Skip to content
NagentNagent
Log inSign upHire your AI Growth Team
NagentNagent
Log inSign up

How to tune AI outbound research for our ICP

8 Minutes read
Published:
AI outbound research is only as sharp as the rules you give it. Turn your ICP into filters, disqualifiers and two or three dated signals, ask research questions with an evidence standard, test on a hand-labelled sample and feed outcomes back.
PB
Pratap BeheraNov 20, 2026·8 min read
How to tune AI outbound research for our ICP

Tune AI outbound research by turning your ideal customer profile into rules the agent can check, not a paragraph it has to interpret. Write hard filters and disqualifiers, name the two or three dated signals that tend to come before a purchase, give the research specific questions with an evidence standard, then test it on a hand-labelled sample of accounts before scaling. After launch, feed meetings and closed deals back so the weights follow what actually converts.

The rest of this guide covers each layer with examples, a test plan and the failure modes to watch. We make Nagent, whose outbound research lead is NORA; how NORA is tuned is in its own section near the end.

What does tuning AI outbound research actually mean?

An AI research agent makes four decisions for every account. Each is a separate dial:

  1. Who to find. Which companies and roles enter the list at all.

  2. What to look for. Which facts and events matter for your offer.

  3. What counts as evidence. Which sources are acceptable, and how recent they must be.

  4. What to do with it. Whether the account is a fit, how it ranks, and what angle the first email takes.

Most "the research is generic" complaints come from leaving dials 2 and 3 unset. The agent was told who to research but not what would change your decision.

How do you turn your ICP into rules an AI can apply?

Start from your best customers, not your hopes. Apollo's ICP guide puts it simply: "A high-impact ICP is built from your best current customers" (Apollo). Then split what you learn into three lists.

  • Must-haves. Industry, company size band, region, and the role that signs. An account missing any of these is out.

  • Disqualifiers. Current customers, open deals, competitors, partners, companies below a size you can serve, regions you cannot sell into, and anyone who opted out.

  • Nice-to-haves. Tech stack, funding stage, team structure, growth rate. These rank accounts; they do not admit them.

Write each as something checkable. "Mid-market SaaS" is vague. "B2B software company with 200 to 2,000 employees, headquartered in India or Singapore, with a head of sales enablement or revenue operations" is a rule.

Label what is evidence and what is a guess. Unify's ICP workflow advises separating attributes your customers support from assumptions to test (Unify). An assumption should be tested, not weighted like a fact.

Which signals should the research look for?

Pick two or three observable changes that usually come before someone buys what you sell. Unify calls these first signals and gives examples by segment, such as compliance hiring in financial services or new facilities in manufacturing.

Good signals share three properties:

  • They relate to your problem. A new head of sales enablement matters if you sell enablement training; a funding round alone may not.

  • They are dated. Set a freshness window, such as 30 to 90 days, and treat older signals as background.

  • They are sourced. A job post URL, a press release, a filing. If the research cannot link it, it should not appear in an email.

Common signal types: hiring for a specific role, a leadership change, funding, a tool added or dropped, a new office or market, a stated pain in a public interview or post.

How should you write the research questions?

"Research this company" produces a summary of the About page. Ask the questions your best rep would ask before a call, and fix the output shape.

A weak brief:

  • "Research Acme and tell me about them."

A tuned brief:

  • "Is Acme a fit on our must-haves? Answer yes, no or unknown for each, with a source."

  • "Has Acme posted a sales enablement or RevOps role in the last 90 days? Give the URL and date."

  • "Which of our three use cases is most likely relevant, and why, in one sentence?"

  • "What is unknown that a rep should ask on the first call?"

Ask for a fixed structure every time: fit verdict per rule, signals with source and date, one suggested angle, and open questions. 11x's Deep Research lets users add directives such as "focus on compliance pain points" and returns cited reports (11x). Clay's Claygent lets you write the prompt yourself, test it on real rows and see a reasoning trace for each decision (Clay). Whatever the tool, the principle is the same: one clear question per field, with a source required.

How do you test the tuning before you scale?

Do not judge tuning on sent emails. Judge it on a sample you have labelled yourself.

  1. Pick 50 accounts: 20 you know are great fits, 20 you know are poor fits, 10 you are unsure about.

  2. Run the research on all 50 without telling the agent which is which.

  3. Compare its verdicts with yours. Count good fits it rejected and poor fits it admitted.

  4. Read every source it cited for ten accounts. Check that each link exists and says what the research claims.

  5. Fix the rule, not the account. If it admitted a poor fit, find the missing disqualifier and add it.

  6. Rerun until poor fits stop getting through, then start small with live outreach.

This takes an afternoon and saves weeks of meetings with the wrong people.

How do you keep the research tuned after launch?

Close the loop with outcomes. Unify suggests ranking signals by meetings per 100 people contacted, showing the sample size beside each rate and not ranking signals with fewer than 100 contacted (Unify). Then:

  • Weekly: read a sample of research and drafts, and log every disqualified meeting with the rule that should have caught it.

  • Monthly: rerun signal attribution and drop signals that do not produce meetings on their own.

  • Quarterly: revisit the ICP against new closed-won and closed-lost deals. Clay's glossary calls the ICP "a living document" to review quarterly or semi-annually (Clay).

What goes wrong most often?

  • Inferred values treated as facts. The tool guessed your target industries from your website, and nobody checked.

  • No disqualifiers. Customers, competitors and students end up in the list.

  • Stale signals. A funding round from last year becomes the email's opening line.

  • Unsourced claims. The research says a company is "expanding into Europe" with no link, and the email repeats it.

  • One ICP for every product. If you sell two things, give each its own profile and signals.

  • Tuning on opens. Opens are unreliable; tune on positive replies, meetings held and deals.

How do AI outbound vendors let you tune research?

From each vendor's own pages, checked 24 October 2026:

  • 11x. Lets you define what a good lead looks like with criteria such as industry, size, tech stack, funding and hiring; its agents read the web to verify fit before outreach, and failed leads are tagged with a reason rather than deleted (11x qualification).

  • Unify. Drafts an evidence-based ICP from your website and CRM customers, labels assumptions, and suggests first signals to test.

  • Clay. Gives you the most direct control: you write each research prompt, test it on real rows, and see a reasoning trace. You also build the logic yourself.

  • Apollo. Filters prospects by firmographic, technographic and behavioural attributes and scores leads by ICP match.

  • Artisan. Its product guide describes a 250M+ contact database filterable by title, industry, geography and company attributes, plus intent signals; it does not describe how the ICP is set in the product (Artisan).

Where Nagent fits

We make Nagent. "Nagent is multiplayer AI for growth: people and AI agents in one shared context, delivering the number." NORA, the AI Outbound Research Lead on SERA's Sales team, is tuned through the same layers:

  • Your ICP in your words. On the Who you sell to screen you describe your customer in plain sentences, and Nagent works out the targeting. Results come back in two groups, From what you wrote and What we worked out, so inferred values are visible. You correct them by editing the description and running it again (how it works).

  • Scored, then revealed. Each company is scored 0 to 100 against the profile, with industry, size and geography weighing most, and contacts are revealed only at 70 or more.

  • Dated, sourced signals. Signals carry a source and date, go stale after 30 days, and a person can confirm or dismiss each one.

  • A cited dossier. Research produces an overview, why the company fits, its signals and pitch hooks, and every claim in a draft must trace to it.

  • Outcomes feed back. Insights breaks results down by industry, persona, signal and score band, and Run tuner proposes weight changes once enough leads have closed.

What it does not do: NORA does not import your CRM history to build the profile, and it writes email only. Work is metered in credits; see pricing.

Frequently Asked Questions

What is the difference between an ICP and a buyer persona?

An ICP describes the company that fits: industry, size, region, triggers. A buyer persona describes the person inside it: role, goals, objections. Research needs both: the ICP to admit accounts, the persona to pick who to write to.

What is signal-based outbound?

Outreach timed to an observable change at the account, such as a hire, funding round or tool change, with the signal as the reason for writing. It works best when the signal is recent, sourced and relevant to your offer.

How many accounts should I test AI research on before scaling?

Enough to see errors. A hand-labelled sample of about 50, mixing known good and poor fits, shows whether the rules let the wrong accounts through.

Should AI research use CRM data or only the web?

Both, if you can. Web research finds signals; CRM data supplies exclusions such as current customers and open deals, and the history of what closed.

See how NORA turns your ICP into scored, cited research: book a demo.

AI agents for SaaS & Technology

Continue learning

Related agents

Agents that match this read.

Browse all 200+ agents
Select Category
Sales

How to Research Accounts for Outbound With an AI Agent

How to research target accounts for outbound with an AI agent: define the ICP, score fit, find accounts, read the buying signals, build a sourced dossier, map the committee and qualify with evidence before anyone is contacted.

Sales Automation

Find ICP-Fit Accounts and Research Them with AI

AI finds ICP-fit accounts by scoring prospects against your closed-won profile, then researching each match into a sourced dossier before outreach. The full loop is profile, discover, research, then write.

Sales

AI SDR booking meetings that don't fit our ICP, how to fix the targeting

When an AI SDR books meetings outside your ICP, the cause is usually a loose list, a metric that rewards any meeting, no qualification step between reply and booking, or copy that attracts the curious. Here is how to diagnose and fix each.

Sales Automation

What is an AI outbound research agent?

An AI outbound research agent finds accounts that match your ideal customer profile, researches each one into a sourced dossier, and drafts the first outreach email for a human to approve before anything is sent.

Sales

How do AI outbound agents personalize messages without sounding generic?

Generic AI outreach comes from thin inputs, not weak models. Good outbound agents research first, write from a dated signal and your own positioning, check every claim against its source and stop when the research is thin.

Sales

Top 10 AI SDR Agents for Outbound Prospecting (2026)

Compare the 10 best AI SDR agents for outbound prospecting, from Nagent NORA, 11x and Artisan to Unify, Clay and Regie.ai, by the job each does best.