AI for sales works best on the stuff around the conversation, not the conversation itself. It speeds up:
- account research and call prep
- first drafts of outreach and follow-ups
- turning calls into CRM notes
- proposal drafts
- pipeline reviews
It doesn’t replace discovery, trust, or negotiation. And the rep reads everything before it goes out, no exceptions. Keep customer data in tools your company has approved, and never let AI invent claims, case studies, or numbers. Below, we go through a deal stage by stage: what to hand over, what to keep, and the prompts that work.
Key terms used in this guide: hallucination and prompt injection.
The Sales Cycle at a Glance
Here’s the whole picture in one table. Everything after it is a closer look at each row.
| Stage | What AI does | What the rep still owns | Tool category | Main risk |
|---|---|---|---|---|
| Research & account planning | Summarizes public info, spots angles | Deciding who matters and why | General AI assistant | Outdated or wrong facts |
| Outreach | Drafts and tweaks messages | Voice, relevance, the send button | Assistant, CRM-built-in AI | Generic, spammy emails; breaking outreach rules |
| Call prep & discovery | Suggests questions and hypotheses | Asking, listening, following the thread | General AI assistant | Scripted calls |
| After the call | Summarizes, drafts CRM notes and follow-ups | Checking accuracy, commitments | Meeting recorder, CRM AI | Misheard details, consent |
| Proposals & objections | Drafts text, plays the skeptical buyer | Pricing, terms, trust | Assistant, roleplay | Invented claims |
| Pipeline & forecasting | Flags stalled or risky deals | Judgment and the forecast itself | CRM AI, assistant | Confident-sounding nonsense |
A quick note on tools. Most reps end up with three kinds: a general assistant (ChatGPT or Claude, say), whatever AI is already inside their CRM (Salesforce and HubSpot both have some), and a call recorder or conversation-intelligence tool (Gong or Otter, for example). Your company decides which are allowed. If you live in Microsoft’s world, we have a separate page on Copilot for sales, and another on Claude for sales. Here we stay tool-neutral.
Research and Account Planning
Research is where most reps feel the quickest win, and using AI in sales research is the safest place to start, because the worst case is a mediocre draft you throw away. You used to spend 40 minutes reading a company’s site, press releases and LinkedIn pages before a first call. Now an assistant can pull the key threads together in a couple of minutes.
The catch: it can’t know what happened last Tuesday unless it can browse, and even then it gets things wrong. So treat the output like a smart intern’s first draft. Useful, but you verify the facts that matter before you say them out loud.
Where it really helps is angles. Ask it why a company might care about your product right now, based on what’s public: a new region, a hiring spree, a product launch. You still decide which angle is real.
I sell [product] to [role] at [type of company]. Here is the public info I’ve gathered about [Company X]: [paste notes, website text, news]. Write a one-page brief: what they do, three possible reasons they’d care about [problem we solve], and five questions I should find answers to. Mark anything you’re guessing as “guess.Check before sending: Verify every fact against the source. Delete anything marked “guess” that you can’t confirm.
Outreach: Drafting and Personalizing Without Sounding Like a Robot
Everyone can spot an AI-written cold email now. “I hope this message finds you well” followed by a paragraph of flattery about your “impressive growth” is a delete.
The fix isn’t a better AI. It’s better input. If you give it nothing real, it makes up something vague. If you give it a specific note (“she posted about moving the support team to a new ticketing system”), the draft gets sharp fast. This is the heart of how to use AI in sales without wrecking your reputation: you supply the human detail, AI supplies the first draft.
A few habits that help:
Keep it short. Ask for under 90 words. Ban phrases you’d never say out loud. Paste two of your own best emails and tell it to match the tone. Then edit anyway. Change one sentence into something only you would write. That’s what makes it yours. For more on tone and structure, see our guide to writing professional emails with AI.
Write a first-touch email under 90 words to [name, role]. Real context I know: [one or two specific, true observations]. What we offer, in one plain sentence: [sentence]. One clear, low-pressure ask. No flattery, no buzzwords, no claims about results. Match the tone of these two emails I wrote: [paste].Check before sending: Is every detail true? Would you say this out loud? Does it include an unsubscribe option and your real identity where the law requires it?
Know the Outreach Rules
AI makes it easy to send a lot of messages, and that’s exactly how reps get into trouble. For commercial email in the US, the FTC publishes a plain-English CAN-SPAM compliance guide. Calls and texts fall under the TCPA, which the FCC enforces. If you’re writing to people in the EU, GDPR applies. I’m not a lawyer and this isn’t legal advice: read the official sources, and ask your company’s legal or compliance team what applies to your outreach.
Call Prep and Discovery: AI Prepares the Questions, the Rep Asks Them
Here’s a thing nobody tells new reps. A list of great questions can make you a worse listener. You stop hearing the answer because you’re waiting to ask number four.
So use AI to widen your thinking before the call, then close the doc. Ask for questions grouped by theme: how they work today, what hurts, who else is involved, what happens if they do nothing. Ask for likely objections too. Pick three or four questions you actually care about. Write them on a sticky note.
Then the conversation is yours. Discovery is where trust gets built, and no tool can fake it for you.
I’m about to run a first discovery call with [role] at [type of company]. They’re likely dealing with [problem]. Give me 12 open-ended questions grouped under: current process, pain and cost, decision makers, timing. Add three follow-up questions for when they give a vague answer. Don’t suggest any question that assumes facts I haven’t given you.Check before sending: Cut it to the four questions you’d actually ask. Make sure none of them sounds like an interrogation.
After the Call: Notes, CRM Updates and Follow-Ups
This is the stage reps usually love the most, because nobody got into sales to type up notes. A recorder plus an AI summary can turn a 30-minute call into a tidy CRM entry and a draft follow-up in minutes.
But recording isn’t free of rules. In the US, consent laws differ by state. Some require everyone on the call to agree. So don’t guess: check your company’s policy, tell people at the start that you’re recording, and get a yes. If you want the basics of the notes side, we cover them in our guide on using AI to take meeting notes.
One more thing. AI summaries get names, numbers and dates wrong more often than you’d like. If the note says “budget approved,” and the buyer actually said “budget might be approved,” that error will follow the deal for months. Read it. Fix it.
Here is the transcript of my call (approved tool, recording consented): [paste]. Write a CRM note with: who attended, the buyer’s stated problem in their words, what they said about timing and budget, next steps with owners and dates, and open questions. Only include what was actually said. If something is unclear, write “unclear.Check before sending: Compare dates, numbers and commitments to your own memory. Remove anything the buyer didn’t say.
Proposals and Objection Handling, Including Roleplay Practice
Proposals are mostly assembly work: summarize their problem, match it to your offer, spell out next steps. AI is good at that first pass, especially if you feed it your discovery notes and your approved product language.
Where it goes off the rails is claims. Ask it for “a case study that proves ROI” and it will happily write one. A fake one. In 2024, a Canadian tribunal held Air Canada responsible for wrong information its website chatbot gave a customer, and the airline’s argument that the bot was a separate entity didn’t fly. You can’t blame the tool. If it’s in your proposal, it’s yours.
Practicing Objections With AI
This might be the most underrated use. Objection practice usually needs a colleague with free time. AI is always available at 8 p.m. the night before a big call. Ask it to play a skeptical procurement manager, tell it to push back hard and stay in character, then ask for feedback afterward.
It won’t feel like a real buyer. It’s too polite, and it caves too easily. But it’s good for rehearsing your first thirty seconds of an answer, which is where most reps stumble.
Play a skeptical [role] at a [type of company] who thinks our price is too high and is happy with their current vendor. Stay in character. Push back after each of my answers, and don’t make it easy. After five rounds, step out of character and tell me which answers were weak and why.Check before sending: Don’t use anything from the roleplay verbatim. Check any facts about your product against official materials.
Pipeline Review and Forecasting: A Second Pair of Eyes, Not the Forecaster
Paste an exported pipeline into an assistant and ask what looks off. It’s surprisingly good at spotting boring-but-important patterns: deals sitting in one stage for weeks, no next step logged, a single contact on a big account, close dates that keep sliding.
What it shouldn’t do is predict whether the deal closes. It doesn’t know your buyer got a new boss last week. You do. Use AI to ask better questions in the pipeline meeting, and keep the forecast as a human call.
Strip out names and anything sensitive first, or use a tool your company has approved for this data.
Here is an export of my open deals (approved tool, anonymized): [paste]. Flag deals that look at risk based only on this data: stalled stages, missing next steps, single-threaded contacts, repeated close-date changes. For each flag, give the reason and one question I should ask the buyer. Don’t predict win probability.Check before sending: Check each flag against what you know about the deal. Don’t share the list as a forecast.
The Five AI Skills That Matter for Reps
The tools will change every few months. The habits underneath them won’t. If you want to build AI skills for sales professionals, these five are worth the time. Everything about AI for sales reps comes back to the first one, and the rest follow from it.
Briefing. Good prompting is just good briefing: who the audience is, what you want, what to avoid, an example to match. Reps who write clear briefs for AI usually write clear emails too. Our guide on writing better AI prompts goes deeper.
Research judgment. Knowing which facts to verify, which sources to trust and when an “insight” is just a guess in a nice font.
Reading data. You don’t need to be an analyst. You need to look at a pipeline export, ask the right question and notice when the answer doesn’t fit what you see in real deals.
Oversight. Reading everything. Catching the invented number. Knowing when a draft is technically fine but wrong for this buyer.
Knowing what to keep. Discovery, empathy, negotiation, saying “I don’t know, let me find out.” The more of the busywork you hand off, the more your human skills are the whole job.
If you’d rather get structured practice than learn by trial and error, take a look at our overview of AI sales training.
Guardrails: The Box You Pin Above Your Desk
Most AI mistakes in sales aren’t dramatic. They’re small, fast and hard to undo. In 2023, Samsung reportedly restricted staff use of generative AI after engineers pasted internal code into a public chatbot. Nobody meant harm. Pasting is easy. Keep these in view:
- Approved tools only. If IT hasn’t cleared it, don’t use it for work.
- No customer data in consumer AI tools. Check each tool’s data and training settings. See our guide on whether AI tools are safe to use at work.
- No unreviewed sends. A human reads every message before it goes out.
- No invented claims, case studies, quotes or numbers.
- Follow outreach laws: CAN-SPAM for email, TCPA for calls and texts, GDPR for EU prospects. When unsure, ask compliance. The European Commission’s data protection pages are a good starting point for GDPR.
- Get recording consent, and follow your company’s policy.
A 30-Day Adoption Plan, One Stage per Week
Don’t try to change everything at once. Pick one stage a week, and measure it.
Week 1: Research and Call Prep
Use Prompts 1 and 3 on your next five calls. Time yourself before and after. Be honest: if the brief still needs 20 minutes of fact-checking, note that too.
Week 2: Outreach
Use Prompt 2 on a small batch, say 30 emails, and keep another 30 written the way you always do. Compare reply rates. Thirty is a small sample, so don’t declare victory over a few extra replies.
Week 3: After the Call
Try Prompt 4 on every call where recording is allowed. Track how long notes take now, and how many errors you catch during review.
Week 4: Proposals, Objections and Pipeline
Run Prompt 5 twice before big calls, and Prompt 6 before your next pipeline meeting. Ask your manager whether the questions it raised were useful.
Measuring It Honestly
Three numbers matter: time saved (stopwatch, not vibes), reply rates (against a control group of your own emails), and quality checks (how many errors you caught before sending). That last one is the sneaky one. If you’re catching nothing, either the tool is great or you’re not reading closely enough. Be suspicious of yourself.
Don’t measure by quota or revenue in the first month. Too many other things move those numbers.
Ready to Use AI in Your Deals?
The reps who win with AI are the ones who can brief it well and check its work. If you want to build that habit in a structured way, Coursiv’s 28-Day AI Certificate Program gives you daily, practical lessons you can use at work right away. You can also read more about the AI Certificate Program. Start with one stage, one prompt, one week. That’s enough to see where it helps, and where it doesn’t.