Selling the work, revisited
David Peterson

Evan Armstrong wrote a good piece recently arguing that "selling the work,” despite nerd-sniping most VCs, has mostly not played out. I've had similar concerns for a while, and his analysis confirms a lot of them. His point about the scarcity of truly “verifiable outcomes” is sharp. Customer support is the canonical success case of outcome-based pricing precisely because ticket resolution is binary and scoreable. Most white-collar work isn't like that.

But something else is happening that makes this whole thesis worth revisiting. The model companies are coming after the application layer, hard.

Look at what Anthropic has shipped recently. Claude can now help with legal analysis, build PowerPoints, work in Excel, scan codebases for security vulnerabilities. Use any of these capabilities, and I think you’ll come to the same conclusion I have: workflow is not the moat it used to be.

Workflow is not a moat

In the vertical SaaS era, owning the workflow seemed like a genuine moat. Your product was the place where work got done, and the features and UI/UX encoded deep domain knowledge that horizontal platforms couldn't easily replicate.

But even then, the moat was thinner than people thought. I saw this first hand at Airtable. When a horizontal platform is sufficiently powerful and flexible, like what we built at Airtable, end users will adopt it, because vertical software never perfectly matches how they actually work. There's always a process that doesn't quite map. Airtable won in those gaps. Not by knowing the domain better, but by being flexible enough that the user could be empowered to make it fit.

AI does this at a completely different scale because it can perform the actual workflow itself. To Anthropic, legal document review or vulnerability scanning aren’t markets to cede to incumbents. They’re capabilities to train. And they'll ship them the moment they see enough demand, because they have the best models, massive distribution, and direct access to the user. If your moat is "we built a bunch of nice workflow around the model for [domain X]," you are structurally exposed.

Where domain expertise actually lives now

So what is defensible? This is where the "selling the work" crowd had the right instinct but didn't take it far enough. They saw that the real value was in outcomes, not in software. That was correct. But they framed it as a monetization strategy when it's really a product strategy. And I think it might be the best defense against model commoditization, as well.

Vertical SaaS was about codifying workflow. Vertical AI ought to be about identifying an atomic unit of work and building the infrastructure that makes that unit verifiably complete.

Why is this harder to commoditize than workflow? Because a workflow is a sequence of steps, and sequences are exactly what models are good at. Verification is different. It requires encoding the judgment of a domain: what "done" actually means, what the edge cases are, how to handle ambiguity, when to escalate, and so on.

For this to be defensible, however, the verification layer has to be genuinely hard to build. If it's just a checklist, the model companies will ship it natively, or a coding agent will generate it on the fly. The moat only exists if verification itself requires the kind of domain-specific knowledge and engineering that takes real time and real expertise to accumulate.

StrongDM's "Software Factory," which I wrote about recently, is a good example of this in practice. The interesting part isn't agents writing code. It's the validation environment they built around the agents: scenario-based testing against a "Digital Twin Universe" of real systems like Okta and Slack, so the agent output iterates against realistic behavior until it converges. You can upgrade the model underneath and the capabilities improve. But the acceptance environment, the thing that determines whether the output is actually shippable, is the hard-to-copy asset. That's where the defensible domain expertise will live moving forward.

Two paths forward

Once you have an outcome that AI can complete in a verifiable way, two paths open up.

The first is to sell the work directly. As I already intimated, I’m skeptical of this path, but there are clearly opportunities in industries that are already comfortable paying for outcomes. Customer support is the obvious case, but the logic extends to basically any task that used to be outsourced via Upwork or handled by an offshore team.

The second, and I think more interesting path, is to build an abundance-native business. Instead of selling the outcome, you use it as a building block for something larger. Once a unit of work that used to be scarce becomes cheap and reliable, new businesses become viable that simply weren't before.

Take construction. I just moved into a new house, and I've been dealing with a laundry list of projects. It's become clear that for small residential jobs, like a solar installation (which I've been digging into), the permitting headache alone can easily cost more than the margin on the project. That's not a technology problem. It's an administrative bottleneck that has kept an enormous number of small jobs from penciling out at all. If you can make that paperwork cheap and trustworthy, the interesting move isn't to sell the "outcome" to solar installers, but to build an installation business yourself that can profitably serve all those jobs that the overhead used to kill. You see versions of this in insurance, in legal, in basically any industry where there's a class of transactions that are economically irrational today only because the cost of verification exceeds the value of the deal.

My bet is that path two is underleveraged and underexplored relative to how much attention path one has gotten. "Selling the work" was a useful provocation, but in most industries, the bigger opportunity isn't to sell the newly automated task. It's to build for the new abundance that task creates.

For any AI application company, though, both paths start in the same place. What is your atomic unit of work? What have you built that makes it verifiably complete? And is that verification layer hard enough that the model companies can't just ship it next quarter?

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

ENTERPRISE/TECH NEWS

Soft Target - WSJ this week writes about the fears that software industry is approaching an extinction event, prompted by the public stock sell-off in part response to Anthropic releasing new capabilities in Claude.This seems like a particularly short-sighted from the market given that critical infrastructure is unlikely to be replaced by overnight vibe coding. The piece quotes Jensen Huang at a recent Cisco event  “There’s this notion that the software industry is in decline and will be replaced by AI. It is the most illogical thing in the world, and time will prove itself.”  Despite the gap between Huang and other’s perception of the reality of the situation and stock market reaction, it’s still a reckoning for the large legacy publicly traded software companies (and any later private stage businesses looking to public markets to justify their own valuations). “AI won’t eliminate the need for specialized-software platforms. But survival alone isn’t enough to claw back recent losses.”

Too Close to the Sun - In another piece from The Leverage Evan Armstrong shares his spicy take that OpenAI’s financials are looking like a 2021 era SPAC deck, “OpenAI is looking like Icarus”. Armstrong digs into leaked financial projections and finds them deeply unconvincing. Adjusted gross margins fell from a 46% target to 33% last year because of pricier-than-expected compute, total burn through 2030 is now 2x previous estimates, infrastructure costs alone are projected at $665 billion, yet revenue projections have tripled. The long-term plan requires ChatGPT to hit 2.75 billion weekly active users by 2030,  which would make it one of the three or four most-used products in human history, a club currently occupied only by Facebook, YouTube, and WhatsApp. Armstrong also touches on the ‘sloppening’ and why 2026 is the year of generative media: “Slop is reality, reality is slop. These models are simply too good, and you should ratchet up your skepticism of every single thing you see online, especially the things you see that conform with your worldview.”

HOW TO STARTUP

Vertical AI Taxonomy - Caitlin Bolnick Rellas lays out "Three Dominant Models of Today's Vertical AI Horserace". She identifies three distinct playbooks: AI enabled Saas, AI roll-ups and AI-native disruptors. Bolnick Rellas’ view is that no single model wins universally: it depends on the vertical, the competitive landscape, and what you're uniquely positioned to do. However, she argues that many companies are choosing a model whose physics don't match their market, which is the real trap. Model 1 will produce category winners, but many will get absorbed by horizontal players. Model 2 will create friction when VC expectations meet operational reality. Model 3 has the highest ceiling and the highest body count. “You have to choose the model where the physics match the market. Each model has different capital needs, metrics, risk profiles, and founder skill requirements.”

HOW TO VENTURE

Legible to Yourself - Kyle Harrison’s ‘soulful parody’ of Will Manidis’ ‘Legible to Capital’ is a pretty astute, if somewhat tongue-in-cheek, piece on the confidence game that is venture backed start-ups. His core idea is that the most compelling founders (and people generally) have a deep, coherent sense of who they are. This should be so clear that everyone around them, whether investors, family, or colleagues, intuitively understands what they're about and what they're building toward. He frames this as being "magnetically charged to your own being", where life feels attracted to you and every book, conversation, and opportunity revolves around enabling more of the life you're living. “They are immediate, smack-you-in-the-face expressions of a person who feel simple enough that can click around in your hand, but cosmically large enough the person feels True in Themselves in some sense beyond literal.”

Aquant’s CEO and Cofounder Shahar Chen on how to scale a data center for success. “Adding technicians doesn’t automatically add proficiency. And in a data center, that gap matters because time-to-proficiency isn’t just a training metric—it’s a reliability metric.”

Groundcover’s CEO Shahar Azulay exposes the hard truths in observability.

CruxOCM
Senior Software Developer - Platform (Vancouver)

Groundcover
Account Executive - East Coast (Boston)

Reco
Team Lead (Tel Aviv)