How to Find B2B Prospects With AI Research Tools

Start from your ten best customers, have an AI assistant turn what they share into a written ideal-customer profile, then use the research modes in ChatGPT, Claude or Perplexity to list matching companies from public sources. Find the right person in a contact database such as Apollo or Sales Navigator, and verify every company and name before contacting anyone.

The research is now the quick part; checking it is the real work. AI research tools will confidently list a plant that closed two years ago or a manager who left last spring. A realistic afternoon produces 20 to 40 checked prospects, each with a named contact and a reason to get in touch now, which is plenty for a small firm doing personal outreach. The method below is written around an illustrative manufacturer of spare parts for packaging machinery, but every step works for any business selling to other businesses.

Follow me on Instagram@sagnikteaches

Step 1: learn from the customers you already have

Your best prospects look like your best customers, and you have more information about them than any database. Pull your top ten customers by profit (not revenue) and describe each in a row, leaving out names and contact details. The spare parts manufacturer's table looked like this:

Connect on LinkedInSagnik Bhattacharya

Then ask an AI assistant to find the pattern. Use a business plan that doesn't train on your content, or switch off model training in your privacy settings first.

Subscribe on YouTube@codingliquids

Here are my 10 most profitable customers (anonymised).
1. What do the top 5 have in common that the bottom 5 don't?
2. Which features look like real buying reasons, and which
   might be coincidence given only 10 rows?
3. Write a one-paragraph description of the ideal customer
   I should look for, and list 3 things that would disqualify
   a company.

An illustrative answer: "Your top five all run flow-wrapping machines over ten years old and order through a maintenance or engineering manager. Four of the five are in snacks, bakery or confectionery. Company size ranges from 95 to 310 staff, so size looks less important than machine age. Possible disqualifiers: plants with new machines under the manufacturer's service contract; very small sites with no dedicated maintenance staff…"

That's useful, but check the second question's answer carefully. In this run the AI claimed "referral" was a buying reason because three of the top five came by referral; with ten rows that's likely coincidence, and the owner knew two of those referrals came from one person. Keep the machine-age insight, which matches what sales conversations say, and drop the referral one. If you want to go deeper on building profiles from real customer data, see creating customer personas with AI from real data.

Step 2: turn the pattern into a search brief

Write the profile as a brief that both you and an AI tool can search with. Here's the manufacturer's, filled in:

SEARCH BRIEF: packaging spare parts
Companies:   Food manufacturers (snacks, bakery, confectionery,
             chilled and frozen) with their own production plant
Size:        About 50-500 staff at the site
Must have:   Horizontal flow-wrappers or tray sealers on site
Strong sign: Machines more than 10 years old; job ads mentioning
             packaging-line maintenance; recent line expansion
Buyer:       Maintenance manager, engineering manager, plant
             manager (smaller sites), maintenance planner
Rule out:    Sites under the machine maker's full service contract;
             contract packers we already supply; sites that
             only do storage or distribution
Region:      [the area you can deliver to within 24 hours]

The "rule out" list is what saves time later. Without it, the AI's lists fill with distribution warehouses and head offices that have "foods" in the name but no production line.

Step 3: build the company list from sources AI can read

AI research tools are good at reading many public pages quickly. Point them at sources that genuinely list the kind of company you want:

• Trade association member directories for your customers' industry (not yours).
• Exhibitor lists from trade shows your customers attend or exhibit at.
• Job adverts, which reveal what equipment a site runs ("experience maintaining horizontal flow-wrappers").
• Trade press and local business news about openings, expansions, new lines and investment.
• Planning and building notices for factory extensions, where these are published.
• Suppliers' and machine makers' case studies naming the plants they've installed kit at.

A research prompt built from the brief:

Using the search brief below, find 20 food manufacturing sites
in [region] that are likely to run flow-wrappers or tray sealers.
Search trade association directories, trade-show exhibitor
lists, job adverts and news from the last 18 months.
For each site give: company, site location, product type,
evidence it has a packaging line (quote it), source URL,
source date. Exclude head offices and warehouses.
If you can't find evidence of a packaging line, leave it out.
Don't guess. Return a table.
[paste search brief]

An illustrative run returns 20 rows. When checked, the result typically splits like this: 14 are good (a real production site with clear evidence), 3 are real companies but the "evidence" is weak or generic ("produces a range of snacks"), 2 are distribution depots despite the instruction, and 1 source link goes to a page that doesn't mention the company at all. That last kind of error is the one to watch for; the AI has linked a plausible page it didn't actually use. Open every source link and see how to check sources and citations in AI research for a faster routine. Then run the prompt again with "exclude these companies" and the list of ones you already have, to get the next batch without repeats.

Which assistant to use matters less than people think; the paid research modes in all the main assistants can do this. Perplexity versus ChatGPT for business research compares the two most common choices.

Step 4: find the person who buys

Titles to search for

From your customer table you already know who orders. For the manufacturer: maintenance manager, engineering manager, maintenance planner, reliability engineer and, at smaller sites, the plant manager. Purchasing or procurement staff often process the order but rarely choose the supplier for urgent spares, so they're a second contact, not the first.

Where to find them

LinkedIn Sales Navigator is the most direct way to see who holds a role at a specific site: filter by company, function (engineering, operations) and title keywords, and save the results as a lead list. Core costs $119.99 a month or $1,079.88 a year (see LinkedIn's plan comparison). It doesn't give email addresses. For those, use a contact database such as Apollo or an email finder such as Hunter, both of which have free plans for testing; the comparison of AI lead generation tools covers prices and strengths.

Don't overlook the company's own website and trade press. A maintenance manager quoted in an industry magazine article about reducing downtime is exactly the person you want, and the article gives you something to mention.

A realistic snag: at a 60-person bakery, there may be no maintenance manager at all, just an owner who also runs the factory floor and one maintenance technician. The search returns nobody, and the prospect gets wrongly discarded. The fix is a rule in your brief: for sites under about 80 staff, look for the owner or operations director instead.

Step 5: find a reason to get in touch this month

A good prospect with no reason to talk to you now is a list entry. A good prospect with a timely reason is a conversation. Ask the AI to look for signals like these for each company on your checked list:

A signal prompt, run per company:

For [company], [site], find anything from the last 6 months
that suggests they may need packaging spares or a new
supplier: new maintenance or engineering staff, job adverts
mentioning packaging lines, expansions, new product lines,
machine replacements, or production problems in the news.
Quote the source and give the date. If nothing, say "none found".

Treat production problems with care. A plant that had a publicised breakdown may well need a reliable spares supplier, but opening with "I read about your line failure" reads as ambulance-chasing. Use it to decide who to contact, not what to say.

Step 6: verify before anything goes in the CRM

Everything the AI found gets checked once, by a person, before it becomes a record. The checks that matter, and how to do each in under a minute:

• The company exists and the site is operating. Website loads, recent news or job ads in the last six months.
• It fits the brief. The quoted evidence actually says what the AI claims (a production line, not a depot).
• The person is still in the role. Their LinkedIn profile or the company website shows the current title.
• The email is verified. Your contact tool marks it valid, not "accept-all" or guessed.
• They're not already a customer or a lost deal. Search your CRM by company domain, not just company name.
• The source is recorded. Note where the contact came from in a "lead source" field.

A mistake that shows why the fourth step matters: the manufacturer's list included a "maintenance manager" whose profile said he'd moved to a competitor's plant eight months earlier. The contact database still had him at the old company. Emailing his old address would have bounced; worse, a later email to him at the competitor would have referenced the wrong plant. Two minutes on his profile caught it.

Once the contact is checked, it goes into your CRM with its source and signal. If your CRM is already full of duplicates, fix that first, or every new list makes it worse.

Step 7: score the list and choose who to contact first

You'll rarely have time to contact everyone in a batch properly, so rank them. A simple rubric, filled in for three prospects:

Contact the 8s to 10s personally this week. Put the 6s and 7s into a slower sequence or wait for a signal. The chilled foods site is a good example of a strong fit with no reason to call: set a reminder to re-run the signal prompt in a month rather than sending a generic email now. For a fuller scoring approach once you have more data, see AI lead scoring for small businesses.

Turning the research into a first message

Research is wasted if the first email could have been sent to anyone. Give the AI the checked record, not just a name, and ask for a short message that uses one fact from it. For the confectionery plant that scored 10:

Write a first email (under 90 words, plain text, no links)
to the new maintenance manager at this site.
RECORD: confectionery plant, 280 staff; runs horizontal
flow-wrappers (job ad, March); she joined in July from a
bakery; we supply flow-wrapper jaws, knives and belts with
48-hour dispatch; we supply 3 similar plants.
Use one fact from the record. One question at the end.
No flattery, no "I hope this finds you well".

Before (what a generic prompt produced): "Dear Maintenance Manager, I hope this email finds you well. We are a leading supplier of high-quality spare parts for the packaging industry and would love the opportunity to discuss how we can support your operations."

After (from the record): "Congratulations on the move to [plant] in July. Your March job ad mentioned horizontal flow-wrappers; we make the jaws, knives and belts for those and dispatch most orders within 48 hours, which three other confectionery sites rely on. When a wrapper goes down, who currently supplies your wear parts, and how long do you usually wait?"

The second version is shorter, specific and ends with a question she can answer in one line. Check every fact in it against the record before sending; an AI will sometimes "improve" a date or a count.

An afternoon's numbers for a spare parts manufacturer

Here's how one four-hour session could look for the illustrative manufacturer, with a sales engineer doing the work and ChatGPT Plus plus an Apollo free plan as the tools:

The quick sum: nine well-researched approaches from four hours of work. If two of them lead to a first order over the next quarter and a typical customer spends $40,000 a year, that afternoon is worth a great deal. Even if none converts this quarter, the other 27 checked records are a pipeline for the next three months. The lesson in the table is where the time goes: the AI saved hours of searching, and nearly two hours still went on checking, which is exactly the part not to skip.

Once you're ready to send those first emails, the set-up and writing rules in writing cold emails with AI that stay out of spam keep a well-researched list from being wasted in junk folders.

Staying on the right side of data rules and site terms

A few rules keep prospect research safe and defensible:

• Work contact details are personal data. Under data-protection law such as the GDPR, a named person's work email and phone number are personal data. Collect only what you need, record where it came from, and be ready to tell anyone who asks. Whether you can email them without prior consent depends on where they are; check with a data-protection adviser before your first campaign.
• Respect site terms. LinkedIn's terms forbid scraping and automated activity, and tools that do either get accounts restricted. Research by hand or with Sales Navigator's own features.
• Keep a do-not-contact list. Anyone who asks not to be contacted goes on it, and you check it before every batch.
• Don't feed personal details into tools that train on them. For the customer-pattern step, anonymise; for contact research, use tools whose terms you've read.

Run the method as a weekly or fortnightly habit rather than a one-off project. Small batches, checked properly and contacted within days, beat a thousand-row spreadsheet that goes stale in a drawer.

Prospecting with AI: follow-up questions

Which AI research tool is best for finding B2B prospects?

For building company lists, any of the paid research modes in ChatGPT, Claude, Perplexity or Gemini will do; Perplexity is quick for cited lists, while ChatGPT and Claude write fuller research reports. None of them is a contact database. For named buyers and verified emails you need a tool such as Apollo, Hunter or LinkedIn Sales Navigator alongside the chat assistant.

How many prospects should I research at once?

Work in batches of 20 to 50 companies. That's small enough to verify properly in an afternoon and to contact personally within a week or two, before the reasons you found go stale. Researching 500 at once usually means half the list is out of date by the time you reach it, and verification gets skipped.

Can I paste my customer list into ChatGPT to find patterns?

Remove what the AI doesn't need first. For pattern-finding it needs company type, size, products bought and spend band, not contact names, emails or phone numbers. Use a business plan that doesn't train on your data by default, or switch off model training in a consumer plan's privacy settings, and check your own privacy notice covers analysing customer records this way.

Further reads

• How to Do Competitor Research With AI in One Afternoon — Knowing who your prospects buy from now sharpens every approach.
• How to Write Sales Call Scripts With AI That Don't Sound Scripted — What to say when a researched prospect picks up the phone.
• How to Clean Up a Messy CRM With AI — Keeping new prospects from turning into duplicate CRM records.
• AI Lead Generation for Small Businesses: What Actually Works — Where outbound research fits among other lead sources.
• How Small Businesses Use AI in Sales: 10 Real Examples — More examples of small firms using AI across the sales process.
• Is It Safe to Put Customer Data Into ChatGPT? — The safe way to analyse customer records with an AI assistant.
• How Consultants Use AI for Client Research Before Discovery Calls — A research routine for discovery calls: scale the prep to the deal, make AI cite everything, check what matters and turn findings into sharper questions.
• AI for Recruitment Business Development: Winning New Clients — A weekly business-development routine for agencies: hiring signals, five-minute account briefs, outreach that names the vacancy, and consented spec profiles.
• AI Lead Generation vs Hiring a Salesperson: A Cost Comparison — Itemised costs of an AI prospecting stack and a salesperson, a caterer's year-one comparison per meeting and per client, and when each wins.
• How to Write LinkedIn Posts With AI Without Sounding Like AI — Start from a two-minute voice memo, let AI tighten without rewriting, and remove the tells readers spot in a scroll, with a before-and-after post and hooks.
• How to Write a Sales Deck With AI That Wins Meetings — Research one buyer, draft the storyline in plain text, turn every headline into a claim, and only then let an AI slide tool do the design.
• What Does an AI SDR Cost, and What Does It Actually Do? — Published AI SDR prices from about $1 a lead to $3,750 a month, the mailbox and staff costs around them, and how to work out cost per booked meeting.
• Is Perplexity Pro Worth Paying For in a Small Business? — Who in a small business gets $20 a month of value from Perplexity Pro, how to test it against the assistant you use now, and the privacy setting to change.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: LinkedIn Sales Navigator plan comparison page; Apollo.io and Hunter pricing and product pages; ChatGPT, Claude, Perplexity and Google AI plan pages. Checked September 2026.