Before there is a deal room, a business case, or a champion, there is a call or an email that creates just enough interest to warrant a deeper dive. That craft, is outbound.
This chapter is built on the outbound philosophy of Kyle Norton, who has built multiple outbound programs at scale, advises hundreds of companies on building them, teaches the material from the stage at Pavilion and SaaStr, and writes one of the most widely read publications in enterprise sales.
In this section, we dive into how to create an effective outbound program, dissect the cold call, and learn how to craft well-written cold emails that earn attention.
The promise of AI was that it would finally make outreach relevant. Every rep would have a research analyst and every buyer would hear from people who had studied their business.
What arrived instead was far more surface-level slop. Buyers now receive outreach that mentions their funding round, quotes their own LinkedIn post, references a competitor, and lands with the distinct feeling of having been assembled by a machine that never understood a word of it. Buyers now discount all personalization as AI slop and outbound has become an order of magnitude more congested than just a few years ago.
"Outbound isn't broken. Your prospects are just drowning in mediocre outreach."
The fix, however, is not less AI. Kyle uses more of it than almost anyone in this guide. The fix is being selective about which parts of the job AI should touch and which parts a human has to own for the output to be valuable to the person receiving it.
Start from how buyers actually react
A quick lesson on human psychology is the backbone of Kyle’s philosophy on outbound. According to psychologist Daniel Kahneman, human thinking runs in two modes:
System 1: is fast, automatic, and pattern matched: the light turns green and your foot moves before you consciously decide anything.
System 2: is slow, deliberate, and effortful. People spend it as sparingly as possible because it costs them time and energy.
An unexpected call from an unknown company gets handled entirely in System 1.
A real conversation requires System 2 attention. With this in mind, the job of every touchpoint is to get the buyer to stop and think for one moment. Treat each micro-moment as a checkpoint that you must pass through to earn the next milestone.
"Checkpoint selling is like those old racing games you played as a kid, where you drive through the gate and earn yourself fifteen more seconds. Nobody is just going to give you two minutes out of nowhere. You have to earn it."
Three things the buyer needs in order to say yes to a meeting
Evidence that someone did real work. One observation the buyer can verify, attached to a downstream impact they feel. Not stating the funding round or a job title. Rather, the problem, followed by the downstream effect that problem creates.
A trusted third party as proof. In enterprise, the fastest route to credibility is borrowing it from someone the buyer already trusts.
Value beyond your product. The buyer has to believe the meeting is worth their time on their own terms, whether or not they ever become a customer.
"People generally care about people, not companies. A specific name is much more meaningful to somebody."
Now that we have stated how the buyer is evaluating this situation, let’s dive into how to do it well.
An outbound program is three competencies: data, process, and talent.
1 Start with the data
Kyle believes that a successful outbound program begins with a high-quality list and an understanding of each unique buyer — so much so that you should take research out of the rep's job entirely.
Any job is a bundle of tasks that can be pulled apart and reassigned. In the case of outbound, that’s enriching contact details, phone numbers, and lists with a hypothesis already attached. Kyle argues this should be produced centrally, by whoever owns your GTM systems, and handed to reps ready to use.
"Nobody got into sales because they love grinding through lead databases."
Related to good data, it’s also important to sort the accounts before anyone dials. Tier targets by how likely they are to buy. In practice, expect roughly 25% to be strong, 50% plausible, and a 25% prospects who will never buy under any circumstances. Orient your team around the accounts that have a chance.
The math makes the case on its own.
Bad list
Well-built list
Dials per day
100
100
Reach a decision maker
3%
13%
Conversations
3
13
Convert to a meeting
4%
16%
Meetings booked per day
0.1
2.1
In both scenarios, we have the same rep, with the same script, and same amount of effort. The rep with good data produces roughly twenty times the output, and the only variable that changed was the quality of the information handed over before the first dial.
"No training, hiring profile, call volume, or clever technique overcomes bad data."
2 Then, hit the phones
Once we have nailed who we are going after and have given the team high quality data that helps them understand their prospect, the phone does the job of booking meetings.
The gold-standard cold-call structure is what Kyle has coined as "PICRIOS" (as he puts it, "the greek god of cold calling"), which stands for:
Pattern interrupt
Credibility
Relevance
Intrigue
Offer of value
Steps
Pattern interrupt
A buyer who answers a cold call is immediately looking for the exit. Almost every cold call opens with "how are you doing today" or "do you have thirty seconds," and system 1 response is to swat it away as quickly as possible.
There are three primary ways to break that pattern:
Silence. Say your name and your company, then stop. The buyer speaks to fill the gap, usually by asking how you are, and answering pulls them off autopilot.
A name. Ask whether they know a specific person: "do you know Tony Stark from XYZ Corp?" Asked cold, that question cannot be processed on autopilot, because the buyer has to search their memory.
A verifiable detail. Open with something specific about their business that no mass email would contain. The one Kyle still remembers named the four cities where he had teams: Cologne, Paris, Hong Kong, Australia. Specificity of that order proves a human did work before dialing, which is itself the interruption.
Credibility
Once the buyer is listening, you have a few seconds to show the call is not random. In the “name” opener we gave above, that doubles as your credibility. Source those names before you dial: list your customers, then check which of them are connected to people at your target accounts. Then open as a question.
Hey John, wanted to give you a shout because we're working with Margaret over at XYZ. Do you know Margaret?
This works even when the buyer has never heard of Margaret. A real name proves the call is not random, and the buyer usually asks who she is, which means they are talking to you.
Relevance
Now prove you understand their business. Naming something you noticed is important, but relevance is one step deeper. Then, use the evidence to build a hypothesis about the problem it creates: I noticed X, so I am guessing you are spending real time on Y. Is that fair to say?
"Every account is not bespoke. You are doing these in batches. There are four or five important problems your company solves, and then you go find all the companies who have a digital footprint that tells you they have that problem."
Intrigue
Solely understanding your buyers problem does not earn you their time. The buyer also has to believe you can do something about it.
One way to accomplish this is to name a result another relevant company got and leave out how they got it. This makes the buyer think for a second and ask “how did they do that?” or perhaps, think “if they’re doing it, I should probably know what they’re doing.”
"What we helped [company] with is they were trying to do X, and then they did Y, and their conversion rate from discovery to demo jumped forty-five percent.”
You are trying to get the buyer to say: “Explain yourself. What do you mean? That sounds too good to be true."
Offer of value
The golden rule of the offer is that whatever you offer has to be worth the buyer's time even if they never buy your product. A demo does not qualify. An artifact, a strategy, or a look at how comparable teams solved the problem does.
Steps
Agreement in principle is not a meeting on the calendar. Close with one specific next step: a day, a time, and who should be on it.
3 Take email out of the rep's hands
Perhaps one of Kyle’s most contested positions is that reps should not be hand-researching and writing their own emails. In his organizations, the research, data enrichment, writing, and sending all sit with a GTM engineer.
Kyle argues that manual personalization by each rep is the slowest, most expensive, and least consistent way to produce an email, and it takes reps away from a uniquely human action of talking to people. With a centralized email system, the rep is handed the problem hypothesis and can use all of their time to call prospects.
This does come with expectation adjustment, however. He is OK if email is no longer booking the meeting. So long as it's priming the prospect for a phone conversation with the rep, it's worthwhile.
Once email is relieved of the booking, email's job becomes familiarity and education.
How to write a good email
The cold call structure still applies to email, it just has to survive a much more crowded environment. In an inbox, you earn each checkpoint one sentence at a time. So we collapse the PICRIOS into four.
On the phone
In the inbox
Pattern interrupt
Subject line and first line
Credibility + Relevance
One observation, one hypothesis
Intrigue
One result, no explanation
Offer of value + Steps
One ask, worth their time either way
1. Pattern interrupt = the subject line + the first line
The subject line is the only part most buyers will ever read. The first line then has to be something a mass send could not contain. Nothing about a funding round, a job change, or a LinkedIn post they wrote. The specific, verifiable detail that proves a human looked at their business creates the interruption.
2. Credibility and relevance: one observation, one hypothesis
State the thing you noticed, then the downstream problem you think it creates:
I noticed X, so I am guessing Y is getting expensive. Fair?
If you have a name the buyer would recognize, it goes here and does the credibility work on its own.
3. Intrigue: a result with no explanation
Name what a comparable company got and leave out the mechanism. The buyer either wants to know how, or decides they should probably know how. The tendency is to explain how. Stick with one result, stated plainly, then stop.
4. Offer and steps: worth their time whether they buy or not
The golden rule from the call applies without modification. A demo does not qualify. The artifact, the teardown, or the look at how comparable teams solved it does. Ask for one thing, make it easy to answer with a single word, and do not ask for the calendar in email one.
What it looks like assembled:
Subject: cologne + hong kong
Message:
John, you have support teams in Cologne, Paris, Hong Kong, and Australia, and your careers page has three more reqs open in APAC.
Usually when coverage spreads that fast, the handoff between regions is where resolution time silently doubles. Guessing that is showing up in your CSAT by now.
Margaret's team at XYZ hit the same wall last year and cut cross-region escalations by 40%.
I put together a two-page teardown of how they restructured the handoff. Worth sending over?
A sequence that does that job:
Email one is a compressed version of the call structure, carrying the best research the GTM engineer has produced for that campaign.
Email two resends it with a single line noting they may have missed it. Historically the highest-reply step.
Email three asks for nothing and gives the buyer something useful from outside your company: a credible independent voice on the problem instead of your case study.
What you get in return is reps whose entire day is conversations, at a dial volume most enterprise leaders would dream of.
Relevance any competitor could have written. The temptation is strongest right after an AI research tool hands a rep something that reads fine.
"We helped Tony with maximizing operating profitability and synergies in his organization. Is operating synergies also a priority for you? It is like, dude, what are you even talking about? Yes, making profit is important for me as a VP of Operations. Duh."
Generic relevance does more damage than none at all, because it claims homework that visibly did not happen. The test: could the buyer have written your observation about themselves before the call? If so, the seller spent their twenty seconds and bought nothing.
Burning the popcorn. The hook lands, the buyer leans in, and the seller keeps talking.
"Burning the popcorn is when I get into a long feature dump, and then they ask about pricing, and then they get scared of pricing, because you talked about economics before they were fully interested."
The discipline is to stop the moment there is enough attention, make one concrete offer, and get off the phone. Two close relatives live in the same place. One is the Johnny Sales Pitch voice, delivered fast and through the nose and sounding junior, which gets sorted into the junk pile before any of the words register. The other is pressure applied to close, which makes the buyer dig in harder the more the seller leans on them.
Buying activity before building the system. Most teams hire a rep before building the information layer that makes a rep productive, then manage that rep on call volume. A rep in that position spends most of the day researching leads and reaching people who cannot buy anything, and books something like one meeting every two days. Activity targets against an unsorted list quietly turn a pipeline problem into a turnover problem, and the diagnosis almost always lands on the reps.
The strongest warnings here come from one of the most aggressive AI adopters in this guide. Kyle has taken manual personalization away from his reps entirely, built an AI tool that briefs every rep before every call, and says that if he were starting a sales organization today he would hire someone to build AI systems before hiring more salespeople. His diagnosis of why so much AI outreach is garbage has nothing to do with the models.
"Slop comes from one-shotting. If you are one-shotting an email campaign, you are done. People want it to be easy, and it is not supposed to be easy."
One-shotting means asking a model for something once and shipping whatever comes back. The premise to abandon is that AI makes the work easy. What it delivers is scale, and scale still has to be paid for with effort.
Lossiness, and why every piece can look fine while the whole thing fails
Most AI outreach gets built as a chain: one step researches the company, another infers what the company cares about, another drafts the email. Each step introduces a little error, and the errors multiply rather than average out.
Ask an AI to crawl your website, work out your value proposition, work out who your ideal customer is, identify your competitors, and then write to a senior executive, and you have stacked five guesses on top of each other. Each guess is defensible alone. The combined result is a sixty-two percent artifact landing in front of a buyer who decides what your company is worth in about four seconds.
What to do instead. Kyle's prescription is not to use less AI, it is to design the chain deliberately. Three moves, and they are cheap to apply to a workflow you have already built.
Be deliberate about chain length. Count the generative steps between your inputs and the buyer, and decide how many you are willing to run before something checks the output. Five unchecked steps is how you get slop. If a workflow needs five steps, it needs a checkpoint inside it.
Make a step deterministic wherever the answer is already knowable. Every inference is a place error enters, so stop asking the model to infer things you can simply tell it. Your value propositions, your ICP definition, your competitor set, and your positioning are facts you own. Fix them as rules and inputs, and the model's job shrinks to the part that genuinely needs generation.
Put the human at the decision points, not at the end. A person approving finished emails is catching errors four steps too late. The human belongs where judgment changes the output: which problems are real, which hypotheses survive, which accounts qualify.
Then there is the part most teams quit before reaching, and Kyle is blunt that it is the actual work.
Nobody is going to hand you a quality bar for your own outbound. You define what good looks like, run outputs against it, and keep grinding on the prompts, the context, and the chain until the output is genuinely good. Most "AI does not work" conclusions are really "I built an MVP, tried it twice, and stopped" conclusions. In his telling, people quit around hour three, and the breakthroughs tend to land around hour six or eight.
Stop letting the machine decide what you should be deciding
The operating principle is creative constraints, another idea Kyle credits to Jordan Crawford: the less open space you leave a model, the less room it has to invent, and the way to leave less space is to supply the answers yourself at the points that matter.
Two versions of the same instruction show the difference.
The version that produces slop asks an AI to review your value propositions and go find companies that are a good fit. That leaves acres of open ground, and the model will wander into it.
The version that works asks the AI to read your internal documentation, your recorded customer calls, and your customer testimonials, and return its best five to ten hypotheses about which problems your company actually solves well. A human reads that list and throws out the ones that are obvious or wrong. Only the survivors pass to the next step.
Three pieces of machinery hold the constraints in place, and most teams skip all three.
Context engineering. Write out the personas, value propositions, beliefs, and positioning by hand and feed them to the AI, instead of letting it infer them from your website.
Evals. Borrow the testing practice from software engineering: define in writing what a good output and a bad output look like, then check outputs against that standard automatically. Include a dedicated step whose only job is to inspect the research another step produced.
A feedback loop. Open the spreadsheet of outputs, write down why each one was good or bad, and feed that judgment back in so the system improves.
