challenging your assumptions. The final message should still be yours. See how to write a cold email.

8. Reply classification and drafting

What it does: classifies every reply and, optionally, drafts a response.

Categories: positive, not now (with date), wrong person (with referral), not interested, unsubscribe, out of office, question.

How: classify with an LLM, update the CRM, stop sequences, route positives to the owner. For drafts, generate a suggested reply for a human to edit and send.

Guardrail: never auto-send responses to positive or nuanced replies. See classify replies and AI reply drafting.

9. Call prep

What it does: prepares a one-page brief before each sales call: company summary, person background, signals, likely priorities, suggested questions.

How: trigger a workflow when a meeting is booked, gather data and send the brief to the rep before the call.

See AI call prep.

10. CRM hygiene

What it does: standardizes job titles, classifies industries, detects duplicates and flags stale records.

Why it matters: clean data makes everything else work, from scoring to reporting. See CRM hygiene automation.

Choosing the right tools

Need Common tools
Run AI across lead lists Clay (AI columns, research agents)
Research and reasoning Claude, OpenAI models
Workflow automation n8n, Make, Zapier
Scraping source content Clay, Apify, custom scrapers
CRM integration Native integrations, webhooks

See Clay AI columns, Claude for sales and n8n sales workflows.

Writing prompts that work

A good prospecting prompt has:

  1. Role and goal: what the output will be used for.
  2. Input: the source content, clearly delimited.
  3. Specific questions with allowed answers.
  4. Rules: length, tone, what not to do.
  5. Output format: JSON or fixed fields.
  6. An escape hatch: “If the information isn’t in the source, answer ‘unknown’.”

Example:

You are qualifying companies for outbound.
Use ONLY the website text below. Do not use outside knowledge.

Questions:
1. sells_b2b: yes / no / unknown
2. has_sales_team: yes / no / unknown
3. target_customers: short phrase or "unknown"

For each answer, include a short quote from the text as evidence.
Return JSON with keys: sells_b2b, has_sales_team, target_customers, evidence.

Website text:
"""
{website_text}
"""

See how to stop AI hallucinations in sales.

Measuring impact

Measure AI prospecting by outcomes:

  • Hours saved per rep per week on research and admin
  • Qualification accuracy: share of AI-qualified accounts a human agrees with
  • Positive reply rate of AI-assisted campaigns vs. your baseline
  • Meetings per 1,000 contacts
  • Error rate: factual errors found in sampled messages

If AI saves time but reply rates drop, quality slipped. If reply rates rise and errors stay low, scale it.

Common mistakes

  1. Vague prompts. “Research this company” produces generic summaries.
  2. No source content. The model fills gaps with guesses.
  3. Free-form outputs. Hard to filter and use downstream.
  4. Using AI to send more instead of send better. Volume isn’t the goal.
  5. No review. Small error rates become many mistakes at scale.
  6. Automating strategy. AI can execute a strategy; it can’t choose your ICP.

Where to start

Pick one use case with high time savings and low risk: account qualification or research are usually best. Build it for one segment, review 50 outputs, refine the prompt, then expand. Once that’s working, add personalization and reply classification. For the bigger question of how much of the SDR role AI should take on, see AI SDR vs human SDR.

Frequently asked questions

How can AI be used in sales prospecting?

Common uses are qualifying accounts against ICP criteria, researching companies and people, detecting buying signals in job posts and news, drafting personalized lines, classifying replies, preparing call briefs and keeping the CRM updated.

What's the best AI tool for sales prospecting?

It depends on the workflow. Clay is popular for running AI across lead lists; Claude and OpenAI models are common for research, classification and drafting; n8n and similar tools connect AI steps to your CRM and sequencer.

Can AI write cold emails?

AI can draft good cold emails and personalized lines when given clear research, a defined structure and constraints. Fully AI-written emails without review tend toward generic phrasing and occasional factual errors.

How do I stop AI from making things up about prospects?

Give the model the source content, ask it to answer only from that content, require it to quote evidence, allow an 'unknown' answer and review a sample of outputs regularly.

Is AI prospecting expensive?

Model costs per account are usually small. The bigger costs are data, the tools that run AI at scale and the time to design and maintain good prompts and workflows.

Will AI replace sales prospecting jobs?

AI is replacing many prospecting tasks, especially research, list building and admin. The parts that need judgment, conversations and relationships remain human, and reps with good AI workflows can cover more accounts with better quality.

Ready to stop doing outbound by hand?

Book a strategy call. We’ll map your ICP, review what’s working (and what isn’t) and show you exactly what we’d automate first.

Book a strategy call

Keep reading