What Smart Sales Professionals Know About Using AI for Prospecting

The way most reps prospect with AI is to ask a chatbot for a list of fifty companies and call it a day. The list is plausible, the names are real, and the conversion rate is a disaster. The reason is not that AI cannot prospect. It is that prospecting is a four-step process and most reps are running step one and skipping the rest.

Salesforce State of Sales found that high-performing reps are 1.7x more likely than underperformers to use AI agents for prospecting specifically. The gap is in the workflow, not the access. Here is the four-prompt chain the high performers are running, and what each prompt is actually doing under the hood.

Prompt one: build the account list with explicit fit criteria

The mistake is asking for “companies that would be a good fit.” That phrase produces the average of every B2B target list on the open web. The fix is to write your ICP in disqualifying language, what would rule a company out, before you ask for any names.

“You are my sales research assistant. I sell [PRODUCT] to [CUSTOMER PROFILE], priced at [PRICE]. Below is my ICP written as disqualifying criteria, what would make a company a wrong fit:

- under [N] employees (too small, no procurement function)

- over [N] employees (too large, our product gaps in enterprise security)

- in [LIST INDUSTRIES] (regulatory misfit)

- using [LIST COMPETING TOOLS] in a multi-year contract

- headquartered in [LIST GEOGRAPHIES] (no local support)

Generate 25 companies that pass all of these disqualifying criteria. For each, include the official company name, headquarters city, employee count band, and the one observable signal that suggests they would care about our problem (recent funding, recent hire, public job ad, recent acquisition, public statement). Do not invent signals. If you cannot find a real signal, mark the row NEEDS RESEARCH instead of guessing. List follows below your output.”

The disqualifying criteria do the work. They force the model to skip the over-mentioned obvious targets and surface mid-tier accounts you would not have researched manually. The NEEDS RESEARCH instruction keeps the list honest, since the model will otherwise fabricate a “recent Series B” for a company that has never raised.

Prompt two: find the right buyer inside each account

A list of fifty company names is not a prospecting list. A list of fifty named decision-makers is. The default behaviour is to ask for “the CMO,” which produces titles not people, half of which the model has guessed.

“For each of the 25 companies above, identify the most likely buyer of [PRODUCT]. Use this priority order:

1. Functional owner whose budget would pay for this (job title contains [LIST OF QUALIFYING TITLES])

2. If multiple, pick the one with the most recent role change (new hires often have budget mandate)

3. If still multiple, pick the one whose public statements (LinkedIn posts, conference talks, podcast appearances) suggest they care about our problem

Output format: company, named buyer, exact title, the source URL where you verified them, and the date of the source. If the source is older than 18 months, mark STALE. If you cannot find a verified source, mark UNVERIFIED and do not invent a name.”

The source URL instruction is the prompt’s whole value. Without it, the model invents plausible-sounding names that look real until you try to email them. With it, you get a smaller list of verified contacts plus an honest count of accounts you still need to research manually.

Prompt three: write the opening message against a specific signal

Cold outreach without a specific signal is spam. Cold outreach against a specific, recent, observable signal is reception. The difference is whether the buyer can verify that you actually researched them.

“Below are five verified contacts from the previous step. For each, write a 60-to-90-word opening email that does this:

- opens with a single concrete sentence referencing the specific signal we identified for their company (recent hire, funding, public statement, etc.)

- second sentence connects that signal to a problem we solve, not a feature we have

- third sentence offers a specific resource (case study, benchmark, comparison) we can send if they reply, not a meeting request

- closes with a question they can answer in one line

Banned: ‘I came across’, ‘I noticed’, ‘congrats on’, ‘just reaching out’, ‘quick question’. No flattery. No name-checking the company unless the company is in the signal sentence. Five emails follow, one per contact.”

The banned-phrases list is doing the heavy lifting. Those phrases are AI prospecting tells. Every email you ship that opens with “I noticed” trains buyers to skip your name.

Prompt four: stress-test the list before any email goes out

Before any of those messages enter your sequencer, run them past the model as a sceptic. This is the prompt almost no rep runs and the prompt that catches the embarrassments.

“Read the five emails above. Assume the role of the named buyer at the named company receiving this message cold. For each email, answer:

- Is the signal claim verifiable? Could the buyer fact-check it in 30 seconds and find we got it right?

- Is the connection to our product believable, or is it a stretch?

- Does any sentence sound AI-generated? Flag specific phrases.

- Would you mark this as spam if you received it?

Score each email 1 to 5 on overall send-readiness. Recommend revisions for any below 4.”

The model is not a perfect judge, but it catches the obvious failures. A claim about a Series B that was actually a seed round. A “connection to our product” that does not match the buyer’s stated priorities. Catching three of these per batch is the difference between a sequencer that works and one that gets your domain throttled.

The counterintuitive insight, and the failure mode

Smaller lists outperform larger ones consistently. A rep running 100 verified, signal-anchored, custom-opening emails per week beats a rep running 500 templated emails by a wide margin, both in reply rates and in closed revenue. AI made it cheap to send 500 templated emails, which means inboxes are full of them, which means the marginal value of each collapses. The work that still costs effort, picking real signals and writing real openings, is the work that still gets replies.

The failure mode follows from this. AI is a multiplier on whatever workflow you point it at. If your account list is wrong, AI will scale up your bad targeting. If your messaging is templated, AI will scale the same template across 500 contacts and you will hit a deliverability ceiling within the month. The fix is to run the four-prompt chain manually for two weeks against a small list, measure reply rates, and only then move into a tool that automates it.

Apollo, Clay, and ZoomInfo handle account discovery and contact verification with fresher data than a chatbot alone. Outreach and Salesloft handle the sequencer side and increasingly the AI-drafted opener. ChatGPT and Claude remain the right home for the stress-test prompt, since the model should be operating outside the tool that wrote the email.

What changes when this clicks

Prospecting stops being the part of your job you avoid until Friday afternoon. The chain takes about an hour for fifty accounts, the messages that go out are specific, and the reply rate stops being depressing. The less visible change is that you start prospecting against more interesting targets, because the disqualifying-criteria prompt surfaces accounts you would never have manually researched. That widening of the funnel, while keeping per-message quality high, is the actual edge AI gives you. Not faster bad outreach. More targeted good outreach, at a pace you could not previously sustain.