Vertical Agents Are the New Compound Startups: Why Owning the Whole Workflow Beats Selling a Tool

Table of Contents

• What "Compound Startup" Actually Meant
• Why Vertical Agents Revive the Model
• The Economics: Bundling Labor, Not Features
• The Compounding Loop That Makes It Work
• Where the Frame Breaks Down
• What This Means for Founders
• What This Means for Buyers
• Insights Most People Overlook
• References

What "Compound Startup" Actually Meant

The phrase got popular through Rippling, whose founder Parker Conrad argued that the conventional VC wisdom, do one thing, do it well, stay focused, was wrong for certain markets. His counter-thesis was that some companies should deliberately build several products at once, tied together by shared data, so each new product makes the others more valuable. Rippling's employee record powers payroll, which powers IT provisioning, which powers expense management, and so on. The compounding came from a shared spine, not from any single app.

That model demanded enormous capital and discipline, which is why it was rare. You had to build five things competently before any of them paid off, and most startups die trying to build one. Investors mostly treated it as an exception that proved the "stay focused" rule.

Here's the thing worth noticing: the compound-startup advantage was never really about having many products. It was about owning a unit of data or workflow that other functions had to route through. The products were just the toll booths. Conrad himself has been clear that the moat is the compounding employee graph, not the breadth of the catalog. Hold that idea, because it's exactly what vertical agents are rebuilding from a different direction.

Why Vertical Agents Revive the Model

A vertical agent doesn't bundle products. It bundles steps in a job. Take a medical-billing agent: it reads the clinical note, assigns the codes, checks them against payer rules, submits the claim, watches for the denial, and drafts the appeal. In the old world, that was a coding tool, a clearinghouse integration, a denial-management dashboard, and three human beings. The agent collapses all of it into one autonomous loop.

When you collapse a multi-step workflow into a single autonomous service, you accidentally recreate the compound-startup structure. The agent has to touch every system the work touches, the EHR, the payer portal, the practice-management software, so it ends up sitting on the same kind of shared data spine that Rippling built deliberately. The difference is that the agent earns that position by doing the work, not by selling you four apps and hoping you consolidate.

This is the real reason the analogy is sharper than it first looks. Horizontal AI tools (a generic copilot, a writing assistant) stay shallow on purpose so they can serve everyone. Vertical agents go deep into one industry's actual operations, and depth in one workflow naturally pulls in the adjacent workflows. A contract-review agent that's trusted to flag risk is one logical step from drafting redlines, then from managing the signature process, then from tracking obligations after signing. Each step compounds the value of the data the agent has already accumulated, which is the same flywheel covered in the cluster piece on proprietary workflow data as a moat.

The Economics: Bundling Labor, Not Features

The compound SaaS startup competed for software budget. It said: stop paying for five tools, pay me for one suite. The savings were real but bounded, software is a small line item compared to payroll.

Vertical agents compete for the labor budget, and that changes the math entirely. The total addressable market for a legal-research tool is the price of legal-research tools. The TAM for a legal-research agent is some fraction of what firms pay associates to do that research. Those are different universes. Analysts at firms like a16z have made the point that AI services companies can price against the cost of human labor rather than the cost of software, which expands the pie by an order of magnitude.

This is why per-outcome and per-task pricing keep showing up in this category, topics the GaaS cluster covers in depth under agent economics. If you're selling the result of work, charging per seat is leaving money on the table; you charge per claim adjudicated, per contract reviewed, per ticket resolved. And because you've bundled what used to be several tools plus the humans operating them, your "cost to displace" is much higher than a feature's. A buyer can drop a SaaS feature painlessly. Dropping the thing that does 40% of a department's throughput is a board-level decision.

The compounding shows up in margins too. A traditional services business has labor costs that scale linearly with revenue, double the clients, roughly double the headcount. An agent company's marginal cost per additional unit of work trends toward inference cost, which keeps falling. So a vertical agent can start with services-like gross margins and walk them up toward software margins as the model improves and the workflow data deepens. That migration, agencies and services firms turning themselves into agent companies to capture this curve, is its own subject in the cluster.

The Compounding Loop That Makes It Work

Strip away the jargon and the loop is simple:

1. The agent does real work in a narrow domain, which generates proprietary data about how that work actually gets done, the edge cases, the exceptions, the "we always do it this way for this payer" tribal knowledge.
2. That data makes the agent better at the work, which earns trust to handle more of the workflow.
3. Handling more of the workflow generates more proprietary data, and pulls the agent deeper into the customer's systems of record.
4. Deeper integration raises switching costs and widens the moat, which funds expansion into the next adjacent workflow.

Notice that each loop turn does two jobs at once: it improves the core product and it opens an expansion path. That dual payoff is the signature of a compound startup. Rippling got it from a shared data model it designed up front. Vertical agents get it from the data exhaust of doing the work, which is arguably a more durable source, because a competitor can copy your architecture but can't copy three years of your customers' exception-handling. McKinsey's research on AI value capture lands on a related conclusion: the durable returns accrue to companies that rewire how work happens rather than bolt AI onto existing processes, and rewiring the workflow is exactly what a vertical agent does by definition.

Where the Frame Breaks Down

I don't want to oversell the analogy, because it fails in instructive ways.

Compound SaaS chose its breadth; agents inherit theirs. Rippling decided which products to build and sequenced them. A vertical agent's expansion path is dictated by whatever the workflow happens to be adjacent to, which is often messier and more regulated than a clean product roadmap. A prior-authorization agent's "next product" might be fighting with insurers, which is a worse business than the one it started in.

Depth can become a trap. The compound startup's breadth was a hedge, if one product stalled, others carried it. A vertical agent that goes all-in on one industry's workflow has no such hedge. If a horizontal platform (or the underlying model vendor) absorbs that workflow as a feature, the agent company has nowhere to retreat. The cluster's piece on when a horizontal platform eats your vertical agent is the cautionary half of this story, and it's a real risk, not a hypothetical.

Outcome pricing cuts both ways. Selling outcomes means you eat the variance. If the agent's resolution rate dips, or a regulator changes the rules, your revenue and your liability move together in the wrong direction. Compound SaaS never carried that operational risk; it sold tools and let the customer own the outcome. Vertical agents in regulated fields, healthcare, legal, insurance, are signing up for a liability surface that pure software companies never had to think about.

The "compound" can be shallow. Some vertical agents bundle steps that look impressive in a demo but don't actually share a data spine. If the contract-review step and the obligation-tracking step don't make each other smarter, you don't have a compound startup, you have a feature checklist with an agent skin. The compounding has to be real, not narrated.

What This Means for Founders

If you're building in this space, the compound-startup lens suggests a few concrete moves.

Pick a wedge workflow where doing the work generates asymmetric data, information a competitor can't easily buy or scrape. Generic document summarization doesn't qualify; the way a specific specialty handles a specific payer's denials does. Your wedge should be the narrowest task that still throws off proprietary exhaust.

Sequence expansion along the data, not along the org chart. The temptation is to expand into whatever the buyer asks for next. The discipline is to expand into whatever workflow makes your existing data more valuable, even if it's a less obvious sell. That's how each new step compounds rather than dilutes.

Get to a system-of-record position deliberately, not accidentally. The agent that merely reads from the customer's systems is replaceable. The agent that the customer's systems now read from, the one that holds the authoritative record of what was done, has the Rippling-style spine. That transition is the whole game, and the cluster's piece on the system-of-record advantage walks through how to engineer it.

What This Means for Buyers

For buyers, the frame is a useful BS detector. When a vendor pitches a vertical agent, ask whether the steps it bundles actually compound. Does handling step three make it better at step one? Or is it four loosely stapled features that happen to share a login?

Also weigh the switching cost honestly, in both directions. Deep integration is the vendor's moat, but it's your lock-in. The same depth that makes the agent valuable makes it painful to leave, and outcome-based pricing means your costs scale with your own volume in ways a flat SaaS contract didn't. That's not a reason to avoid these tools, it's a reason to negotiate data portability and exit terms up front, while you still have leverage. The build-vs-buy calculus the cluster covers elsewhere is the right framework to run before signing.

Insights Most People Overlook

The compound advantage now comes from operating, not architecting. The original compound startups had to be brilliant up front, design the perfect shared data model, then build five products against it. Vertical agents get the compounding as a byproduct of operating, because the data spine accretes automatically from doing the work. This is a lower-IQ, higher-stamina path to the same moat, and it favors founders with deep domain obsession over founders with slick platform vision.

Most "vertical agents" are services businesses cosplaying as software. Plenty of companies in this category are, today, humans-in-a-trenchcoat: people doing the work with AI assistance, billed as if it were autonomous. That's not a scam, it's often the correct starting point, but the compound-startup payoff only arrives if the human share of each transaction trends toward zero. Investors and buyers should track the automation rate over time, not the demo, because a flat automation rate means you've bought a staffing agency with a nicer dashboard.

The model vendors are the real compound startups, and vertical agents are partly their distribution. Every vertical agent runs on someone's frontier model. As those models get cheaper and more capable, the agent's moat shifts away from "we have the smartest model" toward "we have the proprietary workflow data and the integrations." Founders who don't internalize this build on borrowed differentiation. The defensibility has to live in the data and the depth, because the intelligence layer is a rented commodity that improves on someone else's roadmap.

Regulation is a feature, not a bug, for the compound thesis. Conventional wisdom treats regulated verticals as harder. But regulation creates exactly the kind of durable, hard-to-replicate workflow complexity that makes a data spine valuable. An agent that has encoded three years of a specific regulator's actual enforcement behavior has a moat a deregulated competitor could never build. The most defensible vertical agents may be the ones in the most annoying industries.

Compounding can run in reverse. Everyone models the flywheel spinning up. Few model it spinning down. If a vertical agent's automation rate stalls and a horizontal platform commoditizes its core step, the same depth that was a moat becomes dead weight, high integration cost, narrow market, no hedge. The compound structure amplifies outcomes in both directions, which makes these companies higher-variance bets than the steady SaaS they superficially resemble.

References

• Rippling, The Compound Startup
• Andreessen Horowitz, AI Will Bring Down the Cost of Services
• McKinsey, The State of AI