The Last-Mile Problem of AI

Build vs Buy

The core value proposition of the enterprise software industry has always been a trade-off between complexity and convenience. Building your own custom software used to be incredibly expensive and risky, so you instead paid a monthly fee to a company like Salesforce or Intuit.

In exchange for that fee, they gave you a massive, standardized package of features designed to work for everyone. The catch, of course, is that you probably only used a tiny fraction of what you were paying for. You paid for the 100%, but you only really cared about the 10% that tracked your specific sales leads or managed your specific payroll. This arrangement still makes sense, because customers avoided the cost and complexity of a full engineering team, and SaaS companies built highly scalable, recurring revenue businesses.

But that "build versus buy" pendulum is now swinging back with a vengeance. The rise of AI coding tools is making software malleable. If a ten-year-old can vibe code a website for Secret Santa, a business manager can build a custom wrapper around a database that does exactly what their team needs for essentially zero marginal cost.

As an AI consultant recently explained during a Bloomberg podcast:

The whole job of building SaaS is you need to generalize problems and say you build things that are going to work for everybody...

[But what is starting to happen is that] inside these organizations, you can now solve very specific problems that are highly valuable. And not only can you solve them better than generic software, but you can actually, in a lot of ways, do it for less money because you're trying to tackle less stuff. You didn't need the 16 other features you bought it for, the one that you really, really cared about.

The reason that trade-off is shifting so quickly comes down to a change in the plumbing. For the last few years, AI could write code, but you still needed an engineer to actually run it. Getting a block of Python from ChatGPT was one thing; figuring out how to install the right libraries and execute it from the command line was another. It was the last-mile problem of AI development.

But newer tools, like Anthropic's Claude Code, are designed to close that gap. By operating with permission on a user's local machine, the model can not only generate the script but also handle the tedious work of setup and execution. The human's role is less "coder" and more "manager"—supervising an agent that is now both the architect and the construction crew.

The Database Wrapper

Now, this shift exposes the underlying architecture of much of the modern software industry. Many of the world's most valuable companies are essentially just very expensive interfaces sitting on top of a very simple database.

Take Customer Relationship Management (CRM) software. The primary job of a salesperson was to take unstructured data—the details of a lunch meeting or a phone call—and manually type it into a structured database so that the VP of Sales could later ask, "What is the status of this deal?"

Salesforce became a $200 billion company by being the world's most successful place to store that data. But AI is getting exceptionally good at that exact task. It can listen to a recorded meeting, transcribe it, and populate the database fields automatically. When the manual data entry that justified the software's existence is automated away, what is the core value proposition? The software is no longer a rigid product you buy; it is a sequence of logical thoughts you can generate on the fly.

From Talkers to Doers

This gets to the fundamental shift that is reordering the entire software market. In a recent essay, Sequoia Capital suggested that we are moving from AI "talkers" to AI "doers."

The AI applications of 2023 and 2024 were talkers. They were sophisticated conversationalists, great for drafting an email or answering a question. But the AI applications of 2026, Sequoia argues, will be "doers." They are what Sequoia calls "long-horizon agents," autonomous systems that can figure things out by forming hypotheses, testing them, and pivoting without being told what to do next.

This is the very capability that threatens the old SaaS model. You didn't buy Salesforce because you loved its interface; you bought it because you needed a human to do the work of entering data and running reports. The software was a tool for the human doer.

But when the software is the doer—when it can autonomously function as a cybersecurity analyst, or a legal associate—the value shifts from the standardized interface to the agent's ability to execute a task. This is the diffusion that Microsoft CEO Satya Nadella, speaking at Davos just yesterday, warned is necessary to prevent the AI boom from becoming a speculative bubble. For the current boom to be sustainable, he argued, its benefits must be "evenly spread" across industries, not just concentrated within big tech.

The promise of these doer agents is that they can finally deliver those benefits. A law firm can "hire" an agent from Harvey. A hospital can "hire" one from OpenEvidence. They are buying the work, not just the software.

This dynamic helps explain why many traditional software stocks have begun to look like melting ice cubes in the market. The vision of a world where you hire an agent for an hour instead of buying a software suite for a decade is a powerful one.

It rests, however, on two very expensive assumptions. The first is that these AI "doers" will become reliable enough to actually replace the human ones. The second is that the economics of the AI boom—the trillions in debt-fueled data center spending—can sustain itself long enough for these new digital employees to start earning their keep.

Which of those curves will bend first is a different question entirely.

More on Software:

• Software's YouTube Moment is Happening Now (a16z)

On Our Radar

Our Intelligence Desk connects the dots across functions—from GTM to Operations—and delivers intelligence tailored for specific roles. Learn more about our bespoke streams.

OpenAI Internalizes Energy Costs to De-Risk Infrastructure

• The Story: Facing political and community pressure over the immense power requirements of its Stargate project, OpenAI has committed to directly funding the associated energy infrastructure upgrades, such as new power generation and grid support, to ensure its data centers do not increase local consumer electricity prices. (Bloomberg)
• The Operations Implication: This decision fundamentally alters OpenAI's CapEx model for AI infrastructure, shifting the cost and project management of energy generation and grid reinforcement from public utilities onto its own balance sheet. The company must now develop a core competency in large-scale energy project development, including navigating permitting, construction, and long-term asset management, which dramatically increases the upfront capital and complexity required to deploy new compute capacity.

Alibaba Secures AI Power via Nuclear JV

• The Story: To secure the immense power required for its AI expansion, Alibaba has established an energy generation joint venture with China National Nuclear Power Co., signaling a strategic move to control its energy supply chain for data centers. (Bloomberg)
• The Operations Implication: This JV represents a critical de-risking of Alibaba's AI infrastructure supply chain. Instead of relying solely on the public grid and facing price volatility or power availability caps, Alibaba is securing a dedicated, baseload power source through a capital partnership. This strategy allows the company to lock in long-term energy costs and ensure operational continuity for its compute-intensive data centers, while using the JV structure to avoid the full burden of building and managing power generation assets itself.

P.S. Tracking these kinds of complex, cross-functional signals is what we do. If you have a specific intelligence challenge that goes beyond the headlines, get in touch to design your custom intelligence.

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.