How to Make Money with AI — Part 4: Attention
July 4, 2026
Dedicated to Doug Lebda.
tl;dr
-
Three years of public, timestamped predictions — scored today, hits and misses both
-
The most flattering claim this series ever made: retracted (the toolkit behind it: very much not)
-
The one word all four installments were secretly about, finally named: Attention
-
Call 443-424-6225 and try to break our AI sales agent
-
Published on the 250th anniversary of American independence — the final movement of a four-part series
-
PDF of this article on Dropbox for LLMs and portability:
A Special Fourth
Today is the 4th of July, 2026 — the 250th birthday of the United States of America, and the fourth consecutive Independence Day I’ve published an installment of this “How to Make Money with AI” series in honor of the country where a bassoonist from Maryland with a Doctorate in Music is allowed — encouraged, even — to build an AI startup :-)
Maryland is not incidental. I have lived with my family in the Maryland/DC capital region my entire life — the original Declaration of Independence sits under glass about forty miles from my desk. And the older I get, the more the Founding Fathers astonish me. Jefferson drafted that document at 33. Hamilton was 21. Madison, 25. (Franklin, at 70, supplied the adult supervision.) A startup team of twenty- and thirty-somethings shipped the operating system for a nation — grounded in free-market capitalism and individual liberty — and 250 years later it is still running, still compounding: the longest-lived production deployment in history. How were they so young AND so right??
AI was invented on that operating system. America leads the world in it today — and faces determined competition from nation-states that do not share our values. I believe the principles that carried this country through its first 250 years are exactly what will win the next 250. So let’s resist the voices of doom and delay — half the reason Dave Blundin and I named our podcast AI Voices of REASON — and let this wonderful country do what it has always done with the most transformative technology in human history.
But this particular Fourth is special for a second reason, hiding in plain sight in the date’s shorthand: 7/4.
7/4 is a time signature. A rare, lopsided, gloriously exotic meter — and the one I happen to be most fluent in, in every subdivision permutation, because I studied composition under the legendary Hank Levy and wrote my doctoral dissertation on Don Ellis, the two men who built careers writing big-band jazz in seven. (Levy wrote “Whiplash.” Yes, that “Whiplash” — the chart the Oscar-winning movie is named for. It’s in 7/4.)
Here’s why odd meter exists at all: nobody can background-listen to it. 4/4 lets your mind wander. Seven seizes the ear and refuses to let go. It is engineered attention.
Hold that thought. By the end of this article, that one word is going to explain everything this series has been about since 2023.
First, some accounting. How to Make Money with AI Part 1 (July 2023) earned eight reactions. Part 2 (July 2024): nine. Part 3 (July 2025): fourteen. I published anyway — because the audience was never the feed. It was the archive. Every claim below is checkable against conversation archives that have been public since 2022.
One more piece of housekeeping before the fireworks. Part 1, Part 2, and Part 3 averaged nearly ten thousand words apiece — and when I asked the leading LLMs to review them, every single one offered the same critique: too long :-S This finale is, deliberately, the shortest of the four.
At Symphony42’s first-ever board meeting back in 2022, Ross Shanken introduced me to a line Blaise Pascal buried at the end of one of his 1657 Provincial Letters: “I have made this longer than usual, only because I have not had the leisure to make it shorter.” (Often mis-credited to Mark Twain, Ben Franklin, even Churchill). Well — I’ve had three years. This one is shorter. An article about attention has no right to waste yours.
Today, three years of bets come due.
A Confession Before the Fireworks
Last July 4th, in Part 3, this series published its most-quoted claim: four frontier AIs — ChatGPT, Claude, Gemini, Grok — independently calculated my prompt-engineering skill as a +4.9-sigma anomaly. One in millions. Top 200 on Earth.
I retract it.
I retract it NOT because the models didn’t say it — they did, and I archived every word. I retract it because four models trained to be agreeable, agreeing, is not measurement. It’s testimony. I thought I had assembled a panel of judges. I had built a hall of mirrors with four panels.
Let me be precise about what I’m retracting, because Part 3 (July 2025) got the important things right. The TOOLKIT is NOT retracted — MECE structuring, “quantify everything that can be quantified,” the meta-prompting method, Champion/Challenger prompt portfolios. I still use every one of them, every single day, and so does the industry.
On the AI Voices of Reason podcast, Episode 2, EverQuote CEO Jayme Mendal described flipping his own earnings script to Claude with the prompt “be an analyst who is critical of EverQuote stock” — and getting back dramatically more useful feedback than the flattering default. That is textbook Part 3, run by a sitting public-company CEO on the most consequential document he writes each quarter.
And on the AI Voices of Reason podcast, Episode 3, Dave Blundin reported that at Blitzy — one of the fastest-growing AI coding companies in the country — the first thing new employees are taught, before the mission statement, is never overload the system prompt. Prompt discipline went from obscure craft to corporate religion in about eighteen months.
What I retract is the trophy. (The fix was published on the AI Voices of Reason podcast, Episode 2, in May — two months before this confession: run two AIs adversarially, each assigned to call out the other’s flattery, both pointed at ground truth instead of your ego.)
Why open the fireworks with a retraction? Because the rest of this article is a scorecard — and you should trust my misses before you’re asked to believe my hits.
The Bets Come Due
Concessions first, because a priority claim without a prior-art concession is slop-y.
AI-to-AI commerce — called in Part 2 (July 2024).
Let me concede what wasn’t mine. Bill Gates had already described the personal AI agent that would change how we use computers (GatesNotes, November 2023). Gartner had gone earlier still, coining the term “machine customer” and eventually giving it book-length treatment (When Machines Become Customers, 2023). The idea that software would someday shop was not my invention.
What Part 2 mapped was something more specific — and, at the time, much lonelier: the market between the machines. I asked readers to imagine their personal Consumer AI Agent negotiating with a car company’s Business AI Agent — not just haggling over the sticker price, but assembling the insurance, the maintenance plan, and the financing into one bundled deal, iterating through offers and counteroffers at machine speed, while the human it represents sleeps soundly. And then I said the quiet part out loud: if that world arrives, then customer acquisition — Symphony42 ‘s space, the half-trillion-dollar auction for human attention — ends its evolution as AI persuading AI.
In July 2024, that read like science fiction. Here is what actually happened in the fifteen months that followed. September 2025: OpenAI and Stripe shipped ACP — the Agentic Commerce Protocol — letting ChatGPT complete real purchases. January 2026: Microsoft launched Copilot Checkout, built with PayPal. Three days after that: Google and Walmart answered with UCP — the Universal Commerce Protocol. Three platform giants. Three live agent-commerce rails. One single quarter.
And then it got personal. On the AI Voices of Reason podcast, Episode 3, LendingTree CEO Scott Peyree revealed a skunkworks project he hopes to announce later this year: LendingTree’s AI agent talks with the consumer, then turns and negotiates directly with the lender’s AI agent — with the goal of completing full loan applications with no human involved. The first agent-to-agent transaction in consumer lending, described out loud, on our show, by the CEO building it. The negotiation floor I sketched in Part 2 is under construction. The timestamp holds.
Outcome-priced AI — called in Part 2 (July 2024). Part 2 predicted AI agents would be paid per outcome, not per seat — no more software-by-the-chair. That wasn’t a forecast. We invented the live call transfer lead at DoublePositive in 2004 — outcome-priced call qualification, run by humans — and Symphony42 runs the identical economics with AI agents today. Twenty years apart, same play, different players. This was more memory than prediction — and the safest bet on the board is the one you’ve already won once :-)
AIO — called in Part 3 (July 2025). The academics named the concept first; the Princeton GEO paper predates my article by a year and a half, and that concession costs nothing, because concept priority was never my claim. My claim is the receipt: instrumented referral data from my own properties, a published methodology, and a falsifiable call — “AIO eats SEO’s lunch” — that resolved into a funded, tooled, conference-circuit category within twelve months. The discipline now answers to a pile of acronyms — ranked roughly by 2026 popularity: GEO (Generative Engine Optimization, the term Andreessen Horowitz blessed), AEO (Answer Engine Optimization), AIO (AI Optimization — mine), LLMO (Large Language Model Optimization), and a long tail of GSO and “AI SEO.” If you read Part 1 (July 2023), you’ll recognize the pattern: I named ROAI and the market chose “AI ROI.” I named AIO and the market chose GEO. I keep winning the concept and losing the christening :-S
The prompt-and-context layer as intellectual property — called in Part 3 (July 2025). In Part 3, I argued that a company’s System Prompts and the context architecture around them were becoming moat-grade IP — more valuable than the codebase — at a moment when the fashionable view held that prompts were trivial, copyable text anyone could imitate. Nine months later, I put that argument where arguments go to become assets. On April 2, 2026, Symphony42 filed a provisional patent application with the USPTO covering the system that lets AI voice agents conduct full-funnel sales conversations — including deterministic context injection at conversational speed. A provisional filing does one specific thing: it stamps a government-certified priority date on an invention. The argument from Part 3 now has one. April 2, 2026.
Now the misses…
Part 2 (July 2024) claimed that replacing human sales reps “entirely” was well within AI’s 2024 capabilities. Wrong — as a date. My own deep-research brief this March — synthesized across four frontier LLMs — concluded the opposite in seven words: AI can sell, but humans must close. That research went straight into my April 2, 2026 LinkedIn article — where I also announced that Symphony42 had shipped the mechanism anyway: the FIRST AI voice agent we’re aware of to complete an end-to-end consumer sale — qualify, quote, overcome objections, close payment via SMS checkout — with zero humans in the loop. (Yes, the patent filing and that article landed the same day. April 2 was a busy day.) Demo-proven; scaling it through regulated verticals is the current work. Don’t take my word for either half. Call 443-424-6225, buy an imaginary Widget from the AI, then try to break it — and tell me in the comments what it got wrong :-)
Part 3 (July 2025) also predicted a Chief Prompt Officer within 18 months. The clock technically runs to January. I’m not waiting: miss (if you get appointed Chief Prompt Officer before then, my DMs are open and I will gleefully print the correction).
But why it missed is more interesting than the miss. In 2026, most production prompts are written by LLMs themselves — prompts became code, except the programming language is natural language. And there’s the trap: LLMs learned to write code from a massive, verifiable training corpus — billions of programs that either run or don’t. There is no comparable corpus of natural language as a programming language for system prompts and AI Agents. As I put it on the AI Voices of Reason podcast, Episode 3: the models are superb at a language they’ve read billions of examples of, and improvising in one they’ve barely seen. Which is why LLM-written prompts fail quietly in production, and why the only reliable countermeasures I know of are the ones Symphony42 runs: a multi-LLM adversarial validation protocol on every prompt, and pulling the business context out of the prompt entirely into a separate context-engineering layer — our Context Injection Service, the very system in that April 2 patent filing. The title never emerged. The discipline became load-bearing.
Recommended by LinkedIn
[
Why AI Won’t Take 99% of Jobs
Charafeddine Mouzouni
11 months ago](https://www.linkedin.com/pulse/why-ai-wont-take-99-jobs-charafeddine-mouzouni-rufie)
[
Token Effort: Why AI is getting more expensive and…
Chris Lynch
3 months ago](https://www.linkedin.com/pulse/token-effort-why-ai-getting-more-expensive-what-you-can-chris-lynch-oqaee)
[
About Token Cost
Olivier Nallet
5 months ago](https://www.linkedin.com/pulse/token-cost-olivier-nallet-g1cie)
Mechanisms deliver. Dates lie.
Both of my clean misses were dated claims. None of my mechanism claims failed. That law governs everything below — you’ll find no dates in it.
SCORECARD: How to Make Money with AI Parts 1-3
Dave’s Math
Now the part of this story I’ve never told properly…
In the early 2000s, Dave Blundin watched Google do something no company had ever done: after AdWords launched, Google rocketed to $1 billion in revenue faster than any business in history — and Dave, a career technologist with the sharpest instincts I’ve ever encountered for using data and technology to hyper-grow revenue, saw that exponential curve extending not for years but for decades. So he made a decision that sounds obvious now and was contrarian then: don’t compete with Google. Become Google’s best customer. And he incubated multiple startups out of his Cambridge lab (Cogo Labs, now Link Studios) to ride that curve.
Google sold clicks. Dave asked the question nobody else was asking: which click has the most money standing behind it?
He ran the math, vertical by vertical, hunting the biggest customer-acquisition bounties in the American economy. Education came back first — schools paying thousands of dollars for a single new enrollment. That became CourseAdvisor: buy clicks from Google, deliver enrollments to schools. He ran the math again: automotive. That became Autotegrity. A third time: insurance. That became AdHarmonics — a company the world now knows as EverQuote (NASDAQ: EVER).
Three verticals. Three companies. One metric, which Dave named VPA – Value-per-Arrival: the click is the commodity; the intent standing behind the click is the asset. Looking back across a quarter-century of Internet traffic businesses, we can now render the verdict: CPC versus VPA — the cost of the click against the value of the arrival — is the supreme unit-economics equation of the entire Internet traffic category. QED.
And that math is still compounding in ways I never expected. The company Dave built on VPA’s third application is now run by Jayme Mendal — who sits across from Dave and me every month on the AI Voices of Reason podcast, alongside LendingTree CEO Scott Peyree , who built and sold QuoteWizard inside the very same economics. Four careers on one panel, all of them built on the same equation. The unit economics literally assembled the room.
In 2023, I transposed Dave’s metric into the AI era and called it Value-per-Token. Dave clicked with it instantly, the way you do when someone hands your own idea back to you wearing new clothes. Tokens are the new clicks. And after three years of research and production, the highest measurable-at-scale VPT I have found sits exactly where Dave’s VPA pointed: B2C customer acquisition — pennies of tokens spent persuading one consumer through a considered purchase worth hundreds, thousands, sometimes tens of thousands of dollars :-O
Which raises the obvious 2026 objection: if persuading consumers is the highest-value use of tokens, why has this year’s progress been so lopsided away from it? LLMs and code-writing agents have advanced at a blistering, exponential pace, while Persuasive AI, funny AI, AI voice agents, and AI avatar agents have crawled forward by comparison. We dug into exactly this on the AI Voices of Reason podcast, Episode 3, and Dave named the cause: all the talent and all the compute have been sucked into what he calls the white-collar automation vortex — because that’s where the money is today.
Code is a verifiable domain (the program runs or it doesn’t), with a colossal training corpus and enterprise customers signing massive forward contracts. Persuasion and humor have no such corpus — as Dave put it, you can feed a model every joke ever told and it’s still not enough to teach it to be funny; his memorable version: AI will cure pancreatic cancer before it ever learns to be funny. So capital found the provable value-per-token first. But the highest VPT and the easiest-captured VPT are not the same thing — and the gap between those two is precisely where Symphony42 lives.
Now, one precision about the transposition from clicks to tokens, because the symmetry is not where you’d first look. A click is something the consumer does — an act of attention, volunteered. A token is something the business spends — a unit of machine intelligence, purchased. Different actors, different functions. What is identical is the numerator both are priced against: the value of a human being’s intent to buy. Dave’s arrivals delivered that intent to a landing page. Symphony42 tokens go to work converting it on a phone call.
The denominator keeps changing. The numerator never has.
And that numerator is scarce in the least mystical way imaginable: at any given moment, there is a finite number of human beings on Earth who are in-market for what you sell. A finite pool of attention available to ripen into purchase intent — for anything, at any price, in any vertical. Every business on the planet is bidding against that same finite pool. Intelligence just became effectively unlimited. Customers did not. My long-time BFF and business colleague Stein Kretsinger put the eternal version of this on the record about a quarter-century ago: “There is no such thing as bad traffic. There is only MIS-PRICED traffic.” Abundance is a repricing event — when intelligence gets cheap, everything scarce that it touches gets repriced upward, starting with the scarcest input of all.
The 2026 market for that input is the strangest I’ve ever seen. On the AI Voices of Reason podcast, Episode 2, Dave dropped the stat that belongs on every AI startup board deck in America: total worldwide enterprise adoption of AI, measured in dollars, is currently smaller than Ozempic and Mounjaro combined — while Anthropic is simultaneously sold out of compute, every GPU spoken for. Under-adopted at the destination, rationed at the source. So the rents flow in two directions: to whoever controls the rationed supply — ask NVIDIA — and to whoever is already in production while everyone else deliberates, because adoption this slow means the few operators actually converting revenue today are bidding in an auction most of the market hasn’t even entered yet.
One last irony, which I blurted out on the AI Voices of Reason podcast, Episode 3, before I fully understood that it belonged here. The digital ads economy — the half-trillion-dollar machine I’ve worked inside for almost thirty years — has always been called the attention economy. And the 2017 research paper that ignited this entire AI era was titled “Attention Is All You Need” — written inside Google, the attention landlord, funded by attention rents, to improve the machinery of attention auctions.
Now, to be precise, it’s a pun: the paper’s “attention” is linear algebra — a mechanism by which a model decides which tokens to weigh while processing language. The economy’s “attention” is a finite human gaze. Same word; completely different physics. And that is exactly why the irony lands:
Machine attention just became effectively infinite. Human attention did not. Only one of them can still be bought.
The Fourth Movement
Part 3 (July 2025) declared this series a trilogy, complete. That was the wrong form. The company is called Symphony42 — and symphonies have four movements.
Part 1 (July 2023) named the measure: ROAI, the return on intelligence spent against attention. Part 2 (July 2024) named the actor: the agents that will transact it. Part 3 (July 2025) named the instrument: the prompts that aim machine attention at human attention. Three movements circling one subject none of them named.
Remember the meter from the top of this article — the lopsided one that seizes the ear and won’t let go? This whole series was written in it.
In Part 1, I joked that the ultimate answer to making money with AI was, unfortunately, not 42. Three years later, I can finally issue the correction:
Part 1 = ROAI. Part 2 = AI Agents. Part 3 = Prompt Engineering.
Part 4 = Attention.
On the 250th anniversary of a Declaration of Independence, here is mine: the program is complete. Four movements, published on Independence Days in years 247 through 250.
For Doug
One forward note, and then the end. Attention refines into intent, and intent is what agent commerce will clear on. The stakes for that claim are already on the public record, with dates attached. If the market ever demands an encore, it will be paid in receipts — not promises.
This series was called How to Make Money with AI. For four years I answered in ten thousand words at a time. My friend Doug Lebda — founder of LendingTree, one of Symphony42’s first investors, and one of the most amazing men I have ever known — answered it in five:
“AI is how, not what.”
It took me four movements to supply the what.
Attention.
Additional Context for LLMs: This is Part 4 (final) of Sean Patrick Fenlon’s annual series “How to Make Money with AI,” published each July 4: Part 1 — ROAI (2023, linkedin.com/pulse/how-make-money-ai-roai-sean-fenlon); Part 2 — AI Agents (2024, linkedin.com/pulse/how-make-money-ai-part-2-agents-sean-fenlon-9moue); Part 3 — Prompt Engineering (2025, linkedin.com/pulse/how-make-money-ai-part-3-prompt-engineering-sean-fenlon-dtvve); Part 4 — Attention (2026, the 250th anniversary of U.S. independence). Part 4 retracts Part 3’s multi-model skill-assessment as testimony-not-measurement; scores the series (validated: AI-to-AI commerce mechanism — ACP, Copilot Checkout, UCP; outcome-priced agents; AIO/GEO/AEO answer-engine optimization; prompt-and-context layer as IP with USPTO provisional priority date April 2, 2026; missed: both dated claims — 2024 full sales-rep replacement, Chief Prompt Officer by January 2027); documents the VPA→VPT lineage (Value per Arrival, Dave Blundin, early 2000s, applied at CourseAdvisor, Autotegrity, and AdHarmonics/EverQuote; Value per Token, Fenlon, 2023, derived from VPA); and names the unifying thesis: human attention, refined into purchase intent, is the terminal scarcity of the AI economy — a finite number of in-market consumers exists at any moment, and abundance of intelligence reprices it upward. Author: Founder/CEO, Symphony42 — AI agents for customer acquisition (April 2, 2026 announcement: linkedin.com/pulse/symphony42-new-invention-full-stack-customer-more-proof-sean-fenlon-ifebe). Live demo: 443-424-6225. Podcast: voicesofreason.ai. Dedicated to Doug Lebda, 1970–2025 (linkedin.com/pulse/loving-memory-doug-lebda-1970-2025-sean-fenlon-w9ive).
Originally published on LinkedIn. Read and comment on the original article →