ChatGPT for sales: prompts that keep the facts checkable

This guide is for SDRs, account executives and founders who already use ChatGPT, or a similar chat assistant, and want better results for sales work. It covers what ChatGPT does well and badly, a context block to paste first, nine ChatGPT prompts for sales, a worked example including the fact-check, and the data rules to follow. For dedicated tools that do more than a chat assistant, see the guide to AI sales tools.

What ChatGPT is good and bad at in sales

The evidence on writing is strong. In a preregistered experiment, Noy and Zhang (MIT, 2023) gave 444 college-educated professionals realistic writing tasks, including delicate emails, and randomly gave half access to ChatGPT. Time taken fell by 10 minutes (37%) against a control group averaging 27 minutes, and blinded graders rated the work 0.45 standard deviations higher. Weaker writers gained most, and the work shifted from rough drafting toward generating ideas and editing. That's the right model for sales: you supply the facts and the judgment, it does the first draft.

The evidence on facts is the opposite. A 2025 paper by Kalai and colleagues argues that language models hallucinate because training and evaluation reward guessing over admitting uncertainty. For prospecting, the data source matters more than the model. In GTM-Bench, a June 2026 benchmark from Blackpearl Group (Bebop's parent company), the same GPT-5.5 agent scored a net +4,041 with access to a structured company database but only +1,015 when restricted to web evidence.

What not to paste into ChatGPT

Before you paste anything, check your workspace's data controls: who can see your chats, whether conversations can be used to improve models, and how long they're kept. Settings differ between personal and business plans, so use the account your company has approved. Then keep these out unless your company's policy explicitly allows them:

• Customer contracts, pricing exceptions and discount approvals
• Call recordings or transcripts from calls where participants weren't told about recording
• Personal details beyond what the task needs (home addresses, personal phone numbers)
• Anything a customer shared under a confidentiality agreement
• Login details, API keys or exports of your whole CRM

Start with a context block

ChatGPT knows nothing about your offer. Paste this once at the start of a project (or save it as project instructions) and every later prompt gets better.

You are helping me with B2B sales writing. Facts about my business:

WHAT WE SELL: [one sentence, in the buyer's words]
WHO BUYS: [industries, company size, the roles that decide]
THE PROBLEM WE SOLVE: [in the buyer's terms]
PROOF I CAN USE: [customer results I have permission to mention]
WHAT WE DON'T DO: [so you never promise it]
MY TONE: [e.g. plain, direct, no hype, short sentences, US English]

RULES FOR EVERY ANSWER:
- Use only facts I give you. If you add anything, mark it [UNVERIFIED].
- If you don't know something, write "unknown". Never guess names, numbers, titles or dates.
- No exclamation marks, no "I hope this finds you well", no buzzwords.

Nine ChatGPT prompts for sales

Write a cold email under 90 words to [name, title] at [company].
Verified facts (with source and date): [fact 1], [fact 2].
Open with the most relevant fact, connect it to one problem we solve, and end with a low-effort question, not a meeting request.
Give me two versions with different openings.

With web search on, summarize [company name, website domain] for a sales call.
For each point give the source URL and its date. Cover: what they do, size, recent changes in the last 12 months, and likely priorities for a [role].
If two companies have similar names, say so. Write "unknown" for anything you can't source.

Here are my raw notes from a call with [company]: [paste notes].
1) Write a follow-up email under 120 words recapping what they said and the agreed next step.
2) Then list CRM fields: next step, next step date, deal amount, decision makers, risks.
Only include what is in my notes. Mark anything ambiguous with "?".

Play a skeptical [role] at a [company type] who already uses [current solution].
I'll pitch [product]. Push back the way a real buyer would, one objection at a time, and don't give in easily.
After 6 exchanges, tell me which of my answers were weakest and why.

I'm meeting a [role] at a [company type] that has just [trigger event].
Suggest 8 open discovery questions about how this affects [area we help with].
Order them from easiest to most probing. No questions that can be answered yes or no.

Rewrite this email to under 80 words. Keep every fact and the ask. Remove filler, adjectives and anything that sounds like marketing. Keep my voice.
[paste email]

I emailed [name] on [date] about [topic] and got no reply. New information since then: [one new, checked fact or useful resource].
Write a 3-sentence follow-up that adds this new information and doesn't guilt-trip or say "just checking in".

Based on my context block, suggest 5 customer segments I might be missing.
For each: why they might need us, the trigger that would make now a good time, and the job title most likely to own the problem.
Label all of it as hypotheses for me to test. Don't name specific companies.

Suggest 8 subject lines for this email, each under 6 words, lowercase, specific to the content, no clickbait or false urgency.
[paste email]

Worked example: from checked facts to a sent email

Illustrative example. The companies, people and facts are invented to show the method.

The seller: Fernway People (fernway.example), an HR software vendor. The prospect: Lakeshore Logistics (lakeshore.example), checked by the rep on its careers page and news page. Verified facts: opened a second warehouse in August 2026 (company news page); hiring 14 warehouse associates and its first HR generalist (careers page, checked 20 October 2026).

The rep pastes the context block, then prompt 1 with those two facts. ChatGPT's first draft:

Hi Dana, I saw Lakeshore opened a second warehouse in August and is hiring its first HR generalist. When onboarding moves from one site to two, paperwork and scheduling usually stay on spreadsheets longer than they should. Following your recent Series B, that's likely a priority. Is onboarding at the new site still handled by hand?

The check: the rep traces each claim back to a source. The warehouse and the HR hire match the inputs. "Following your recent Series B" appears nowhere in the facts: ChatGPT added it, and Lakeshore has raised no such round. The rep deletes the line, confirms Dana's title on the company's leadership page, and sends a four-sentence email. This is why the context block asks for [UNVERIFIED] tags: a model that follows the rule makes the check faster, but you still read every line.

Test whether ChatGPT-written emails work better

Don't judge by how good a draft looks. Split one list at random, send version A to half and version B to the other half, and compare positive reply rates. Mind the sample size: to detect a lift from a 3% to a 5% reply rate with reasonable confidence (5% significance, 80% power), you need about 1,500 emails per version. With 100 per version, a difference of two or three replies is noise. If your volume is lower, test bigger differences (a new angle, not a new adjective) and judge over several weeks.

The sending rules still apply

An email ChatGPT drafted is still your email. The FTC's CAN-SPAM guidance says the law "makes no exception for business-to-business email": use honest headers and subject lines, include your postal address and a clear opt-out, and honor opt-outs within 10 business days. Don't let a prompt talk you into a misleading subject line.

Checklist before you send anything ChatGPT drafted

• Every fact traces back to a source you opened, with a date.
• Names, titles and company names are confirmed outside ChatGPT.
• No promises, prices or customer results you haven't approved.
• Nothing marked [UNVERIFIED] or "unknown" is left in.
• It sounds like you, and you could explain every sentence on a call.
• The email has your postal address and a working opt-out.

Common mistakes

• Asking ChatGPT for leads or email addresses. It may produce plausible names and addresses that don't exist. Use a data source for contacts; Bebop, for example, finds matching companies and contacts and explains why each fits, then you verify.
• No context. "Write a cold email for my software" gets the same email every other seller gets.
• Pasting confidential data into a personal account. Use the workspace your company approved.
• Sending the first draft. The first draft is where invented details live.
• Judging prompts by feel. Measure replies on a split list.

Next step: if you're ready to go beyond a chat assistant, see how a dedicated AI sales assistant handles call notes and CRM updates, and how to pilot one.