Chapter 15 · Part Four: The Machine That Makes Average
This is the chapter I most wanted to write, and the one that’s easiest to get wrong.
The wrong version is a tools roundup. Here are the twelve apps I use, here’s my prompt library, here’s a screenshot of my workflow. That chapter is out of date before it’s published, and it teaches the reader to buy things.
So this one is built differently. It covers the systems running in my marketing function, described as patterns rather than products, each one the same way: what it does, what it replaced, what it costs, what to watch for, and how it scores on the six-month test. I’ve left out what’s specific to my company. You don’t need it. The patterns transfer, and you can build every one of them yourself.
And it ends with one I built that I never shipped, because a chapter of nothing but successes would be the exact victory lap this book argues against.
Here’s the idea that runs under all of it, and it’s the most transferable thing in this part of the book:
The unit of work is a document, not a subscription.
Every system below is a skill: a written set of instructions that a model follows, with sources to read, steps to take, rules to obey, and a format to deliver. Not software I bought. A document I wrote and keep improving. Which means a reader with no budget can copy all of it, and Appendix F has the versions you can run.
1. Voice profiles
What it does. Lets a model draft in a specific person’s voice well enough that the person edits rather than rewrites. We keep one for each executive who publishes, including me.
What it replaced. A ghostwriting process where somebody on my team spent an hour interviewing an executive, an hour drafting, and a day waiting for edits that rewrote half of it.
How it’s built. Not from their published writing. Published writing has already been edited, often by marketing, so it’s partly our voice already. The useful sources are the ones nobody polished: how they write in chat, how they answer an email at 11 PM, and above all their corrections to drafts. When an executive rewrites a sentence, the before and after is the most precise description of their voice you’ll ever get. A good profile records what they always do, what they never do, the words they hate, and how their punctuation behaves.
What to watch for. Parody. A profile that only lists a person’s habits produces a draft that uses every habit in every paragraph, and it reads like an impression rather than the person. The fix is a section most profiles don’t have: the imitation tells, a list of the things a copycat overdoes. Use their signature phrase once a month, not once a paragraph. Real voices are mostly ordinary sentences with a few distinctive ones.
Crutch or multiplier. Conditional, and the condition matters. It’s a multiplier when the ideas are theirs and the profile gives them their time back. It becomes a crutch the moment it’s used to publish when nobody had anything to say.
2. The weekly content system
What it does. Every week, an agent reads the news across our market, finds the stories worth having an opinion on, filters them, and proposes a short plan: which stories, which person should speak to each, and the angle. Then it drafts against the right voice profile. A human picks, edits, and posts.
What it replaced. Somebody scrolling for two hours on Monday, or more honestly, nobody doing it and posts going out when an executive happened to have a thought.
The part that matters. The filter, not the drafting. Ours asks whether the person posting can be the protagonist of the take: do they have a real point of view, earned from something they’ve done, or would they just be summarizing the news? A post that summarizes the news is the crutch column of Chapter 14. A post where somebody who has done the work says what they think is the multiplier, and the filter is what tells them apart.
What to watch for. The system finding the same stories every competitor’s system finds. If your agent reads the same feeds as everybody else’s, it’ll surface the same top five stories. Give it sources the others don’t read, and tell it to rank the strange story above the big one.
Crutch or multiplier. Multiplier, as long as a human still chooses. The day it starts posting without anybody deciding, it becomes a machine for producing the category’s opinion on schedule.
3. The Monday review
What it does. Every Monday, a set of agents runs against our CRM, our internal search and our documents, and hands me a read on where we are against our goals, how spend is pacing, and what we should be building that nobody has asked for yet.
What it replaced. That used to take somebody about two days, when it happened at all.
The honest claim. It didn’t save two days. That’s the crutch version of the story and it’s not the interesting one. What it actually did is give me a proactive read I never had before. The old version of this was reactive: somebody asked a question in a meeting and somebody else went to find out. Now the answer, and three questions I hadn’t thought to ask, are sitting there before the week starts.
What to watch for. Believing it. An agent that reads your CRM can only report what’s in your CRM, and every CRM has holes, stale stages and deals that exist mostly as optimism. The review is a starting point for a conversation with the people who own the numbers, not a replacement for it. Humans on the loop, always: the agent proposes, a person decides.
Crutch or multiplier. Multiplier. It produces a read that didn’t exist, and it changes what gets decided on Monday.
4. The competitive library
What it does. One page per competitor, always current. A sweep runs every week to catch what changed: new messaging, new pages, new claims, new hires, new ads. A deeper rebuild runs monthly. Alongside it sits a deal ledger, a running record of every deal where that competitor showed up, what happened, and why.
What it replaced. The battlecard, which in most companies is a document somebody updates the week before a sales kickoff and nobody trusts by March.
The decision that made it work. Wins and losses, including the loss reasons and who we lost to, are visible to every rep. Not gated to leadership. It would have been easy to keep the losses private, and it would have made the library much less useful. My reasoning was simple: it’s how we learn. A rep walking into a deal against a competitor deserves to know exactly how the last three went, especially the ones that went badly.
What to watch for. Sameness, again. A library that only tracks what competitors say will make you better at responding to them, which is a polite way of saying it’ll pull you toward them. Use it to find what everybody in the category is saying, and then do Chapter 2’s exercise: work out what nobody is saying.
Crutch or multiplier. Multiplier. The alternative wasn’t a slower library. It was no library.
5. The living timeline
What it does. A page that tracks one fast-moving topic in our market over time, every entry with a date, a source, and what we think it means. A scheduled agent checks for new entries every day and proposes updates. It runs internally today, and the plan is a public version with a short weekly recap as its own post.
What it replaced. The annual report on the same topic, which is out of date the week it ships, and which Chapter 5 already suspected of being annual for no reason.
Why it’s different. A report is a moment. A living resource is a place people come back to. It’s a better asset for search, for sales, and for anybody trying to understand the topic, and it’s a reason for a buyer to return that has nothing to do with being sold to. It also forces us to have an opinion every week rather than once a year, which is uncomfortable in exactly the right way.
What to watch for. The agent adding things that are wrong. Fast-moving topics are full of bad reporting, and an agent that reads the news will happily repeat it. Every entry gets a source and a human check before it goes up.
Crutch or multiplier. Multiplier. Keeping a resource like that current by hand would take a person’s whole job.
6. The internal hub
What it does. One place for a new rep to find the current version of anything: the deck, the competitive page, the sales plays, the partner material. Agents keep it updated from where the work actually happens, and a search box answers questions across all of it.
What it replaced. Document sprawl. Fast hiring, a dozen folders, three versions of every deck, and a chat tool where the answer to every question exists somewhere and finding it takes twenty minutes.
How it’s going. Honestly, it’s new. It’s in beta with our go-to-market team, there’s a feedback channel where people flag wrong versions and ask for things that are missing, and the measure of whether it works is whether new reps stop asking where things are. I’m including it before I know the answer on purpose, because the alternative is waiting until every example in this chapter is a proven success, and you already know how I feel about that.
Crutch or multiplier. Too early to call. Ask me in six months, which is the point of the test.
What all six have in common
- Each one is a document. Sources, steps, rules, format. Anybody on my team can read it, change it, and run it, and some of my team run skills like these without ever opening a terminal.
- Each one does work that wasn’t getting done. Not faster versions of old work. Proactive reads, current libraries, research nobody had time for.
- A human decides in every one of them. The agents propose and assemble. People choose what’s true, what gets said, and who says it.
- None of them required a budget. The cost is the time to write the instructions well, and the willingness to keep rewriting them when they’re wrong.
The one I shelved
Last September, we wrote a book.
It was called PLAID: The Operating Model for AI-Native Companies, and it took us less than a week. We were proud of that. It says so in the book.


It had ten operating principles, a marketing framework with its own acronym, a Michelin star metaphor running through every chapter, and an epilogue that ended “The revolution isn’t coming. It’s here.” It read like a book. It had the shape of a book.
We never published it. Nobody told us not to. It just never became the priority, and every time I opened it I could feel why. We’d been selling for a few months. Most of the stories worth telling, we couldn’t tell publicly yet, so the ones that made it in were “a Fortune 1000 retailer” and “a manufacturing company,” and the claims were the kind you make before you can prove them. We were writing the operating manual for a company that hadn’t operated long enough to need one.
So I stopped. Not because it failed. Because it wasn’t good enough, and I knew it, and nobody else was going to say so.
The ideas lasted. All of them. They’re how we run the company now. The book didn’t, because the ideas came from doing the work and the book came from a week with a machine.
Read it again with Chapter 13 in mind and it’s almost funny. Not AI-enabled, not AI-enhanced, AI-native. A comparison to the Model T. A section headed “This Changes Everything.” It’s the wall of homepages at book length, and I wrote it. With a model doing the structuring and smoothing, we got the most probable book about building an AI company, and the most probable book is the one everybody in the category would have written.
It’s also the other side of Chapter 8. With the car, Ed was the one person who didn’t nod, and I ignored him. With the book, the person who didn’t nod was me, and this time I listened.
AI made it possible to write that book in a week. It couldn’t make us ready to write it.
That’s the limit of everything else in this chapter. A model can produce the artifact. It can’t produce the proof that the artifact deserves to exist: the customer stories you’re allowed to tell, the results you can stand behind, the years of doing the thing. When the artifact arrives before the proof, it doesn’t look early. It looks finished, and that’s what makes it dangerous.
The Comfort Check
Chapter 15: building with AI
1. The comfortable assumption. Using AI well means buying the right tools.
2. The true objective. Do work that wasn’t possible before, with people still deciding what’s true.
3. The redesigned approach. Write the work down as skills anyone can run and improve, aim them at work nobody was doing, and don’t ship the artifact before you have the proof.