• The Angle
  • Posts
  • The search for the AI-native firm

The search for the AI-native firm

The Angle Issue #327

The search for the AI-native firm

Here’s my entire thesis on the next 10 years: AI adoption is going to be fast. AI transformation is going to be slow.

Getting everyone in a company to chat with ChatGPT or Claude is one thing. Redesigning the organization around the assumption that AI can do much of the work is something else entirely. That will take new products, new business models, and new organizational structures.

Today’s AI-powered enterprise software winners may be growing at an extraordinary clip, but they are mostly just "SaaS 2.0." Whether horizontal (Lovable, Gamma) or vertical (Harvey and Legora, Rogo and Hebbia), these companies are largely the same as their SaaS predecessors. SaaS 1.0 products were opinionated interfaces on top of databases. SaaS 2.0 products are opinionated harnesses on top of models.

Of course, with frontier models under the hood, SaaS 2.0 products can do far more, and charge more per seat as a result. But they're still selling software per seat per month. And as the cost of building software collapses, customers are starting to push back on price and think differently about the age-old “build vs. buy” question. That doesn't mean these software companies disappear. But it does mean that the software itself is becoming less important.

As an aside, this is why I think the SaaSpocalypse is real. The standard pushback on vibe coding in the enterprise has been “nobody will rebuild Salesforce.” And that’s true. Because nobody would build that gigantic piece of garbage. Starting from scratch, you’d build the 1% that’s actually relevant to your business. And today’s frontier coding models can already do that! This is doubly true for SaaS 2.0, built on the same models its customers can now use directly. Software is just getting too easy to build.

Deep down, I suspect the companies know this, which is why they're increasingly trying to "charge for outcomes." But defining an outcome, verifying it, and pricing it is hard. A closed support ticket is easy. But what about a PRD? A marketing brief? A legal doc review? And once you've defined it, by the way, you've made it far easier for your customers to shop around. If everyone can produce roughly the same work, what stops a competitor from showing up and undercutting you?

So we have a technology that's increasingly capable of doing the work. But we're stuck trying to figure out how to sell it as software. Maybe the problem isn't how we're charging for the software. Maybe it's that we're still selling software at all?

Selling the work

This is the logic behind one of the biggest theses in venture right now, from Sarah Tavel's "Sell Work, Not Software" to Sequoia's "Services: The New Software": instead of selling software to the businesses doing the work, why not own the work itself?

One approach is the AI roll-up. Buy a bunch of services businesses. Deploy AI across them. Improve margins. Grow faster. Rinse and repeat. It makes sense. You're buying an established customer base, with existing revenue, and a working operation. Instead of convincing accounting firms to buy some new piece of accounting software, you just deploy it yourself and capture the upside.

But there are two problems. First, the capital. Roll-ups run on debt and equity, and that money comes with expectations. You have to service the debt and sell within five to seven years. If AI transformation really is slow, that's a short clock. Second, and more fundamentally, you're betting you can layer on enough AI, while keeping the lights on and the cash flowing, to meaningfully improve productivity. But what if taking full advantage of AI requires more than just cost-cutting and software deployment, but a full re-architecture of the firm itself?

Okay, fine. So don't buy a services company. Build one from scratch.

This is the AI-native services firm, another increasingly popular thesis, championed by Emergence. Pick a category where AI can do a meaningful share of the work. Build the company from the ground up around AI, and sell the actual work to customers rather than selling them the tools. As the old saying goes, “people don’t want to buy a quarter-inch drill, they want a quarter-inch hole!”

The advantage is that you get to design the entire organization from scratch. The problem, of course, is now you're running a services firm! You didn’t get into tech and raise gobs of venture capital to do that, did you? AI might make each employee enormously more productive, but that doesn't mean the rest of the business scales any better. You will still need to hire people to sell, manage customers, and handle edge cases. So if you succeed, your tech-enabled services firm may end up way more “services”-heavy than tech. And, as all your engineers up and leave (because they joined a tech company not a staffing firm), your technical edge dwindles as well.

So we're left with a bit of a conundrum. SaaS 2.0 can scale software, but may struggle to capture the value of the work. AI roll-ups can buy the work, but may struggle to fundamentally change how it's performed. And AI-native services firms can capture the value of the work, but may struggle to scale the organization.

Finding the bottleneck

I wonder if we're looking at this the wrong way. Maybe the more interesting question is: if you were building a brand new organization today to solve a particular problem, knowing what AI can do, how would you do it?

I'd start with three questions. The first is where the bottleneck is. This is the argument of The Goal, Eliyahu Goldratt's book on manufacturing and the theory of constraints. His core insight: a system's throughput is governed by its bottleneck. You can make everything else dramatically more efficient, but if you haven't addressed the bottleneck, you haven’t increased output.

Everyone is going to be talking about Goldratt and the theory of constraints for the next decade. I guarantee it. Because this is going to be a huge problem for AI deployment. Imagine you automate 90% of the administrative work at your newly acquired services firm, only to discover that actually the number of customers you can serve is still fundamentally constrained by the availability of licensed professionals on the team. You've made the firm more productive, kind of, but you haven't actually changed how fast it can grow. The question isn't how much work AI can automate, but whether it removes the constraint. And if it does, where does the bottleneck move next?

The second question is where the economies of scale are. Which parts of the business benefit from centralization, and which don't? In The Nature of the Firm, Ronald Coase famously argued that firms exist because coordinating activities inside an organization can be cheaper than coordinating them through the market. But AI is changing both the cost of producing work and the cost of coordinating it (as Alex Imas argued in his recent paper, Organizing Intelligence). Activities that once had to live inside the same firm might now be performed more efficiently by independent organizations (or AI agents) working together. In this imagined near-future, it’s not obvious to me what should get centralized and what should remain local and fragmented. The right structure should reflect where scale creates an advantage and where coordination is cheapest, but seeing as AI capabilities are a moving target, that structure isn’t at all clear right now. Regardless, it seems doubtful that the dominant strategy will be to reproduce the traditional boundaries of a software company or a services firm.

The third question is what compounds. Does adding customers improve the product? Does it generate data that makes production cheaper or better? Does it create distribution advantages, purchasing power, or room for further automation? Or are you just adding another customer and another fraction of an employee? I’d argue that a truly AI-native firm must find ways to compound.

When factories first adopted electricity, most simply swapped a large electric motor in for the steam engine and kept everything else. The same machines, driven by the same overhead line shafts, laid out the same way. It took decades before factories put a motor on each machine and arranged the floor around the flow of work instead of around the shaft. 

The lesson is obvious enough to slap us in the face: the opportunity afforded by electricity wasn't to replace the steam engine, it was to redesign the factory around what electricity made possible. We're making the same mistake with AI. We're putting AI into existing software, giving employees copilots, and automating individual tasks inside organizations designed around human labor. Even AI roll-ups and AI-native services firms are, in a sense, attempts to fit a new production technology into structures we already understand.

Ultimately, what we’re doing is mistaking adoption for transformation. And the biggest gains will not be from making existing organizations more productive, but from designing new kinds of organizations that couldn't have existed before.

I don't think anyone knows what those look like yet. But if you think you might be building one, I want to hear about it.

David Peterson

FROM THE BLOG

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.

WORTH READING

HARD TECH

On the road. Anthropic is targeting up to $64.4 billion in IPO proceeds, with a proposed November 19 Nasdaq debut. Its prospectus reports an approximately $81 billion annualized revenue run rate and nearly $900 billion in compute commitments. This could reset the entire valuation framework for private AI companies and test whether public markets will underwrite the capital intensity of frontier AI.

Jensen out. Nvidia's co-founder stepped down following reports concerning personal relationships and an alleged Chinese espionage connection. CFO Colette Kress became interim CEO. Nvidia has shed roughly $1.9 trillion in market capitalization since early October. Nvidia is not merely a semiconductor supplier; its financing and commercial relationships underpin much of the AI buildout. Leadership disruption adds uncertainty to the entire ecosystem.

HARD MARKETS

Going direct. Anthropic reportedly secured orders for approximately 1.7 GW of gas-fired generation equipment to support its data centers. This is another reminder that frontier AI companies are increasingly procuring their own generation rather than relying exclusively on utility capacity. The opportunity extends well beyond reactors and turbines to grid interconnection, power management, cooling, and flexible demand.

Autonomy in Ukraine. UForce secured a major financing - $700M at a $6B valuation - to expand autonomous systems production across air, land, and sea for Ukraine and allied militaries. Defense autonomy is attracting growth-stage financing on a scale previously reserved for major software companies. Manufacturing capacity, field-proven systems, and procurement access are becoming central competitive advantages.

NVIDIA competition heats up. Google introduced TPU 8tx, designed for confidential computing, and TPU 8i Max, designed for higher-capacity inference and large-context workloads. Google says its TPU business now serves nearly 100 customers. Nvidia's alternatives are becoming more credible, and hardware specialization for different AI workloads may change infrastructure economics.

HOW TO STARTUP

Data for sale. Mercor, originally an AI recruiting platform, has become a major provider of expert human input for training and evaluating frontier AI models. Its latest Series D, $750M at a $20B valuation, doubles its valuation from roughly a year ago. Human expertise remains a valuable input to frontier-model development. The company illustrates how scarce human judgment can become infrastructure for training and evaluating AI agents, rather than merely a labor-market service.

Flagship lands in the UK. Flagship is launching its first company-creation laboratory outside the US, planning to recruit approximately 50 scientific entrepreneurs and operators and attract at least £500 million of additional funding. This is an example of venture shifting toward systematically creating companies around AI-enabled production processes, rather than simply financing existing startups.

HOW TO VENTURE

The death of venture. Another great piece from Neumann, in which he argues that as venture funds grow larger, and they’re forced to either deploy more capital or make more investments and be less involved, they undermine venture’s fundamental advantage: making small, staged bets on genuinely uncertain markets where the upside is unknowable. He draws parallels to the 1980s venture boom and subsequent collapse, arguing that smaller funds willing to embrace genuine uncertainty will ultimately inherit the opportunity.

The McDonaldization of venture. Harrison argues that venture is increasingly dominated by massive asset managers whose incentives favor scale, efficiency, predictability, and consensus over original thinking and contrarian conviction. As capital concentrates in a handful of firms and AI companies, the industry increasingly optimizes for deploying enormous sums into widely accepted opportunities rather than discovering genuinely novel ones. His central concern isn't that this model can't generate returns, but that an industry designed to finance radically new futures is becoming structurally incapable of imagining them.

PORTFOLIO NEWS

Groundcover expands its eBPF sensor with deeper Kubernetes and infrastructure visibility.

Opdo raises $2.3M to automate optical engineering with AI.

PORTFOLIO JOBS

Reply

or to participate.