How to Start Using AI in Sales: A Stage-by-Stage Guide

"The spread in AI maturity is enormous," a former software CEO told us recently. "You have the forward-leaning early adopters, and then a huge long tail of companies still stuck in 2015."

Most sales teams we meet sit somewhere in between. Someone uses ChatGPT for emails, someone else has a note-taker in their meetings, and the manager has been asked by the board what the AI plan is. When they search for where to begin, they find lists of forty tools.

This guide does it the other way around. It walks through the sales process one stage at a time and, for each stage, says what AI does reliably today, what to keep in human hands, and what to test first. It ends with a 30-day plan you can start on Monday.

The short version: start where the buyer can't see the AI, and move toward where they can.

Why the order matters

Most teams start with cold email, because that's where the demos are. It's also where buyers notice first. "There's too much written outreach, and too much of it is AI-generated and mass-produced," a head of sales at a B2B tech scale-up told us. A founder who gets pitched constantly put it more bluntly:

"A rep called me to sell an AI that finds customers. I just asked him: 'Then why are you calling me?'"

An AI mistake in your back office costs you a correction. An AI mistake in front of a buyer costs you the buyer. So build trust with the tools where errors are cheap, and move toward the buyer as you learn what they get right.

This guide follows that order:

1. After the conversation: notes, CRM updates, follow-up tasks
2. Before the conversation: account research and meeting prep
3. Building the list: who to call
4. During the call: live prompts and objection handling
5. Writing to buyers: emails and follow-ups
6. Managing and coaching: deal review and win/loss

Three rules that apply at every stage

Before the stages, three rules. If a tool breaks one of them, it doesn't matter how impressive the demo was.

1. It shows its source. Every claim the AI makes about a company, a person or a conversation should come with where it came from: the link, the sentence, the timestamp in the call. We watched a sales manager test an AI agent that listed retailers that had "recently opened stores." Within a minute he noticed the examples were from 2024 and asked where the information came from. That's the right reflex. Without a source you can't check the answer, and an answer you can't check gets either trusted blindly or redone by hand.

2. A human approves before anything reaches a buyer or changes what the team treats as fact. Drafting is cheap. Sending is not. Neither is a deal marked "decision-maker confirmed" because the AI misheard who was in the room. A sales and growth consultant summed up what most experienced sellers feel:

"I would never dare let three AI agents loose in my inbox to automate everything. I want control."

The goal isn't to remove the human. It's to make the human's part a quick yes or no instead of twenty minutes of typing.

3. It writes into one record. If the AI's output lives in a chat window, it's gone the next time someone opens the deal. AI only compounds when notes, tasks and research land where the next person will look: on the account, in the CRM, in the deal history. Otherwise you've moved your sales knowledge from people's heads into people's chat histories, which is the same problem with a new interface.

Stage 1: After the conversation (start here)

What AI does well today: transcribing calls and meetings, summarizing them, pulling out commitments (who promised what, by when), filling CRM fields, and creating follow-up tasks with the buyer's own words attached.

What to keep: your read of the deal. The AI can tell you what was said. Only you know what you think is really going on. Write that one line yourself.

What to watch: language. Test in the languages you actually sell in, with your accents, your product names and the mix of Swedish and English that happens in every Nordic tech meeting. Check how it handles names and numbers, because those are the details that end up in a follow-up email.

Why start here: the errors are visible to you, not the buyer. The payoff is immediate: a sales-process consultant told us admin "kills the buying tempo in any conversation" and that he uses AI so he can spend as much time as possible in front of customers. Above all, this stage creates the data every later stage depends on.

Your first test: for two weeks, run AI notes alongside your own on the same calls. Then apply the six-month test: would a colleague who has never met the buyer understand the deal from the AI note alone? Measure the time from end of call to follow-up sent, and the share of calls with a logged next step.

More on why this matters: admin is a tax your sales team pays twice.

Stage 2: Before the conversation

What AI does well today: turning a company's website, annual report, news, job ads and your own previous conversations into a one-page brief. Before a follow-up meeting: what has happened since last time, and which commitments are still open.

What to keep: deciding what matters, and verifying anything you'll say out loud. A country sales lead at an enterprise software company put the bottleneck well:

"AI research is fast. The slow part is verifying every contact and every phone number before a seller can pick up the phone."

What to watch: stale sources (the 2024 store openings), the wrong legal entity in a group, and confident filler when the AI found nothing. A good research tool says "nothing found." A bad one writes three plausible paragraphs.

Your first test: an AI brief before every meeting for two weeks. After each meeting, mark the brief as used, ignored or wrong. If more than one in ten is wrong, fix the sources before you scale it.

Stage 3: Building the list

What AI does well today: turning a description of your ideal customer into search filters, sorting messy job titles into actual roles (titles are made up), reading a web page or a pasted list and matching it to real companies, and finding your competitor's customers from what suppliers and customers publish.

What to keep: the definition of who you sell to, and the link to real records. Ask a chat assistant for "50 logistics companies in Sweden" and you get plausible names, some of which have closed, merged or were never called that. Use AI to interpret and rank, and anchor every company in a registry record with an organization number.

What to watch: coverage. A Nordic BDR lead told us their AI tool and data vendor found "maybe a fifth" of the CFOs, PMOs and controllers they needed in one sector. The rest was manual digging. AI can't find a person who isn't published anywhere, so ask any tool how it handles the people it can't find.

Your first test: take your last 20 won customers and have your current method and an AI-assisted method each produce 50 lookalikes. Have a rep score 20 from each list blind. Pick the method by the score, not by how fast it was.

Stage 4: During the call

What AI does well today: live prompts when the buyer raises an objection, reminders of the discovery questions you haven't asked yet, and surfacing the answer that worked for a colleague last week.

What to keep: the conversation. Prompts should be something you glance at, not something you read. A rep reading a script off the screen loses the call.

The condition: live coaching is only as good as what it's fed. Generic AI gives generic answers. Feed it your real objections and the answers your best reps actually use, and it becomes the place where your sales process stops living in three people's heads (where should your sales process live?). The battle card format is a good starting point.

What to watch: resistance from senior sellers. A sales trainer described technical sales engineers who "don't jump for joy at a CRM, and they don't jump for joy at AI either." Don't start with them. Start with new reps, where live prompts cut ramp time, and let the results do the convincing.

Your first test: write down your five most common objections and your best rep's answer to each. Run live prompts with your newest reps for two weeks and compare their meeting-booked ratio to the two weeks before.

Stage 5: Writing to buyers

What AI does well today: first drafts of follow-ups based on what was actually said, rewording, shortening and translating.

What to keep: the send button, and any claim that you personally noticed something about the buyer.

Why AI emails look like AI emails: the same structure ("I noticed that your company..."), compliments without substance, and a pain point that could apply to anyone. The fix is not a better prompt. It's better input. A follow-up built from the buyer's own words in yesterday's call reads as human because it's about something they actually said. A cold email built from a website scrape reads like every other cold email in their inbox.

What not to do: automate cold first-touch at volume. The flood of AI-written outreach is exactly why buyers in several of our recent meetings said the phone has become stronger, not weaker.

Your first test: AI-drafted follow-ups after held meetings only. The rep edits and sends every one. Track reply rate and time-to-send against the previous month.

Stage 6: Managing and coaching

What AI does well today: listening to more conversations than any manager has time for, flagging deals where the next step is missing or a competitor came up, and classifying why deals were won or lost from what was actually said (win/loss that changes behavior).

What to keep: the coaching conversation, and the call on any individual deal.

Why it comes last: it needs stages 1 to 5 to produce data. A manager dashboard on top of empty CRM fields is a dashboard of empty fields. Teams that start here buy an expensive mirror and find nothing in it.

Your first test: after a month of recorded calls, have the AI list the ten most common objections with examples. Compare that list to what the reps say in the weekly sales meeting. The gap is your coaching agenda.

A 30-day plan

Week 1: pick one stage and one team. For almost everyone that's stage 1. Write down your baseline before you change anything: time from call to follow-up, share of calls with a logged next step, and each rep's own estimate of admin hours per week.

Week 2: run it alongside the old way. Don't remove anything yet. Reps keep their own notes; the AI does its version too.

Week 3: check the output against the three rules. Could you see the source? Did a human approve before anything went out? Did it land in the record where the next person will look? When something fails, fix the setup, not the reps.

Week 4: decide, then write it down. Keep, change or drop. If you keep it, write it into the process (how a call ends, what gets checked before a follow-up goes out) so it doesn't depend on who happens to like AI. Then pick the next stage.

A CEO of an outsourced SDR consultancy gave the best advice we've heard on this:

"With AI right now you have to pick one solution, commit to it and build a structure. It's far too easy to plug in eight different tools at once that don't talk to each other."

Five mistakes we see

1. Starting with a tool instead of a stage. "We bought an AI SDR" is not a plan. "We want follow-ups out within an hour of every meeting" is.
2. Every rep with their own prompts. The best rep's prompt library is the new version of the best rep's head: valuable, invisible, and gone when they leave. We wrote about where this ends up in the "I can do everything with Claude" high.
3. Measuring time saved instead of outcomes. Hours saved are easy to claim and hard to verify. Follow-up speed, held meetings and booked ratio are not.
4. Eight tools that don't talk to each other. Each one saves ten minutes, and the copying between them costs twenty.
5. Skipping the language test. A tool that's brilliant in English can be mediocre in Finnish. Test with your own calls before you sign.

Five questions to ask any AI sales vendor

• Where does each fact come from, and can I click through to the source?
• What does it do without asking me first, and can I turn that off?
• Where is our data processed and stored, and is it used to train models?
• How does it handle the Nordic languages, names and mixed-language conversations?
• What exactly reaches our CRM: which fields, in which direction?

If a vendor can't answer all five in the first meeting, you'll find out the answers the hard way later.

Where Funnelfeedr fits

We built Funnelfeedr around the same order and the same three rules. In short, stage by stage:

• After the conversation: calls in the dialer and meetings in Google Meet, Teams and Zoom are transcribed in Swedish, Norwegian, Danish, Finnish or English, then summarized. Action items become tasks with the buyer's quote attached and sync to HubSpot or Pipedrive. A connected Gmail or Outlook inbox is read the same way.
• Before the conversation: a pre-meeting brief is prepared automatically, covering who they are, what to talk about, what to watch out for, and its sources.
• Building the list: ideal customer profiles and lookalike search on Nordic company data anchored in the business registries, search by role instead of job title, AI Import for pasted lists and logo walls, and agents that find, for example, a competitor's customers, with a one-sentence reason and the source link when it came from the web.
• During the call: live coaching in the dialer surfaces your own objections and the answers that work when the buyer raises them. The objection list learns from calls, meetings and email, and a new objection goes to approval before any rep sees it.
• Writing to buyers: AI drafts replies and follow-ups from the actual thread. A person always reviews and sends.
• Managing and coaching: deal qualification (BANT, MEDDIC, MEDDPICC or SPICED) filled in from conversations, with the evidence for each answer, plus warnings such as no next step scheduled or no decision-maker engaged.

AI suggestions that change what the team relies on (qualification, stakeholders, knowledge, objections, competitors) go to an approval queue first, and you can switch a type to auto-apply once you trust it. Follow-up tasks are created directly, because a task is a reminder for the rep, not a decision. And if your team already works in Claude or ChatGPT, the MCP connector lets them search companies and build lists in Funnelfeedr from there.

Want to see where AI fits in your own process? Book a demo and we'll walk through it stage by stage →