A Lever Made of Agents

AI as a new source of operating leverage where the durable advantage comes from process power, not prompts.

Editorial note: I’m taking on a few additional companies looking to find the highest-leverage place to put AI to work in their business, and build it. If you’re interested, you can find out more here: https://taylorpearson.me/ai-roadmap

When the internet showed up, every newspaper in the country had the same idea: put the paper online. Readers everywhere, no printing costs. It looked like free money.

US newspaper ad revenue peaked around $49B in 2005, down ~80% in fifteen years.

Using newspapers as an example of technological disruption is a cliche at this point, but I’ll be lazy and use it because it’s a clean example of how technology changes supply/demand dynamics and what the result is.

Supply went towards infinity: anyone could publish anything online. Demand for digital media went up but not nearly enough to compensate. We learned again that the price of anything abundant goes one direction: toward zero.

A lot of the excitement around agentic AI is about vibe-coded apps and one-shot prompting. This is 100% super cool! It’s how I started playing with things and it got me hooked on using Agentic AI. It is, however, abundant.

To borrow another phrase from the economists: solve for the equilibrium. The equilibrium price of anything you can build with a one-shot prompt is the token cost to build it because anyone else can build it too.¹

AI will likely have a lot of far reaching implications that I have not even considered, but let me propose one that based on my personal are a of expertise.

I’m a systems guy. A lot of my career has centered on building systems and processes at small/early-stage companies. I like the Toyota Production System, Theory of Constraints, and SOPs.

So, when I started using Agentic AI tools like Claude Code, Codex, and OpenClaw, what got me excited about them was the potential to use a new technology to build business systems in new and defensible ways.

The question I’ve been asking as I work with these tools is how do you use AI to do this in a way that builds sustainable competitive advantage? What is the new capability unlocked by this technology and how do individuals and businesses best leverage it?

Most people’s answer so far is a chatbot in a browser tab. You open Claude or ChatGPT, paste something in, get something back, paste it where it goes, and do it again. For most use cases, you end up about twenty percent more efficient on the things that were mildly annoying to begin with, and none of it touches the work that actually moves the business.

A key shift is the move from chatbots to agents. In As We May Work, I argued that agent harnesses (Claude Code, Codex, OpenClaw, Hermes, etc.) give AI two new and important abilities that transform how AI can be used.

1. Long-term Memory. It can read and write notes to its future self (context, decisions, history) that persist across sessions.
2. Agentic-ness. It can execute multi-step work autonomously while you’re doing something else.

The harnesses started as tools for programmers. They’re becoming the place where work happens, a new operating system layer for knowledge workers.

In The AI Knowledge Work Stack, I argued that a clear architecture is emerging for working with these tools.

The bottom two layers of the stack are the AI model (Opus, Fable, GPT, DeepSeek, Gemini, etc.) and the agent harnesses. Models and harnesses at this point are pretty clearly their own product categories and it doesn’t make much sense to roll your own for all but the largest businesses.

The durable leverage in this architecture for most businesses lives in the local scaffolding layer: what sits on top of the harness, pulls context from your other services, and tells the agent how your work works. ²

This is your AI business brain: the thing you own that makes the company more effective and defensible.

What this looks like is going to be different for every company, but I think it has a few core form factors at this point:

• Memory — Files where knowledge of the business accumulates: the current state of every project, the history of what’s been done, and the reasoning behind decisions. It’s what makes starting a session with the AI feel like talking to someone who already knows you and your business and what makes it defensible.
• Skills — Something akin to standard operating procedures (SOPs) written for an AI instead of an employee: how to draft the newsletter, close the monthly books, turn sales-call notes into follow-ups. The agent pulls in the right one when the task calls for it.
• Bespoke software — Everything from large software (e.g. a custom CRM) to small custom scripts and tools for the work that needs the same answer every time: pricing, reports, the bookkeeping math.

The newspapers adopted the internet just fine; what they never asked was what it would do to the economics of their business. AI puts the same question to every business: when your competitors have the same models you do, where does durable advantage come from?

7 Powers

In his book 7 Powers, Hamilton Helmer has a framework for the seven sources of economic power: the things that generate persistent above-market returns in a business.

1. Scale Economies — per-unit costs decline as volume grows (Amazon, Costco).
2. Network Effects — value to a customer increases as the user base grows (Facebook, Visa).
3. Counter-Positioning — you adopt a model an incumbent can’t mimic without cannibalizing their existing business (Netflix’s no-late-fee subscriptions against Blockbuster, whose model relied on late fees).
4. Switching Costs — customers expect a greater loss from switching than the value they’d gain from a competitor (SAP; QuickBooks, once it holds five years of your books).
5. Branding — higher perceived value for an objectively identical offering (the same diamond costs more in a Tiffany box; Coke against store-brand cola).
6. Cornered Resource — preferential access to something coveted (drug patents, Pixar’s “Brain Trust” of star directors).
7. Process Power — your organization and activity set produce lower costs or superior product, and matching it requires extended commitment (Toyota’s production system).

The first two of these (scale economies and network effects) are structurally closed to small and early-stage businesses (sub-50 people). Scale economies require scale and network effects require a network. You can grow into both, but no one has them out of the gate and very few businesses ever get them.

Two more are usually filed under big-company powers but exist for some small businesses in weaker forms. Any service business that accumulates deep client knowledge (the accountant who already knows your books, the IT firm that manages your whole stack) is building switching costs. A liquor license, an exclusive dealer territory, or a fifteen-year lease at below-market rent on the best corner in town can be a cornered resource. In my experience, not that many businesses have these and they are sufficiently weak at most business’s scale that neither is usually decisive on its own.

Counter-positioning is sometimes available (Vanguard against actively-managed mutual funds, Dollar Shave Club against Gillette). But it requires a structural conflict at an incumbent, and most startups and small businesses compete in fragmented markets where there is no incumbent to counter-position against.

That leaves two powers available to essentially every business: Branding and Process Power. This is what most businesses actually rely on. They do something good reliably (process power) and have a reputation for it (branding).

Most businesses start because the founder is good at something or has a good idea. They get to some meaningful stage because the founding team is good enough at sales/marketing plus some other competency.

The rainmaker agency founder is better at paid ads than anyone they’ve hired and can sell. Suddenly there’s a team of media buyers now, but they are still the one closing every major account, setting the strategy, and when a campaign underperforms it’s their gut that knows which lever to pull.

Or the e-commerce founder who sells equipment to gym owners: there’s a team running fulfillment, ads, and customer service, but he’s the one who knows the customer well enough to see what they’ll want next and every winner in the catalog was his call. At this stage, the founder is basically what’s defensible. The founder is also the constraint: the business can’t grow past his calendar, and he can’t take six weeks off.

This is a pretty well understood and common phenomenon. Books and approaches like The E-Myth, Work the System, Traction (AKA EOS), and Scaling Up are all variations on solving this same core problem and helping businesses build process power.

The Four Levels of Organization

When you break down what those approaches prescribe, you can think of it as four levels of organization: 1. Software 2. SOPs 3. Expertise 4. Culture

• Each level is know-how living in a different place.
• Culture lives in myth: how-we-do-things-here, absorbed by being somewhere, usually written down nowhere.
• Expertise lives in a person’s head: it does the job reliably, but it can only partly explain itself, and it walks out the door at five o’clock.
• An SOP lives on paper: expertise made explicit enough that a competent hire can run it.
• Software lives in code: it is knowledge made so explicit that a machine can execute it with no judgment at all.

The dimension running up the pyramid is explicitness: how fully the knowledge has been (or can be) spelled out.

The more explicit the know-how, the cheaper it is to run so each level is also a different cost structure. – Knowledge in a head you pay for every time it runs: an expert’s salary. – Knowledge on paper you pay an expert for once, and then it runs on lower wages. – Knowledge in code you pay for once, and the marginal run is close to free.

Finance people call this operating leverage: costs that stay fixed while volume grows. It’s why a software company’s margins expand as it grows while a service firm’s margins usually don’t. When the next customer costs almost nothing to serve, growth compounds into profit instead of into headcount. A business with no operating leverage has to buy every new dollar of revenue at the same cost; a business with a lot of it gets each new dollar cheaper than the last.

Getting systematized is a way to generate operating leverage. It’s a financial engineering move sold to entrepreneurs as a stress reducer. More operating leverage is easy to convert into more vacations and people like piña coladas more than financial engineering.

No matter how you frame it, the result is that every procedure you push up the pyramid converts a little bit of payroll into an asset. This means either higher margins (and less stress) or the opportunity to reinvest for growth.

Moving work up the pyramid has a one-time cost. An expert documenting a process might spend three times as long as just doing it, because they have to make explicit every call they normally make without noticing.

But there’s a clear threshold where it’s worth it. Take a task you do every week that takes about an hour, and assume your time is worth $100 an hour. Three hours to write the SOP is a $300 investment. Hand the task to someone at $30 an hour and you save $70 a week, about $3,600 a year.

The SOP pays for itself in about a month and returns roughly 1200% in the first year. An index fund compounding at ten percent takes about twenty-five years to return that much.

Many businesses have dozens or hundreds of those tasks in them, each one running on someone’s memory. Most owners have never added them up. Not all know-how can (or should be) made explicit, but it’s a waste to not push things as far up the pyramid as possible.

Through the lens of operating leverage, each of these books is a strategy for moving work between the layers, and together they form a rough sequence.

The E-Myth (Michael Gerber) and Work the System (Sam Carpenter) are where many businesses start as they both focus on building out written documentation. The move is the same in both: get the work out of the founder’s head (Expertise) and into SOPs so that someone other than the founder can do it.

Carpenter wrote Work the System about his telephone answering service, a business that was almost entirely people executing procedures. Document the procedures well enough and you can run the business on people you can actually afford to hire. The most deterministic of those SOPs migrate up another level, into software. Larger companies push the same idea further and build internal tools.

Gino Wickman’s Traction (and Verne Harnish’s Scaling Up, its equivalent for somewhat larger companies) is the next level. Most of its artifacts are SOPs for managing the people who run the SOPs: the standing meeting agenda, the weekly Scorecard of KPIs, quarterly priorities, a mechanism for cascading strategy through the organization.

They are playbooks for coordinating the bottom two layers: expertise and culture (e.g. right person in the right seat, core values explicit enough to hire and fire against) because at that size the constraint is no longer documenting the work; it’s keeping the humans aligned while they do it.

Technology and New Forms of Operating Leverage

Over the last couple decades, software tools have gotten cheaper and more accessible. Underneath the palm trees, Tim Ferriss’s The 4-Hour Workweek was the manifesto of the internet’s deployment era with two storylines.

The first is a leverage story: email marketing, self-serve ad platforms, and cheap offshoring gave a small, internet-based business the kind of operating leverage that used to require an enterprise.

The second is a threshold story about converting that operating leverage into growth within existing businesses and by lowering the minimum viable scale of a business.

Chris Anderson called the lowered minimum scale on the demand side the long tail. The internet aggregated niche demand that had never been worth serving. Suddenly it was possible to start a small business with customers in 100 different countries.

Combined with new tools plus offshoring on the supply side, a small business could suddenly serve what seemed like a small niche profitably.

I wrote a book on this a decade ago, The End of Jobs, which argued that the long tail had made a whole new type of business viable. One business owner I profiled, Andrew Youderian, built a business around CB radios, a niche too small for any physical store, by writing online guides that answered buyers’ questions, work that used to require a salaried person on the floor.

The specific playbook in 4HWW got arbitraged away fast. A chunk of the book was, in practice, “Google SEO and AdWords are underpriced.” That stopped being true about as fast as you’d expect, but the broader phenomenon it pointed at ran for two decades.

Some niches stayed small and others turned out to be new categories. Rogue Fitness started in a garage selling equipment to other garage gym owners and now owns the category. RXBar started in a basement making protein bars for CrossFit gyms and sold to Kellogg for $600 million.

No-code tools like Zapier entered the picture in the 2010s and were symbolic of the tooling available to non-technical businesses to push tasks that used to require humans into software.

It was suddenly viable to set up an abandoned checkout flow that sent a personalized follow-up email and added a tag in the CRM. A brick-and-mortar shop could never do the equivalent; it would have meant paying an employee to trail every browsing customer out of the store asking follow-up questions. Ignoring the fact this would be creepy, it’s also expensive.

The same payback math applied. A task someone does an hour a week at $30 an hour is a ~$120-a-month line item; an afternoon of building converts it into a thirty-dollar workflow, and it pays for itself in about a quarter.

The wage line became a software line, businesses got more operating leverage, and automated follow-up went from impossible to table stakes in about a decade.

All of these historical tools for creating operating leverage still work and still matter. But every book in that canon is grounded in the same assumption about the structure of the pyramid and the tools available. And we just got a new tool: thinking machines.³

The Five Levels of Organization?

Agentic AI is capable of executing lots of tasks that were historically only possible with a human. It also makes it dramatically cheaper for humans to build software by making software developers more productive and unlocking software development for lots of non-developers. In short, they are a new source of operating leverage.

Agentic AI offers both an opportunity and a threat to building a business with process power in the same way that the internet unlocked new business models and challenged old ones.

Process power works like compound interest. It’s the accumulation of a thousand tiny improvements. It’s things like a better way to greet a customer and a checklist that catches one more defect. Any one of them is trivial and easy to copy.

What turns that pile of improvements into something a competitor can’t easily copy is the emergent value of all the integrated components. The thousand small things start to depend on one another until the system stops being legible from the outside. When GM got a full tour of Toyota’s plant and copied every visible procedure, they still couldn’t reproduce the results.

Helmer’s term for the barrier is hysteresis: the advantage can only be built by walking the path because each improvement is the soil the next one grows in. Bashō said it more cleanly four centuries ago: “Do not seek to follow in the footsteps of the wise; seek what they sought.” A competitor can copy your footsteps, your current procedures, but the footsteps aren’t where the power is. Process Power works the same way: you don’t design it from the top down, you craft it through small choices over time.

You Get A Process Power and You Get a Process Pwoer

What agentic AI uniquely enables is a shift in the level of granularity and scale at which individuals and companies can build process power. Historically, one person could only manage so many processes before needing to hire someone else to take over managing them.

The rise of agents has changed that calculus. One person or a small group can now oversee an army of agents and custom software that allow them to operate at a historically unprecedented scale.

 

Here’s an example of what that looks like in miniature. My monthly newsletter is a curation of the best things I read each month. As I read, I highlight and annotate whatever catches me.

When it’s time to draft the newsletter, I used to manually go through all the articles and pick out the ones to put in the newsletter. Now, an AI agent using a skill file ranks the month’s reading by how much I marked it up, writes a report on each top candidate, and hands me ten to twelve to choose from.

I pick the four to eight worth sharing and write the editorial take. The same skill routes leftover highlights to the article drafts and projects they’re relevant to, turning my reading notes into actionable tasks for things I’m working on.

Formatting the newsletter for each platform (X, LinkedIn, Email, etc.) used to eat a couple hours of fiddling; now a skill file and a small script do it in two minutes. A separate skill generates a custom image optimized to each platform’s algorithm and size requirements. That’s a little bit of operating leverage.

The margin story of AI is to replace humans with tokens. The growth story is to use tokens to make new forms of work and human agency economically viable. You see both here.

There are agents enabling more margin by going through the docs and formatting them faster. There are also agents unlocking new capabilities that never happened: integrating every note or highlight into a relevant project as reference material and generating custom images.

Other Than the Sum of Its Parts

Scale the same shape up to a full service business. An independent financial advisor with fifty client households could have an agent maintain a folder per household: account structure, the plan, the tax picture, and a timeline of every decision and the reasoning behind it.

A pre-call skill preps every meeting. A post-call skill turns her notes into action items and files the why. When she figures out the right way to handle a client’s car-purchase question, that method becomes a skill that runs against all fifty households. The portfolio math lives in small scripts that can pull from custodian APIs. None of the pieces is special on its own, but the combination of them can create a huge amount of operating leverage.

It’s the SOP math run across the whole business. Each one may be another small build throwing off twelve-times return, and there are dozens of them hiding in most businesses. Stacked up, the time and cash they free adds up to something significant.

That freed-up time and money has two places to go. One is margin: the payroll that just became an asset and the hours that are saved drop straight to the bottom line.

The other is recombination into new possibilities.

This is a one-person business, but it scales again to a larger business where other people have access to the company’s institutional memory and skills. For the founder who was the constraint, the job goes from operator to designer: you stop being the person every decision routes through and start building the system that handles them. A procedure you design once runs every week, across every client, without you in the loop.

The opportunity for both companies and individuals is to retool how they think about and execute process power in a post-AI world, what I’ve called freestyle work.

The First Percent

The 4-Hour Workweek moment was about the realization that there was a new source of operating leverage for existing businesses and a whole new type of business that became possible.

That’s the same set of questions agents re-open. What becomes viable when the SOPs execute themselves and the custom software gets built in an afternoon? That playbook hasn’t been arbitraged away yet and we’ll see it play out over the next few decades.

*I am taking on a few additional companies to work with to find the highest-leverage places to put AI to work in their businesses. If that’s useful to you, you can see how it works here: https://taylorpearson.me/ai-roadmap.

 

1. That’s not to say there won’t be lots of intermediate opportunities to use this as an acquisition channel. There for sure will be!
2. “Lever” comes from Old French levier — literally the agent noun of lever, “to raise.” So “a lever made of agents” is etymologically self-referential: a lever is an “agent.”
3. Yes, this is a Butlerian Jihad reference. No, it doesn’t matter if you don’t get it.