Your 10:30 call is with a VP of operations you've never spoken to. You have a name, a company and 12 minutes. You could open six tabs, or you could hand the raw material to a model and get a one-page brief with the questions worth asking.
That second path works, but only if you control what goes in. Most of what makes AI outreach feel like spam is one mistake: asking the model to produce a "personalized" line from nothing. It then invents one, or writes something that fits ten thousand companies. The prompts below flip it. You supply real facts, the model organizes and drafts, and you decide what gets sent.
These are untested templates. I didn't run them against a live CRM or real prospects, so treat the first few uses as drafts you'll tune. The worked example further down is labelled illustrative.
What should I never ask AI to do in sales?
Three rules cover most of the damage.
- Don't ask it to find facts from memory. Models don't know that your prospect's company raised money last month, and they'll still sound confident about it. Paste sources in. If you use a model with web search, require a URL per claim and click it.
- Don't let it write the personalized line unsupervised. The common complaint from the buyer side is the opener that could apply to any company: "I noticed [Company] is doing exciting things in [industry]." Vendors that sell outreach tools say the same thing in their own marketing, calling it "personalization theater." Use the model to propose three possible hooks, each tied to a source, then pick one yourself.
- Don't automate the send on cold first touches. Draft with AI, send with a human. A wrong first name or a made-up product feature costs more than the five minutes you saved.
The rest of the post assumes those three.
Account research prompt: sources in, brief out
Use this before you write a word to anyone. Paste raw text from the company's site, a recent press release or news item, the prospect's LinkedIn headline and about section, and any relevant job post. Job posts are underrated: they tell you what the team is hiring to fix.
You are helping a sales rep prepare to contact [PROSPECT NAME], [TITLE] at
[COMPANY]. I sell [PRODUCT/SERVICE] to [TYPICAL BUYER].
Use ONLY the source text below. Do not use outside knowledge about this company.
If something I ask for isn't in the sources, write "not in sources".
Return:
1. What the company does, in two sentences.
2. What this person likely owns or is measured on (say "inferred" if it isn't
stated).
3. Three possible outreach hooks. For each: the hook, the exact quote from the
source that supports it, and why it might matter to someone in this role.
4. Two things I should NOT assume or mention (anything uncertain or too private).
5. Three questions I should ask on a first call.
SOURCES:
[paste text, label each one with where it came from]
The quote requirement is the control. If the model can't point at a sentence, the hook is invented. Read the quotes before trusting item 3.
If you're researching many accounts, keep the prompt and rotate the sources. Don't keep one long chat open across accounts, because details from account A will leak into account B. Start fresh each time.
Cold outreach that a person would answer
A reply-worthy first email is short, has one reason for contact, and asks for something small. Give the model that structure and your actual selling point, not a request to "write a cold email."
Draft a first email from me to [PROSPECT NAME], [TITLE] at [COMPANY].
Facts you may use (nothing else):
- Hook: [ONE hook from the research step, with its source]
- What we do, in one sentence: [PLAIN DESCRIPTION, no adjectives]
- One relevant proof point: [A REAL customer result or fact I can stand behind,
or "none"]
- The ask: [15 minutes next week / a yes-no question / permission to send a
one-pager]
Constraints:
- Under 90 words, plain text, no bullet points, no formatting.
- Subject line under 6 words that says what the email is about. No tricks, no
"Re:", no fake urgency.
- No flattery, no "I hope this finds you well", no "quick question" unless it
really is one.
- Do not invent numbers, customer names, or features. If the proof point is
"none", leave it out.
- Write like a colleague, not a brochure. Contractions are fine.
Give me three versions with different openings.
Three versions help because you'll usually hate one and fix another. Then edit by hand. Read it out loud once. If you wouldn't say a sentence to the person across a table, cut it. For more on stripping out the tell-tale phrasing, see how to prompt AI to sound human.
The compliance bit you can't skip
If you're emailing from or into the US, CAN-SPAM applies, and per the FTC's compliance guide it has no exception for B2B email. The requirements include accurate sender information, a subject line that doesn't mislead, a valid physical postal address, a clear way to opt out, and honoring opt-outs within 10 business days. The guide lists penalties of up to $53,088 per violating email as of 2026-10-08; the amount is adjusted periodically, so check the page for the current figure. Hiring someone else to send for you doesn't move the responsibility off you.
Other countries have their own rules, many stricter on consent. I haven't verified those here, so ask your legal or compliance team which regimes cover your list. Put the unsubscribe line and address in your template, not in the AI prompt, so it can't get dropped or reworded.
Call prep: the one-page brief
Research tells you who they are. Call prep tells you what you'll do in the room. Paste in your research output, any previous emails with the person, and your own notes.
I have a [DISCOVERY / DEMO / RENEWAL] call with [NAME, TITLE] in [TIME].
Stage of deal: [STAGE]. What I know so far: [NOTES].
Create a one-page call brief:
1. Goal of this call in one sentence (what must be true when I hang up).
2. Opening: two sentences that reference something real from my notes.
3. Five discovery questions, ordered. Mark the one most likely to uncover
budget, timeline, or who else decides.
4. The two most likely objections for this stage and a short honest response
to each.
5. Red flags to listen for.
6. A proposed next step I can ask for before the call ends.
Use only facts from my notes. Where I'm missing information, list what to ask
instead of guessing.
The last line turns gaps into questions, which is what you wanted from the brief anyway.
A tip from the other direction: after the call, don't ask for "a summary." Paste your messy notes and ask for the facts split from your interpretation. More on that below.
Objection handling prompts that don't produce slogans
The generic AI answer to "it's too expensive" is a three-step framework about value. Real objections are more specific. Give the model the real sentence, the context and your constraints.
A prospect just said: "[EXACT QUOTE]".
Context: [STAGE, who they are, what they've said about their problem].
What's actually true about our offering: [FACTS: pricing structure, what's
included, what's not, real alternatives they might use].
1. List three possible reasons someone says this. Which is most likely given the
context?
2. Write a 2-3 sentence reply for the most likely reason, in spoken English.
3. Write one question I can ask to find out which reason it really is.
4. Tell me if the honest answer is that we're not a fit.
Item 4 matters. A tool that always finds a way to keep the deal alive will coach you into wasting your quarter on prospects who will never buy. Let it say no.
You can also run the model as the buyer for practice. The setup is the same as the mock interviewer approach: one objection at a time, no feedback until you ask, and a rubric for the scoring.
A worked example (illustrative)
This is a made-up scenario to show the shape of the process, not a transcript from a real run.
Situation: you sell a field-service scheduling tool. The prospect is a head of operations at a mid-size HVAC company. From their careers page you pasted a job post for a "dispatch coordinator" and noted their website lists a customer promise of same-day service.
Research step output (shape):
Hook 1: Hiring a dispatch coordinator.
Source quote: "Coordinate daily schedules for 30+ technicians."
Why it might matter: dispatch is probably manual and growing.
Hook 2: Same-day service promise.
Source quote: "Call before noon, we're there today."
Why it might matter: schedule changes would hit this promise.
Not assumed: how they schedule today (not in sources).
What you do with it: pick hook 1, because it points at a problem they're already paying to solve. You don't claim to know their scheduling is manual. The email says: "Saw you're hiring a dispatch coordinator for 30+ technicians. We build scheduling software for exactly that setup. Open to 15 minutes to see if it's relevant?" Then you add your address and unsubscribe line from the template.
Notice what the model did not do: guess at their software stack, invent a customer result, or flatter. That's the system working.
Follow-ups that add something new
"Just bumping this to the top of your inbox" tells the reader nothing. Each follow-up should carry one new thing: a relevant fact, a short answer to a likely question, a resource, or a clear way to say no.
Draft a follow-up to my email below. It's been [N] days with no reply.
My earlier email: [PASTE]
One new, true thing I can add: [A relevant article, a customer-agnostic insight,
a change in availability, a smaller ask].
Constraints: under 60 words, no guilt, no "just checking in", no fake urgency.
Include an easy way out ("If this isn't a priority, tell me and I'll close the
loop").
After a meeting, use a different prompt for the recap email. Paste your notes with decisions and open items, and ask for: what we agreed, what each side does next, and a date. Keep the model away from promises. If your notes don't say you'll deliver a feature by March, the email shouldn't either.
Turning call notes into CRM updates
Messy notes go in, structured updates come out. The risk is the model "tidying" your uncertainty into certainty.
Here are my raw call notes. Output:
- Facts the prospect stated (with the note line each came from)
- My interpretations (label clearly)
- Next steps with owner and date (only those I actually mentioned)
- Open questions
- Suggested CRM field updates: stage, close date, competitors, decision makers
Do not upgrade "maybe" to "will". Where the notes are unclear, write "unclear".
NOTES:
[paste]
Check the "interpretations" list against what you remember. This is where a prospect's "we might look at this next quarter" quietly becomes "Q1 close" in your forecast.
Where this goes wrong
| Failure | What it looks like | Fix |
|---|---|---|
| Invented hook | A funding round, product launch or quote that didn't happen | Require a source quote per claim; open links |
| Generic opener | Could be sent to any company in the industry | Ask for hooks tied to the role, not the company name |
| Creepy detail | References something personal or non-public | Add "Do not mention personal social media, family, or non-professional activity" |
| Tone drift | Sounds like a different person on email three | Paste two of your real emails and say "match this voice" |
| Over-promising | Follow-up commits you to dates or features | Instruct: no commitments unless I list them |
| Context bleed | Details from one account show in another | New chat per account |
What not to automate
- The decision to contact someone. Account selection and timing are judgment calls about fit.
- Anything with pricing, legal terms or delivery commitments. Draft with AI if you like, but a human checks every number.
- Sensitive replies. Complaints, escalations and cancellations deserve your own words.
- First touch at volume. Buyers can see the pattern, and the damage lands on your domain and your name.
- Customer data in tools you haven't cleared. Ask your security or IT team which tools are approved. Use placeholders if unsure.
If you want to go further on structuring prompts like these, giving context is the lesson on what to include, and the role prompting playbook covers personas. Other business roles have their own sets: small business owners has a sales section aimed at owners rather than reps, and email automation prompts covers inbox workflows. More copy-paste starting points live in the prompt library.
