The Human-in-the-Loop Principle

I want to tell you about a farmer.

Daniel Priestley, an entrepreneur I’ve learned a lot from, uses a farming analogy that I think about constantly.¹ He describes it like this: in the agricultural age, the farmer had to know when to plant the crops. Then the farmer had to go out, toil the soil, put the seed in the ground. But after that, the soil did most of the value creation. At the end, the farmer had to know when to harvest, and then turn the harvest into something and take it to market.

The farmer did steps one and two, the soil did the middle work, and then the farmer did the final steps.

AI is the same.

It’s very good at doing the middle, but not good at knowing what to do in the first instance, or knowing when to stop, or how to take it to market. The human’s job is to do the beginning and the end: know what to plant, prepare the soil, then harvest and bring the output to the world. The system handles everything in between.

Every workflow in this book follows that pattern. In fact, every system I’ve built in the last three years follows that pattern. And every time I’ve seen a team struggle with AI, it’s because they tried to hand the beginning or the end to the machine and keep the middle for themselves.

That’s backwards, and it’s why this chapter exists.

§What Humans Do Better

There are four things that humans are consistently better at than AI, and I don’t see that changing in 2026 or 2027. Maybe not ever, but I’m not interested in forever predictions. I’m interested in what’s true right now.

1. Knowing what matters

This is strategy.

• Which accounts should we target?
• What does our ICP actually care about?
• What’s the one message that will resonate with this buyer at this stage?
• Which of the fifteen possible blog topics will actually drive pipeline versus just generate traffic?

AI can generate options and analyze data, but it can’t tell you which option is right for your specific business, your specific market, your specific moment. That requires judgment built from experience, and experience is something AI reads about but doesn’t have.

2. Knowing when to stop

This is editorial judgment. An AI workflow can produce five articles a day, but…

• Which three should actually get published?
• Which two need another pass?
• Which paragraphs are technically correct but tonally wrong?
• Where does the AI default to generic corporate language when the piece needs to sound like a specific human being?

AI has no internal sense of “this is good enough” or “this isn’t right yet.” It will produce output confidently regardless of whether that output is excellent or mediocre. The human in the loop is the quality filter.

3. Knowing who to talk to

This is relationships.

• Deciding which customer to interview for a case study.
• Knowing that the VP of Engineering at Account X is the real decision-maker, not the CTO listed on LinkedIn.
• Sensing that a prospect is interested but nervous and needs a specific kind of reassurance.

AI can research people and pull data about them, but it can’t read a room, build trust, or exercise the kind of social intelligence that turns a cold lead into a warm conversation.

4. Saying things only you can say

This is lived experience. AI can generate any content: blog posts, emails, landing pages, social captions, case studies, in any voice, on any topic, at any length. What it cannot do is tell you…

• What it was like to walk out of a conference room where the whole audience disagreed with you and feel certain that you were right.
• What it felt like to kill 140,000 monthly visits and watch the pipeline number climb.
• It can’t describe the specific moment when you realized that workflows were fundamentally different from chat, and that distinction changed your career.

Those stories, those specific details from real experience, are the only defensible moat left in a world of infinite content. And they’re the one thing that makes your marketing feel like it came from a person, not a prompt.

Priestley said something in that same conversation that I keep coming back to: “Find something that only you can say.” He was talking about personal brand, but the principle applies to every piece of content your company produces. AI can write the structure, the research summary, the supporting arguments. But the thing that makes a reader stop scrolling, the specific story from the specific person who actually lived through it, and that has to come from a human.

§What AI Does Better

It would be dishonest to downplay what AI actually brings to the table, because the gains are real.

1. First drafts at scale

A human writer staring at a blank page might produce one blog post in a day. An AI workflow can produce five structured drafts in less time, each informed by competitive data, ICP profiles, and search intent analysis. The drafts aren’t finished; they need human review. But starting from a structured 80% draft is a fundamentally different starting point than a blinking cursor.

2. Pattern recognition across large datasets

When I built the content engine at Copy.ai, one of the most valuable things the system did was identify recurring themes across sales call transcripts. A human reviewing 50 calls would catch the obvious patterns. The system caught subtle ones: specific phrases that prospects used repeatedly, objections that appeared in a particular industry segment but not others, feature requests that clustered around a use case we hadn’t prioritized. Humans are great at pattern recognition within small datasets, but AI is great at pattern recognition across large ones.

3. Consistency at volume

If you need 200 personalized outreach emails that all follow the same structural template but each reference something specific about the target account, AI does that faster and more consistently than any human team. And the consistency isn’t just about speed but also reliability. A human writing the fiftieth email of the day will inevitably phone it in. The system doesn’t get tired, doesn’t lose focus, doesn’t skip the research step because it’s running behind schedule.

4. Speed on structured tasks

Enriching a lead record with firmographic data, generating a meeting prep brief from a CRM record, reformatting a podcast transcript into five different content formats, tagging content by persona and buying stage. These are structured tasks with clear inputs and defined outputs (assembly work, not creative work). AI does them in minutes instead of hours.

5. Connecting structured data to personalized outputs

This is the one most companies underestimate: AI’s real power is in taking structured data (your value prop library, your customer language database, your account research) and combining it into outputs that are relevant to a specific person at a specific moment. That connection between data and output, at scale, is something humans simply can’t do manually.

§The Four-Stage Human-in-the-Loop Content Model

Every content workflow in this book follows a four-stage model. Once you see it, you’ll recognize it in every chapter.

Stage 1: Strategy (Human): The human decides what to create, who it’s for, what outcome it should drive, and what quality standard it needs to meet. This is the “knowing what to plant” stage: topic selection, ICP targeting, competitive positioning, and defining what “good” looks like. No AI involvement here, or at least much less (using Chat to help ideate strategy, for example).

Stage 2: First Draft (AI): The system generates the first draft based on the strategic inputs from Stage 1. This might be a blog post, a sales email, a case study outline, a landing page, or a set of social posts. The draft is informed by structured data: customer language, competitive analysis, value propositions, brand voice guidelines. It’s a targeted production step that builds on the foundation of the human set.

Stage 3: Editing and Review (Human): The human reviews the AI-generated draft. This is the editorial layer, the “knowing when to stop” stage. Is the tone right? Does it sound like our brand or like generic AI output? Are the claims accurate? Does the structure serve the reader? Is there anything that would embarrass us if a customer read it? The human edits, refines, adds personal stories or specific details that only they can provide, and approves the final version.

Stage 4: Publishing and Distribution (System with Human Oversight): Publishing (pushing to the CMS, scheduling social posts, adding to the content library, tagging by persona and buying stage) and distribution (triggering email workflows, surfacing content to sales reps, feeding into ABM sequences) are handled by the system. The human doesn’t do this manually; the system does it based on the rules the human defined. But the human maintains oversight: monitoring performance, adjusting distribution rules, and intervening when something isn’t working.

Human, AI, Human, System. That’s the loop.

The stages don’t have equal time allocations. In a mature system, Stage 1 might take 20 minutes (selecting topics and defining the brief). Stage 2 takes 5 minutes (the workflow runs). Stage 3 takes 30 to 45 minutes (the editorial pass). Stage 4 takes almost no active time because it’s automated.

Compare that to the old way: a human doing all four stages manually.

• Research (2 hours)
• Writing (3 hours)
• Editing (1 hour)
• Formatting and publishing (1 hour).

Seven hours for one piece of content (and that’s in a perfect world, on a good day).

With the four-stage model, you’re at roughly one hour per piece, and the bottleneck has shifted from production to judgment. That shift is what lets one person produce five articles a day instead of one.

§The Model Extends to Everything

The four-stage model applies to every go-to-market function in this book, not just content. Here’s how the same pattern works across different contexts.

1. Sales outbound: Stage 1 (Human): Define the target account list, value prop library, and personalization parameters. Stage 2 (AI): Workflow researches each account and generates personalized outreach. Stage 3 (Human): Rep reviews each email for accuracy, tone, and the personal touch. Stage 4 (System): Sequence management, follow-up timing, response tracking.

2. Inbound processing: Stage 1 (Human): Define lead scoring criteria, response templates, and routing rules. Stage 2 (AI): Workflow enriches the lead, scores intent, generates a personalized response. Stage 3 (Human): Rep reviews for high-priority leads; lower-priority responses go out automatically within defined parameters. Stage 4 (System): CRM updates, meeting scheduling, rep notifications.

3. Podcast repurposing: The human records the conversation (Stage 1; this is the irreplaceable part, the lived experience, the relationship). The workflow processes the transcript and generates ten assets (Stage 2). The human reviews each asset for accuracy, brand voice, and quality (Stage 3). The system distributes across channels, adds to the content library, and surfaces to sales (Stage 4).

4. ABM campaigns: Stage 1 (Human): Select target accounts, define the campaign thesis, choose the angle. Stage 2 (AI): Workflow researches accounts, generates personalized landing pages and outreach. Stage 3 (Human): Review personalization for accuracy (wrong data = embarrassment). Stage 4 (System): Launch sequences, track engagement, trigger follow-ups.

Every function follows the identical pattern: Human sets the strategy, the system does the production, the human reviews the output, and the system handles the distribution.

The middle gets automated, but the beginning and the end stay human.

§Good In, Good Out

There’s a principle that governs the quality of every system in this book, and I want to state it plainly because it’s the most common point of failure I see:

The quality of your AI output is bounded by the quality of your inputs.

You’ve likely heard the old programming maxim “garbage in, garbage out.”

If you ask AI to generate a case study and you give it a rambling, unstructured customer interview with no clear narrative arc, you’ll get a mediocre case study. Give it a well-conducted interview with clear questions, specific metrics, and a defined story structure, and you’ll get a draft that’s 80% of the way to publishable.

The same applies to sales outreach. When your value prop library consists of three vague bullet points copied from your website’s hero section, the outreach will be generic and unconvincing. When it’s structured by persona, pain point, and use case, with specific proof points and customer language attached to each, the outreach will be specific and credible.

Brand voice is the clearest example. Ask AI to produce content in your brand voice without ever defining what your brand voice is, and you’ll get content that sounds like every other AI-generated blog post on the internet. Provide the system with annotated examples of what “good” looks like and what “bad” looks like (this is exactly what my tone samples document does for Systems-Led Growth), and the output will be recognizably yours.

The human-in-the-loop starts before AI ever runs. The human’s most important job is defining the inputs, not editing the output.

Stage 1, the strategy stage, is where the quality of the entire system gets determined. The topics you choose, the ICP you define, the value props you write, the brand voice examples you provide, the quality templates you create for the AI to pattern-match against: all of that is input. And the system will only ever be as good as those inputs.

I learned this the hard way. Early on at Copy.ai, I built workflows that produced technically competent content that sounded like nobody. Well-structured, grammatically correct, SEO-optimized, and completely forgettable. The problem wasn’t the AI, no matter how much I wanted it to be. The problem was that I hadn’t given it enough signal about what “good” looked like for our specific brand, our specific audience, our specific voice.

When I started providing the system with real examples (here’s a blog post that performed well, here’s the tone we’re going for, here’s language our customers actually use, here’s a common mistake to avoid), the output quality jumped dramatically. The system didn’t get smarter overnight, but I had finally started giving it better inputs.

Good in, good out. It’s the simplest principle in this book and the one most people say out loud while skipping it entirely.

§The Bottleneck Shifts But Doesn’t Disappear

I want to be direct about something that’s easy to miss in all the enthusiasm about AI-augmented workflows: the human-in-the-loop requirement has a cost.

When I ran a content engine producing five articles per day, I still reviewed every piece, and that review was the bottleneck. Some days the drafts were 90% there and the review took ten minutes each; other days they were 60% there and I spent 45 minutes rewriting sections. The production bottleneck was gone, but the editorial bottleneck was very much present.

Sales outreach workflows showed the same pattern: the generation was fast, but the rep review was the bottleneck. Some reps loved the drafts and barely changed them. Others rewrote every email from scratch, which defeats the purpose of the system (and frankly tells you something about whether the inputs need work, not whether the system is broken). Podcast repurposing was similar; the transcript-to-assets conversion was nearly instant, but reviewing ten assets for accuracy, tone, and brand voice still took an hour.

In every case, the bottleneck moved from production to judgment, and that’s a better bottleneck to have. I’d rather have too much material to review than too little to publish. But it’s still a bottleneck, and if you pretend it doesn’t exist, you’ll burn out just as fast as you did doing everything manually. The exhaustion just comes from a different place.

For a skeleton-crew operator, this means being ruthless about where your editorial attention goes. Not everything needs the same level of review:

• A high-stakes sales proposal to a six-figure account gets a thorough pass.
• A mid-funnel blog post gets a careful but faster review.
• A social post gets a glance.

Triaging your review time is a skill, and it’s one that only becomes important when your system produces enough output to require triage.

Again, that’s a good problem to have, but it’s still a problem, and ignoring it is how people go from “AI is incredible” to “I’m drowning in AI output” in about three months.

§What Human-in-the-Loop Doesn’t Mean

A few things this principle is not, because I’ve seen every one of these misinterpretations in the wild:

It’s not an argument against AI.

I’ve seen people use “human-in-the-loop” as a way to slow-walk AI adoption. “We can’t let the machine do anything unsupervised!” That’s not what I’m saying. The whole point of the system is that the machine does most of the work and the human does the parts that require judgment. If you’re spending more time in the loop than you were spending doing the task manually, something is broken in your workflow design, not in the principle.

It’s not an argument for perfection.

Not every piece of content needs to be a masterpiece, and not every email needs to read like it was hand-crafted by a copywriter. The human review pass is about catching errors, ensuring brand consistency, and adding the specific details that make things feel personal. The goal is good enough to publish, not flawless.

The loop will get wider over time.

As your system matures, as your input quality improves, as your quality templates get more refined, you’ll find that more outputs are publishable with less review. A system that’s been running for six months produces better first drafts than one that launched yesterday, because the inputs have been tuned.

The human-in-the-loop stays, but the time spent in the loop decreases.

And it’s not an excuse to avoid building the system.

“I can’t automate this because it needs a human touch” is usually a sign that someone hasn’t thought carefully enough about which parts need the human touch and which parts don’t. Almost everything has a decomposable structure: parts that require judgment and parts that are assembly. Automate the assembly, but always, and I mean always, keep the judgment.

§The Farmer’s Job

Back to the farmer.

The farmer doesn’t plant the seed and then dig it up every day to check on it. The farmer doesn’t do the soil’s job. The farmer plants well, trusts the process, and shows up at the right time to harvest.

That’s what human-in-the-loop means in practice.

• Design the system well.
• Define quality inputs.
• Set the workflows running.
• Then show up at the right moments to exercise judgment (reviewing the output, catching the errors, adding the human detail, and deciding what goes to market)

You’re not the machine, but you’re also not above the machine. You’re the farmer. You do the beginning and the end so the soil can do its work in between.

That’s the last foundational idea before we get tactical. You now have the full framework:

• The Wonka metaphor (pipes before the chocolate).
• The Iron Triangle (AI reshapes the constraints).
• The three levels (chat, workflows, agents).
• The human-in-the-loop principle (humans do the beginning and the end, the system does the middle).

Part Two starts now. Let’s lay some pipes.