ChatGPT Prompts for B2B Sales: 12 That Actually Work in 2026

Most ChatGPT prompts for B2B sales circulating in 2026 don’t work in production because they ask the model to do creative work without giving it the constraints that keep its output useful. “Write me a cold email” produces generic output. “Write me a cold email to a SaaS founder” produces slightly less generic output. The prompts that actually work in real sales operations look very different — they’re highly constrained, fact-grounded, and frequently include the model’s role explicitly stated. We’ve shipped about 60 such prompts into client campaigns at AFF Lab over 2024-2025; this article walks through 12 of the most useful ones across four common sales jobs, with the prompt-engineering rules that make each one reliable.

The category split below mirrors how AI fits into real B2B sales workflows — see our pillar on AI in B2B sales for the broader stack. Four jobs where ChatGPT (or any frontier LLM) actually earns its keep: research, personalization, reply triage, follow-up. The prompts below assume GPT-4-class capability; for older or smaller models you’ll need more verbose constraints to get equivalent output.

Working ChatGPT prompts for B2B sales share four properties: a clear role assignment (system prompt or explicit “you are…”), constrained context (only use facts I give you, no extrapolation), a defined output format, and an example of the desired output. Prompts missing any of these four produce generic or hallucinated content.

We won’t waste space repeating “ChatGPT can help your sales team” filler. Below are the specific prompts and the rules behind them.

Prompt category 1: prospect research

Use case: take a name, company, and LinkedIn URL — output structured intelligence the SDR can use in personalization.

Prompt A — One-paragraph prospect summary.

You are a B2B sales analyst. Given the LinkedIn profile and company information
below, produce a 3-sentence summary covering: (1) the prospect's role and what
they likely care about professionally, (2) one specific recent event or post
from the data, (3) one plausible business reason they would care about our
offering, which is [YOUR OFFER].

Use ONLY the facts in the data below. Do not extrapolate beyond what's stated.
If a fact isn't available, write "unknown" — do not invent.

DATA:
[paste LinkedIn bio, recent posts, company description]

Why it works: explicit role, explicit constraint against extrapolation, defined output length, instruction to use “unknown” rather than guess. The “unknown” instruction is what stops hallucinations — without it, LLMs default to confident invention.

Prompt B — Industry-specific ICP fit check.

You are evaluating B2B prospect fit. Our ICP is: [ICP description in 2-3 sentences].

Given the company below, score 1-10 their fit with this ICP. Then in one
sentence each, list (a) the strongest reason they fit and (b) the strongest
reason they don't. Only use information present in the data.

COMPANY DATA:
[website snippets, employee count, industry tags, tech stack if known]

Why it works: forces both positive AND negative reasoning, which catches false-positives that simple yes/no prompts miss.

Prompt category 2: personalization

Use case: write the personalized opening of a cold email.

Prompt C — Constrained opener.

Write the opening paragraph of a cold email to [NAME] at [COMPANY].

REQUIREMENTS:
- Reference one specific fact from the data below (cite which one)
- Maximum 2 sentences
- No flattery, no "I noticed you", no "given your work in"
- Conversational, not formal

FACTS YOU CAN USE:
[3-5 specific facts: recent posts, role context, company news, etc.]

OUR OFFERING (one sentence): [one-sentence offer]

Why it works: banned-phrase list at the prompt level prevents the LLM’s most-overused conventions. Forcing it to cite which fact it referenced reveals when the LLM is being lazy.

Prompt D — Subject line generator with constraint.

Generate 5 cold email subject lines for an email to [ROLE] at [COMPANY TYPE].
Topic: [your offering, 1 sentence].

CONSTRAINTS:
- 4-7 words each
- No "Re:", no "Fwd:", no questions
- Specific not vague (mention concrete role/company/industry where possible)
- Vary structure between the 5 — don't all start with the same word

Output as numbered list, nothing else.

Why it works: variation constraint stops the model from producing 5 near-identical lines.

Prompt category 3: reply triage

Use case: classify cold email replies so SDRs only see the ones that matter.

Prompt E — Reply classifier.

Classify the following cold email reply into exactly ONE of these categories:

1. POSITIVE_INTEREST — asks for more info, requests a call, expresses curiosity
2. SOFT_DECLINE — polite no, "not right now", "maybe next quarter"
3. HARD_DECLINE — explicit no, asks to be removed, angry
4. OUT_OF_OFFICE — automated absence reply
5. WRONG_PERSON — says they're not the right contact, redirects you
6. BOUNCE_OR_SPAM_TRAP — automated rejection language
7. UNCLEAR — ambiguous, needs human review

Reply: [paste reply text]

Output: ONLY the category label, nothing else.

Why it works: explicit category definitions, single-token output, “UNCLEAR” escape valve so the model doesn’t force-fit ambiguous cases.

Prompt F — Reply sentiment + next-action.

For the following B2B cold email reply, output:
1. Sentiment: positive / neutral / negative
2. Next action: REPLY_NOW / SCHEDULE_FOLLOWUP / REMOVE_FROM_LIST / ESCALATE_HUMAN
3. One-sentence reason

Reply: [text]

Format output as JSON with keys: sentiment, next_action, reason.

Why it works: structured output makes integration with downstream automation trivial.

Prompt category 4: follow-up suggestions

Use case: when an SDR needs to write a follow-up, give them 3 options to pick from.

Prompt G — Three follow-up variants.

The prospect [NAME] at [COMPANY] opened our cold email but didn't reply.
Original email subject: [subject]
Original email opener: [opener]
Time since send: 5 business days

Write 3 follow-up email variants. Each:
- Maximum 4 sentences total
- Different angle (Variant 1: add value; Variant 2: reference specific fact;
  Variant 3: explicit pivot to "is this not the right time?")
- No "just bumping this up", no "checking in", no "circling back"

Output: numbered list, each variant clearly separated.

Why it works: three deliberately different angles force the model to think about which approach fits the specific prospect, rather than producing three near-identical messages.

The five rules behind every working prompt

The specific prompts above are useful, but they age. Frontier models change every quarter, what was best practice in early 2025 isn’t quite right by late 2026. The rules behind the prompts last longer:

1. Assign a role explicitly. “You are a B2B sales analyst” or “you are a copywriter for cold email” puts the model in a constrained mental space that produces more useful output than no role assignment. This isn’t a placebo — measurable difference in output quality.

2. Constrain to verified facts. Always include “use only the data below, do not extrapolate” or similar. LLMs default to invention; only explicit constraint stops it.

3. Define output format precisely. “3 sentences”, “JSON with keys X, Y, Z”, “numbered list, nothing else.” Vague output requests produce vague output.

4. Include an “unknown” escape valve. When asking the model to extract or classify, always give it a way to say “I don’t have enough information.” Without that valve, it invents.

5. List banned phrases explicitly when copy-quality matters. “Don’t use: I noticed, given your work in, just wanted to” etc. LLMs default to certain conventions; banning them at the prompt level is the only thing that consistently stops them.

These five rules cover roughly 80% of the difference between prompts that produce useful B2B sales output and prompts that produce slop. The remaining 20% is testing — every prompt should be run on 10–20 real prospects before you trust it in production. What looks good in a single sample often falls apart on the eleventh case.

If you’ve gotten this far, the natural follow-up is the operational playbook for running AI cold outreach in production — that’s where these prompts plug into the actual workflow.