What Companies Won't Automate: Where AI Value Actually Accrues

TLDR: The companies raising billions to build smarter AI models may be solving the wrong problem. The question isn't just whether AI can automate a task, it's who captures the value when it does, and which functions companies will actually allow to be commoditized. Whether AI efficiency gains come from internalized or outsourced processes will depend on whether the task is a competitive differentiator (core) or a variance-insensitive function (supporting). This distinction will largely determine where economic value accrues and the timelines of automation. For investors, supporting functions represent platform opportunities while core functions will be reliant on services.

Distinguishing Core from supporting function

There is a running assumption that if a task can be automated, someone will build a company to do it. However, this misses the more important question: not of what AI can replace, but who captures the real value of that automation, and what functions companies will be willing to allow to be commoditized.

Company operations can be divided into two categories. Core operations are those where incremental differences in performance translate directly into competitive outcomes. Supporting operations are threshold-based: in opposition to core operations, once a minimum standard is met, additional improvement captures little incremental value.

In Practice

Core functions of a real estate firm include asset selection, underwriting assumptions, and operational improvements, areas where small differences can materially impact the defining function of the firm, providing returns to investors. Supporting functions include areas like compliance, reporting, and fund administration. If multiple funds all share the same compliance, reporting, and fund administration platforms, the respective relative returns will have no persistent differentiation beyond baseline. These company functions are threshold-based: once the bar has been cleared, further improvement has little to no impact on returns.

By contrast, if multiple firms outsourced their core functions to the same firm, expected returns would converge. Each firm would identify the same assets, underwrite them identically, submit the same bids, and ultimately produce the same return on the assets. This would turn investment performance into a commodity differentiated predominantly by fees, setting up funds for fee compression.

Scale, Decision Layers, and Commoditization

Importantly, shared tooling does not eliminate all differentiation; it shifts it. Non-automatable differentiators like scale can preserve differentiation by lowering unit costs or improving deal access even when shared tooling across all firms causes decisions to converge. This shift underscores a core claim: once operations are commoditized, competition migrates away from judgment toward structural advantages.

In practice, commoditization is rarely all or nothing. Firms will often outsource shared primitives while maintaining a proprietary layer on top of them. For companies like Optum's AI prior authorization tool, it only serves to automate approvals and reviews but refuses to automate denials, automating the plumbing, not the judgment layer. Importantly, this convergence will occur not from a shared infrastructure layer, but at the decision layer. As long as firms retain control over how information is weighted and acted upon, variance will persist. The relevant distinction is not shared technology, but shared judgment. For firms that outsource too aggressively due to short term cost pressures, over time they will be penalized through commoditization and reinforcing separation between internalized core automation and externalized threshold automations.

When Platforms Work

The distinction between core and supporting functions predicts where AI adoption will occur first. It will occur externally where variance in performance beyond the bar does not matter: compliance, HR, and internal reporting. HR is already moving in this direction. Workday's Paradox is targeting this category in interview scheduling, an area where differentiation has no impact. In contrast, AI applied to core functions like capital allocation, pricing, or strategy is more likely to be developed or heavily customized by internal teams, not because external systems are incapable, but because sharing them risks commoditization. This dynamic is unusually visible in investment firms where returns are directly comparable, but the same logic applies across markets.

A friend was automating debt research for investing firms but found themselves running rapidly into the issue of each firm demanding customized workflows and outputs, not willing to default to a standard option. Similarly, another major financial institution is refusing to utilize any outside platforms for all of their key AI initiative. This is in part due to heavy customization requirements, but also because the organization wants to utilize their proprietary data, not a standard solution. However, the same company is actively utilizing a variety of platforms to help support employees in the mundane, unimportant tasks. This purchasing dynamic is reflected broadly, platform companies that are thriving are serving these supporting functions. Additionally, the demand of many supporting functions (compliance, HR, and internal reporting) can be pooled across industry. While there may appear to be a risk of non differentiation, switching costs away from the platform, standardization, and trust can still drive meaningful monopoly power.

Internal Development and FDEs

By contrast, core functions are being automated internally or through the use of FDEs. Companies like Applied Compute and Palantir have thrived not by providing a general platform but instead by offering companies a better way to unlock data. The FDE model works for core automation because it solves the commoditization problem without requiring companies to build everything themselves. Building internal AI capabilities from scratch is expensive, slow, and requires maintaining specialized teams that most firms don't have. But sharing a platform with competitors creates the convergence risk outlined earlier. FDEs sit in the middle: they provide the expertise and infrastructure to automate core functions while keeping the resulting systems proprietary to each client.

This in-house ownership and customization allows for the gains from automation to be realized without the risk of commoditization. A real estate firm using FDEs to automate underwriting isn't just getting a faster version of a standard model. They're getting a system trained on their historical deals, calibrated to their risk preferences, and integrated with their proprietary data sources. The firm next door using the same FDE provider gets a fundamentally different system, trained on different data and optimized for different strategies. The decisions don't converge because the systems aren't shared, they're bespoke.

Meanwhile, each deployment requires integrating with proprietary data, custom feature engineering, and ongoing refinement all of which serve to create high switching costs and long lasting customers. Once an FDE has spent months embedding itself into a firm's operations, understanding its data infrastructure, and training models on its proprietary information, switching providers means starting over. This creates the kind of lock-in that justifies premium pricing.

This is visible across verticals. Goldman Sachs isn't buying off-the-shelf AI tools for trading or risk management, they're building proprietary systems internally and partnering with Anthropic FDEs for six months to provide capabilities they can't develop in-house. Harvey isn't selling access to a legal AI, they're embedding custom systems within individual law firms. In each case, the value proposition is the same: we'll help you automate a core function, but the resulting system is yours, not a shared resource.

Economically, companies relying on FDEs are trading breadth of offering for defensibility and higher contract values. A platform company might sign hundreds of customers in a year, each paying relatively modest subscription fees. An FDE-heavy company might close a dozen deals, but each contract runs into the millions and lasts for years. This creates slower adoption and tighter constraints as the companies scale, but it maintains the exclusivity required by companies for these most critical functions. You can't sell the same bespoke system to everyone, which means growth is limited by how many custom deployments you can support. But for functions where commoditization is unacceptable, that's a feature, not a bug.

The Trajectory

Over time, however, even core functions are likely to face automation and commoditization from outside platforms. It seems to me that this will start where variance matters the least. For instance, baseline mapping may converge across providers while differentiation becomes concentrated in non-commoditized dimensions like brand and distribution.

If this trajectory holds, AI adoption is unlikely to act as an equalizing force and instead a concentrator. Supporting automation favors scale, while gradual core automation will ultimately advantage firms with durable, non-automatable differentiators like capital access, talent, and scale. The value brought about by AI will therefore be captured not by what it is able to automate, but by the dimensions of competition firms are willing to commoditize.

For Investors

To distinguish whether a company is automating core vs. non-core functions, ask two questions: Would a 1-2% improvement in this function reliably change firm outcomes? And if two competitors used the same platform, would their decisions converge? The answers determine not just the business model, but the economics, growth trajectory, and competitive dynamics of the company automating it. Non-core companies win via unit economics, doing the needed thing cheaper and faster with minimal implementation overhead. Success looks like Stripe or Gusto: horizontal platforms serving thousands of customers with standardized workflows. The moat comes from scale economies, network effects, and operational excellence. These companies grow quickly because each new customer looks roughly like the last. Margins compress over time, but volume makes up for it.

Core automation looks different: fewer, higher-ticket contracts, longer sales cycles, and FDE-heavy delivery. Contract values are an order of magnitude higher, and once a system is embedded, switching costs are prohibitive. Success here looks less like a SaaS company and more like a high-end consultancy with software leverage. The distinction also predicts where pricing power persists. Non-core platforms face constant commoditization pressure. If your value prop is doing a threshold task efficiently, someone will always do it more efficiently. Core automation captures a share of the value it creates, allowing outcome-based pricing that's sustainable as long as gains remain exclusive. The critical distinction for investors isn't whether AI replaces human judgment, but whether that replacement is exclusive or common. Exclusive automation drives differentiation. Common automation drives convergence. A company selling compliance automation can run lean with high gross margins. A company selling core automation needs embedded teams and custom development cycles. Both can be valuable, but they're fundamentally different investments with different risk profiles, growth constraints, and exit paths.