AI prospecting: agents that find accounts, check them and draft the outreach

Most pages ranking for this search are written by a company selling a database, a sequencer or an AI SDR, and they define AI prospecting as whatever their product does. Tulina sells something too, an agentic RevOps workflow builder that runs prospecting as agents over your own tools, so read this page with that in mind. It explains the method first, in enough detail to run it without us, then shows where we fit. The two articles that go with it take up the two questions this one leaves open: what an AI SDR really does, and which AI prospecting tools do which job.

What is AI prospecting?

AI prospecting is prospecting where software that can read and reason does the research steps: deciding which companies are worth a message this week, who to write to there, whether they are already a customer or in someone's sequence, and what to say first. The output is the same as a good SDR's morning: a short list, each name with a reason, and a draft ready for a person to check.

It is narrower than "AI sales automation", which also covers call notes, forecasting, CRM updates and inbound chat. Prospecting is the part before the first conversation.

The job splits into five steps, and AI changes each one differently:

The first four steps are where AI saves the hours. The fifth is where it costs the most when it goes wrong, which is why the whole page keeps it human.

What is an AI sales agent, and is it the same as an AI SDR?

An AI sales agent is any AI system that performs sales tasks on its own, with tools, rather than answering questions in a chat. The term is broad: the top results for it include phone agents that answer inbound leads, website chat agents that qualify visitors, and research agents that work inside a CRM. An AI SDR is one kind of AI sales agent, the outbound kind, sold as a replacement for the sales development representative's job: source leads, write and send outreach, handle replies, book meetings.

The distinction matters when you buy. An AI SDR product takes the send and the replies off your hands, in your reps' names, on mailboxes it manages. An AI prospecting agent, in the sense this page uses, does the research and stops at a draft. What an AI SDR really does compares the two against an SDR's actual day, with the vendors' own prices.

How does AI prospecting work, step by step?

As a loop, not a list. A list is built once and decays from the day it is exported. A loop reruns every morning on the signals of the day before, so a company that posted the job you sell into on Tuesday is in a paused campaign on Wednesday.

1. Start from a signal, not a filter. A database filter on industry and headcount returns every company that could buy. A signal returns the ones with a reason to buy now: a job posted for the role your product helps, a technology adopted or dropped, a funding round, a new leader. Hiring and technology signals are the cheapest to read because they come from public job posts.
2. Score the accounts on what the search already returned. Before any paid lookup, an agent reads the fields the signal search gave back (headcount, industry, the job's text) and rates each account against a written rubric. Accounts below the line are dropped here, when dropping them is free.
3. Find the people by reading titles, not matching keywords. A people search by title is a keyword match, and titles are written by the people who hold them. An agent reads every title at the account and picks the two or three who own the problem. The measurement below shows how much that changes.
4. Check the CRM and every live sequence before buying anything. Match by email or company domain. Drop customers, open deals, and anyone a colleague is already working. This check is free; a revealed email for a customer is not.
5. Enrich only what survived, cheapest source first. One provider, then the next where it misses, stopping at the first verified email. Keep which provider found it and when next to the value.
6. Draft one opener per person, citing the signal with its date. "Saw you are hiring two SDRs in Austin (posted 21 October)" can be checked; "I noticed your impressive growth" cannot, and buyers can tell.
7. Stage, do not send. Load the leads into a paused campaign with the opener as a variable. A person reads the batch and launches it.

The order is the method. Run it backwards, enriching a whole export and filtering after, and the bill is set by the size of the export instead of the size of the list you actually wanted.

How much does a buying signal narrow the list?

More than most teams expect. We measured it on 29 October 2026 in our own workspace, on a segment a sales tool might target: US companies of 51 to 500 employees whose job posts mention HubSpot.

Two things follow. The signal did more work than any filter: one condition, a job posted in the last month, removed 96% of the segment, and what is left is the 4% with a reason to talk this month. And the keyword filter on titles found two of the seven leaders in the first 60 profiles (plus one further down the list). It missed "Sales Development Team Lead", "BDR Team Lead", "SDR Lead", "Business Development Lead" and "Director of Business Development", which are the people who would own an SDR tool. Reading the titles is a judgment step, and it is the step an AI agent does better than a filter.

What it cost: two TheirStack searches returning one company each, about 6 credits by TheirStack's published rate, and three Apollo people searches, which return identities without contact details and spent no Apollo credit. Nothing was revealed, enriched or sent. One segment on one day, with one company read in depth: a measure of how the method behaves, not of anyone's coverage.

How do you start AI prospecting yourself?

You need four things: a source of signals, a contact database, your CRM, and an AI assistant that can reach them. Each of the steps above can be done by hand with an assistant before anything is automated, and doing it by hand once is how you find out whether your rubric is any good.

1. Write the ideal customer profile as a rubric. Five or six criteria an account can pass or fail, with the evidence each one needs. "B2B software, 50 to 500 people, sells to mid-market, hiring SDRs" is a rubric; "high-growth companies" is not.
2. Pick one signal and pull a week of it. Job posts for the role your product helps are the easiest place to start; most databases and job-data providers expose them.
3. Score the accounts in a chat. Paste the rows and the rubric into Claude or ChatGPT and ask for a pass, fail or unsure on each criterion with the evidence quoted. Read the unsure ones yourself.
4. Find two or three people per kept account, by reading titles rather than filtering on them.
5. Check every person against the CRM before revealing an email.
6. Draft openers that cite the signal and its date, and read every one before it goes anywhere.
7. Send by hand, from a sequencer you already use, and count replies by signal, so the next week starts from the signal that worked.

This costs time, not tools: a few hours a week for a few dozen accounts. It stops scaling when the same steps have to run every morning across several segments, and that is the point where the steps become a schedule.

How do you run AI prospecting with Claude or ChatGPT, on a schedule?

With Tulina, Claude or ChatGPT runs the steps above as an agent: your contact database, your signal source, your CRM and your sequencer, each through its connector, in the order the method sets, with the spend shown before anything is bought. Run once from a chat, it builds one list. Saved as a process and given a schedule, it reruns every morning as a hosted agent and leaves the campaign paused for a person.

Paste this into Claude or ChatGPT once your database, CRM and sequencer are connected:

Build this week's prospect list from this sentence: "US B2B software
companies, 50 to 500 employees, using HubSpot, that posted an SDR or
BDR job in the last 30 days. Two or three people each: whoever runs
sales development or revenue operations."

1. Print the filters you read from the sentence and wait for my go.
2. Search for companies with that technology and that hiring signal.
   Give me the count first. Score each one on what the search returned
   and keep the top two tiers.
3. For each company kept, list the people and read their titles.
   Pick two or three who own sales development or revenue operations.
   Do not reveal anyone yet.
4. Check every person against my CRM by email or company domain, and
   against running sequences. Drop customers, open deals and anyone a
   colleague is working.
5. Find a work email for each person left, cheapest connected provider
   first, first hit wins. Show me the plan and its cost for one person
   as a dry run before running the rest.
6. Write one opener per person that cites the job post and its date.
7. Load them into my paused lemlist campaign with the opener as a
   variable. Do not launch anything.
8. Report: companies found and kept, people found, dropped by the CRM
   check, emails found by provider, and what each step cost.

The process behind that prompt, with every step and the tools it uses:

The process in Tulina

Build a target list from a sentence

Logistics companies in France, 250 to 1,000 employees, hiring supply chain roles. Source 40 accounts.

One-sentence brief in chat

See it in a demo

Claude runs the rest autonomously

1. 1

   AI shortlists accounts and their people.

   Apollo.io

   Apollo.io

   Matching companies are scored on what the search already returned, and each one kept gets two or three people fitting the persona described.

2. 2

   AI drops contacts the CRM already has.

   HubSpot

   HubSpot

   Anyone the CRM shows as a customer, an open deal or in a colleague's sequence is dropped, matched by email or else by company domain.

3. 3

   AI gathers signals for what survived.

   TheirStack

   TheirStack

   Each remaining company gets its open roles and technology, recent funding, moves and expansions, and details from its own pages, each with its source URL.

   Serper

   Serper

   Apify

   Apify

4. 4

   AI finds verified emails and numbers.

   Apollo.io

   Apollo.io

   Each contact gets a verified address and direct number, a second source fills what the first missed, and anyone still unfound is marked unreachable, not guessed.

   FullEnrich

   FullEnrich

5. 5

   AI writes each opener from a signal.

   Each opener names one dated finding about that company and cites its URL, and a company with nothing found gets a plainer version, never an invented hook.

6. 6

   AI stages the campaign, updates the CRM.

   Lemlist

   Lemlist

   The sequence sits as a draft with each contact's own opener, the qualified list is in the CRM with its signals as a note, and the launch is left to a person.

   HubSpot

   HubSpot

See the full process

For the version that starts from the signal itself rather than a sentence, Outbound from buying signals scores new hiring and technology signals each morning and fills a paused campaign. For a list you already have, Enrich contacts with email and phone runs the provider cascade alone.

Which tools do you need for AI prospecting?

Fewer than the listicles suggest, and probably ones you already pay for. The loop needs a signal source, a contact database, an email finder, a CRM, a sequencer and something that runs the steps in order. Some products bundle several of these (Apollo is a database, finder and sequencer in one seat), some are one layer (Clay for research and enrichment), and AI SDR products bundle all of them with the send. AI prospecting tools, sorted by the job they do lists them with prices read on each vendor's own page.

What AI prospecting can't do, and what gets you in trouble

• It cannot fix a vague ideal customer profile. An agent scores against the rubric it is given. A rubric of adjectives returns a confident list of the wrong companies, faster.
• Personalisation without a source is worse than none. A model asked to "personalise" will write a compliment about the company's growth. Make every opener cite a fact with its date and URL, or do not send it.
• Volume is the risk, not the saving. An agent can draft 500 emails as easily as 50. Sending them from one domain is how a company's mail starts landing in spam for everyone, sales and support included. Keep daily volume per mailbox low and warm new domains before using them.
• Data protection rules do not change because a model wrote the email. In the EU a cold B2B email rests on legitimate interest and a clear way to object (GDPR recital 47 names direct marketing as a possible legitimate interest); in the US, CAN-SPAM requires a truthful sender and a working opt-out. Keep the source and date on every record so you can answer "where did you get my address?"
• Signals are public, so your competitors read them too. A job post for an SDR team will bring several vendors into the same inbox in the same week. The reason to write gets you opened; the offer still has to be relevant.
• Replies need a person. A reply asking about price, a referral to a colleague, or an angry opt-out each need judgment that a wrong answer turns into a lost account.

Common questions

Is AI prospecting legal under GDPR and CAN-SPAM?

The rules are the same as for prospecting by hand, because the law looks at the data and the message, not at who wrote it. In the EU, cold B2B email to a named work address usually rests on legitimate interest, which needs a relevant offer, a clear way to object and a record of where the data came from. In the US, CAN-SPAM asks for a truthful sender and subject, a postal address and a working opt-out. An agent makes the record easier to keep, since every row can carry its source and date, and it makes mistakes faster, which is why the send stays a human step.

How much does AI prospecting cost?

Three meters add up: the data (contact and company lookups, usually billed per record found), the model (usually a fraction of the data cost per account researched and drafted), and whatever runs the steps (a seat, a platform plan or your own time). The data is almost always the largest line. That is why the order of the steps matters more than the choice of model: a list scored on free fields before anything is revealed costs a fraction of one enriched first and filtered after.

Will AI prospecting replace SDRs?

It replaces most of the research and list work that fills an SDR's day, and very little of the conversation. The teams getting results use agents to find, check and draft, and keep a person on the send, the replies and the calls. The AI SDR page in this guide goes through what an autonomous agent really does against the job description.

Can I do AI prospecting with ChatGPT or Claude alone?

For research, yes: either will read a company's site and news and draft an opener. For prospecting as a process, not on their own, because a chat cannot search a contact database, check your CRM, write to a table or run again tomorrow. Connected to those tools through MCP, and run on a schedule, the same assistant becomes the agent this page describes.