The Boring Businesses That Will Dominate the AI Era

Tina He is back at Every. The writer and investor—now at Pace Capital—explored what happens when AI agents become your primary users last spring. Now she’s going deeper: In a world where AI picks vendors based on ruthless logic, she identifies five kinds of businesses that become irreplaceable. These aren’t the companies with the best models or the slickest interfaces—they’re the ones that own the boring yet essential infrastructure AI must flow through but cannot replace. If you’re building for a future where your customer is an algorithm, this is your map.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

At 2:17 a.m., your customer fires you and your customer relationship management software. The reason? Their sales rep calculated that your CRM’s $30,000 annual contract only delivers $12,000 in value. There was no meeting and no negotiation—because the sales rep was an AI agent.

Agents don’t need training like human users of software, and they don’t have loyalty. They evaluate economics in milliseconds and switch the moment it makes business sense, even in the middle of the night. It’s pure ruthlessness.

When AI models commoditize—when the latest GPT, Claude, and Gemini models are all making similar decisions based on similar capabilities—the competitive edge for companies shifts from having the best model to having the infrastructure between algorithmic decisions and real-world consequences. This is the infrastructure that AI can’t route around: the boring, essential systems that control access to data agents need, workflows they must execute, and regulations they can’t bypass. And the next few years will determine who owns this layer while everyone else is still building better AI agents.

This infrastructure breaks down into five archetypes. I’ll show you how to recognize which one fits your business, where there are opportunities to build outsize businesses, and how to build a competitive advantage before this window closes. If your edge is a model or a coat of user interface like an app or chatbot, it isn’t a moat. What endures is infrastructure AI must flow through, but cannot replace.

The new infrastructure for software

Before we get into the first archetype, there’s one shift to explain. Think of it like the early internet: Before companies could build Google or Amazon, we needed HTTP and TCP/IP—the invisible communications rules and protocols that let computers talk to each other. We’re in that moment again, but for AI agents. New protocols—AG-UI (agent-user interaction) for agents talking to users, Google’s A2A (agent-to-agent) for agents talking to each other, MCP (model context protocol) for agents using tools—have become the standard pipes and wires. The archetypes I’ll walk through next are maps to where businesses are built as this infrastructure solidifies.

When the AI itself becomes cheap and interchangeable, the edge becomes what you feed it and what you can do with its outputs—data it can’t find elsewhere, and actions it can’t take alone.

Knowledge compounders: Data that learns while you sleep

The first winners are what I call knowledge compounders. These companies control organized data that agents need and that improves continuously through real-world usage, and requires years of operations to replicate. This data isn’t widely available, so most AI agents wouldn’t have it (and it hasn’t been used to train the major models).

Every time a player makes a strategic move in a video game or a doctor confirms that a diagnosis made by AI is indeed correct, the dataset grows stronger. That learning stacks up in ways that competitors can’t shortcut through capital alone (similar to the concept of compound engineering).

The best companies in this archetype create environments where customers unknowingly generate training data through usage. Medal is a platform where gamers can record and share clips of gameplay. Medal’s community creates over 2 million gaming clips per day—more than 700 million clips per year—each rich with signals about player behavior, reaction timing, and strategic decision-making. At typical data-collection and labeling costs, building an equivalent dataset from scratch would require hundreds of millions of dollars, yet Medal’s users generate it organically in their quest for recognition.

Another example of a company building an edge in data is Mercor, which provides feedback from experts such as lawyers and doctors to AI labs to help improve their models. Even if each generation of frontier models is getting better, in domains such as medicine, law, or finance, the cost of errors is high enough that “good enough” AI isn’t acceptable, and ongoing human judgment is essential. Mercor is betting that demand for human-n-the-loop verification persists even as raw capabilities improve.

Knowledge compounders aggregate human judgment at scale, like gamers expressing preferences or doctors confirming diagnoses. When AI agents need to check if their outputs match reality, not just sound plausible, they have to consult these datasets. The alternative is hallucination.

Examples: Explorium (B2B data infrastructure), Mercor (human-in-the-loop verification), Medal.gg (user-generated training data)

The workflow commons: Templates that capture how work happens

AI agents don’t have eyes. They don’t care about whether or not your software interface is easy to use or aesthetically pleasing. The winners in this category are those that build headless architecture—software designed for direct machine-to-machine connections rather than human interfaces—and build shareable workflows that capture which tools to connect, in what sequence, with what parameters.

n8n is a Berlin-based workflow automation platform that illustrates how this model can scale. The platform hosts more than 7,000 community-built workflow templates, each one a shareable blueprint capturing what worked in production. A marketing team in São Paulo can import a template that a growth engineer in New York refined through dozens of iterations. The template works because it survived real-world usage.

Between early 2025 and the fall, n8n’s valuation climbed from roughly $300 million to $2.5 billion. Major companies including Vodafone, food delivery giant Delivery Hero, and Microsoft rely on the platform to orchestrate AI‑powered operations.

The inflection was not driven by a better interface; after all, agents, not humans, are becoming the key users of software in general. The shift came when the accumulated workflows became the foundation for AI agent coordination. When someone builds an n8n workflow connecting a CRM to email to Slack—specifying “if deal closes, notify team; if any step fails, retry twice”—they capture institutional knowledge about how work happens. That’s thousands of templates and hundreds of integrations that an agent can’t reverse-engineer.

This is the flywheel: Users contribute workflows. The platform learns what sequences closed deals, saved hours, or failed. More users arrive for those proven patterns. If you leave, you’re abandoning a shared library that keeps getting better.

As AI models become interchangeable, the layer that stitches them together becomes more valuable. This layer pulls from any provider, plugs into any tool, pulls from any data source, and handles everything models don’t: tracking what happened, managing failures, and recovering from errors.

The success of the transition from a human workflow builder to an agent coordination layer will be the decisive factor in how big these companies can get.

Examples: n8n (enterprise workflow orchestration), ComfyUI (generative AI workflows), Roboflow (robotics workflow)

Reality’s gatekeepers: The toll booths between AI and consequences

The third archetype exists in areas where “move fast and break things” is a fireable offense: regulatory approval, banking relationships, compliance infrastructure. When an AI agent wires money to the wrong account, you can’t iterate your way out of it. When it submits a fraudulent insurance claim, you can’t A/B test the regulatory response.

This is the layer where silicon logic collides with real-world consequences. Companies that own these “bridges” don’t compete on technical sophistication, but on the boring yet essential guarantee that the wire transfer has arrived correctly in a customer’s bank account.

The business model is toll-booth economics: small fees on massive volume that compounds to billions. Stripe, for example, takes roughly 2.9 percent plus 30 cents per transaction. Plaid charges every time an app connects to a bank account. Deel collects a fee on every international contractor payment it processes. The dull is what drives the returns. The alternative to paying the toll is building your own bridge, which could take five years and hundreds of millions annually for compliance—and you still need regulatory approval to touch the payment rails.

The Synapse bankruptcy showed what happens when you don’t own the infrastructure. Synapse sat between fintech apps and actual banks, handling transactions so startups didn’t have to build direct banking relationships. Mercury, a business banking platform serving startups, was one of its biggest clients. When Synapse collapsed in 2024 after missing funds and discrepancies in ledgers were discovered, roughly $85 million in customer funds got stuck in limbo. Nobody could tell whose money was where. Hundreds of thousands of people lost access to their savings for weeks; some still haven’t recovered their funds.

Mercury survived because it had already begun migrating to Column, a bank built to serve fintechs directly—no middleman or ambiguity about who holds the money. Trust, once earned, hardens into rebar, painful to rip out and expensive to rebuild.

Examples: Stripe and Column (financial rails), Deel and Bridge (international compliance), Plaid (financial data access)

The marketplace: Trading floors for algorithmic buyers

Traditional marketplaces like Amazon or Airbnb are built for eyeballs. They rely on visual catalogs, reviews, and human intuition. But AI agents don’t have eyes, and they don’t browse. They connect through code, and they negotiate on price and terms in milliseconds.

When a buying agent from a logistics company meets a selling agent from a supplier, it needs a standardized way (such as the agentic commerce protocol developed by Shopify and Stripe) to discover the price, verify delivery, and arbitrate disputes. This creates a new kind of marketplace—not a visual catalog, but a high-speed trading floor for AI services.

Initially, companies in this archetype act as “traffic directors.” An agent needs to complete a complex task, and the platform routes that request across dozens of models and data providers, each with different speeds, prices, and reliability. The platform solves for the optimal route. In the example of an agent drafting a contract, it runs compliance checks and pulls recent case law: “Send the reasoning to Anthropic’s Claude Opus 4.5 for legal reasoning, the code execution to a specialized Llama instance for executing the compliance, and the search query about recent rulings to Exa.” Without the platform, a developer stitches this together manually.

As usage scales, the traffic director becomes a marketplace. Not Airbnb, where you browse and choose—more like Uber, where the platform picks based on data you can’t see. Every routed request teaches the platform something: which models actually perform, not just which ones claim to. If model A advertises 99 percent accuracy but fails 15 percent of legal queries, only the platform knows. That knowledges gives it leverage over pricing, traffic, and terms. Buyers stay for reliability guarantees; sellers stay for the volume.

However, this archetype carries the highest risk. It bets entirely on fragmentation.

If foundation models converge on similar capabilities and costs, the value of a routing marketplace collapses. But if the AI landscape stays messy—with models excelling at different tasks, quality varying by use case—then the marketplace becomes essential infrastructure.

If fragmentation prevails, these platforms become the universal translators of value. They won’t just route traffic; they will force every buying and selling agent to speak their specific dialect of trade.

Examples: OpenRouter (multi-provider LLM routing), Composio (tool integration abstraction), Context7 (default dev tool stack)

Vertical transformation: Replace the workflow, not the worker

Why optimize a job when you can eliminate it entirely?

Companies in this category pick industries with high labor costs, complex knowledge requirements, and regulatory barriers—then build AI that handles the entire workflow end-to-end. It’s the same quality at one-tenth the cost. They own the AI, the data, the tools, and the compliance—the entire operation from start to finish.

The economics are dramatic. McKinsey charges $2 million for a three-month strategic review involving eight consultants. An AI system might deliver comparable output in 48 hours for $20,000. Traditional law firms dedicate 65 percent of revenue to salaries; AI-driven solutions target human oversight at 5 to 10 percent of costs. A $20,000 two-week legal review can now be executed for $2,000 in two hours.

But undercutting on price isn’t enough to win. A simple app built on top of ChatGPT can undercut you on price but has no proprietary moat. By owning the entire workflow—the domain knowledge, the compliance infrastructure, the liability—companies create a lasting advantage.

Legal tech company Harvey is a good example of how this works: Lawyers work alongside its engineers, so the product includes insider knowledge specific to how law firms work that generic models can’t infer. These kinds of businesses are defensible because they can prove exactly which expert reviewed which output, when, and why—a paper trail that a weekend prototype can’t provide and a regulatory prerequisite in many of these industries.

While for now, Harvey’s contract limits how much they are on the hook for in the case of a malpractice claim, and law firms must accept the risk of there being errors in any AI output. In the future, however, a vendor like Harvey might be able to absorb this risk into its pricing. Finally, these companies have the essential elements large companies look for when procuring software, such as secure login systems, compliance with data storage rules that are unique to specific countries, and security certifications.

The companies that prove this in a vertical have hard-to-beat advantages—accumulated data, customer trust, and regulatory relationships. But “wins permanently” overstates it. They win time: a head start that matters if they keep executing and evaporates if they don’t.

The test will come when a major firm faces a malpractice claim involving AI-assisted work. Whoever can demonstrate a defensible process will dominate the next decade of enterprise AI in regulated industries.

Examples: Harvey (legal), Sierra (customer success), Rogo (financial research)

The common thread: Each archetype compounds

The most interesting combinations haven’t emerged yet. A company accumulates proprietary data, lets users contribute to this data, then becomes the exchange where that data is priced and traded. A workflow platform adds compliance capabilities such as audit trails, using built-up know-how to navigate regulations automatically.

Beneath the combinations lies a deeper pattern: Every archetype improves the more it’s used by humans and agents. This is fundamentally different from the software model we’ve known for decades. Traditional SaaS companies shipped updates weekly, maybe monthly. These new businesses ship the latest version every second. The product your customer uses at 3 p.m. is measurably better than the one they used at 9 a.m. because customer interaction data creates new data, refines a workflow, or strengthens a compliance record.

This gives them an advantage over the AI labs themselves. OpenAI can train a better model in six months. These companies acquire something harder to copy: institutional memory from millions of real-world transactions. A better model can be trained with enough compute. Institutional memory requires time, customers, and stakes—things money can’t shortcut.

One divergence worth noting: Some companies win through horizontal breadth, others through vertical depth. n8n’s workflows apply across industries—marketing, operations, engineering, and finance. Harvey’s value is concentrated in a single domain, where regulatory complexity and liability create natural barriers. Both approaches build on themselves, but they build different things: One accumulates integration knowledge across use cases; the other accumulates domain expertise within a single use case. The horizontal play bets on ubiquity; the vertical play bets on irreplaceability.

The model alone is rarely the moat.

At 2:17 a.m., your customer’s AI agent will evaluate your value. The winners aren’t building better models or stickier interfaces. They’re building whatever takes the longest to rebuild.

Tina He is a writer and investor at Pace Capital interested in systems of behaviors and beliefs enabled by new markets, networks, and technologies. She led developer tools at Base after being the cofounder and CEO of Station Labs, which Coinbase acquired in 2024. To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.

For sponsorship opportunities, reach out to [email protected].

Help us scale the only subscription you need to stay at the edge of AI. Explore open roles at Every.