Skip to content
← All posts

AI Agents

How to build an AI research agent for B2B account research

Build an AI agent that researches B2B accounts automatically: inputs, sources, prompt design, structured output, scoring and human review.

Victor Mendonça
Victor Mendonça

Founder, OretonGTM · ·3 min read

The most expensive part of good outbound used to be research. An SDR spending ten minutes per account can research maybe a few dozen accounts a day, and the quality varies with their energy level. An AI research agent does the same work for every account, consistently, in seconds.

This is how I build them.

What the agent should produce

Start from the output and pick the tools later. For each account, I want a structured record like this:

Field Example
What they sell Payroll software for mid-size companies
Who they sell to HR and finance leaders, 100–1,000 employees
Go-to-market motion Sales-led, SDR team, hiring AEs
Relevant signals Raised Series B, 4 open SDR roles
ICP fit score 1–5 with a one-line justification
Suggested angle Ramp new SDRs faster with automated research
Evidence Links or quotes supporting each claim

Structured output is what makes the research usable downstream for scoring, routing and personalization.

My take: An AI agent without structured output is just a very fast intern writing essays nobody reads.

Step 1: Gather the inputs

LLMs are only as good as what you give them. Before the model runs, collect:

  • Website content: homepage, product, pricing and customers pages
  • Job posts: especially sales, marketing and ops roles
  • News and announcements: funding, launches, leadership changes
  • Enrichment data: headcount, industry, tech stack, location
  • Persona context: the title and seniority you’re targeting

In Clay, each of these is a column. In n8n, each is a step calling a scraper or API.

Step 2: Design the prompt

The research prompt I use has five parts:

  1. Role: “You are a B2B sales researcher preparing account briefs for [your company].”
  2. Offer context: two sentences about what you sell and the problems you solve.
  3. Inputs: only the gathered content for this account.
  4. Output schema: the exact fields above, in JSON.
  5. Rules: use only the provided inputs, cite evidence, return “unknown” instead of guessing.

The “unknown” rule is the most important line in the prompt. It’s what keeps the agent honest.

Step 3: Score and route

Use the structured output to act:

  • High fit + signal: move to Tier 1, deeper personalization, alert the rep.
  • Good fit, no signal: standard automated sequence.
  • Low fit: exclude, or monitor for future signals.

This plugs straight into signal-based outbound and TAM tiering.

Step 4: Use the research everywhere

The same research record feeds:

  • Personalization of the first line and the angle. See AI personalization in cold email
  • Call prep for SDRs and AEs
  • CRM records, attached to the company for anyone who touches the account

Step 5: Keep a human in the loop

Before scaling, review a sample of research records by hand:

  • Are the facts correct and supported by evidence?
  • Is the fit score consistent with your judgment?
  • Are “unknown” answers used when they should be?

Then keep reviewing a small weekly sample. Models, prompts and websites change.

Cost and scale

Research agents are cheap per account compared to human time, but costs add up at scale. Run the agent only on accounts that already passed your ICP filters, and cache results so you don’t research the same account twice.

Building and maintaining research agents is the core of our AI research agents service, always with human review before anything reaches a prospect.

Frequently asked questions

What is an AI research agent in sales?

It's an automated workflow that uses an LLM to research companies and contacts by reading websites, job posts and news, and returns structured insights that sales teams use for prioritization and personalization.

Which tools can I use to build an AI research agent?

Clay is the most common choice because it combines data sources and AI columns. n8n is a good option for custom logic and integrations. Both can use models like Claude or GPT.

How accurate are AI research agents?

Accuracy depends on the sources you feed the agent. Grounding the model in scraped content, asking for evidence with each claim and allowing 'unknown' answers reduces hallucinations significantly.

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