Why Vertical Agents Beat Horizontal Platforms (And When They Don't)

Table of Contents

• The core claim, stated plainly
• Why depth beats breadth in agentic work
• The reliability math nobody puts on a slide
• Economics: who actually captures the value
• The moat question: where defensibility really lives
• When vertical agents lose
• A decision framework for builders and buyers
• Insights Most People Overlook
• Frequently Asked Questions
• Conclusion
• References

The core claim, stated plainly

A horizontal agent platform promises to automate "any workflow." A vertical agent promises to do one job, review a commercial lease, code a prior-authorization request, reconcile a month-end close, and do it the way that industry actually does it. On paper, horizontal sounds like the bigger business. In practice, the buyers who write the biggest checks for agentic AI-as-a-service keep choosing the narrow tool.

The reason is simple and a little unglamorous: the model is the easy 20%. The same frontier model sits underneath the legal agent, the healthcare agent, and the generic "build-your-own" platform. What separates a product that gets renewed from a demo that impresses once is the other 80%, the domain rules, the edge cases, the integrations into a system of record, the audit trail a compliance officer will accept. That work is irreducibly specific. You can't abstract it into a general platform without throwing away the exact thing the buyer is paying for.

This piece is one node in a larger map of the vertical-agent landscape. If you've read our coverage of legal contract-review agents or prior-authorization agents, you've already seen the pattern up close. Here we zoom out to the strategic question underneath all of them: when does narrow win, and when does it lose?

Why depth beats breadth in agentic work

Software has had this fight before. The horizontal-versus-vertical tension goes back decades, Salesforce versus a dental-practice CRM, SAP versus a restaurant POS. The general lesson held: vertical software wins when the workflow is high-stakes and idiosyncratic, horizontal wins when the workflow is universal and commoditized. Agentic AI inherits that lesson and sharpens it.

Sharpens it, because agents act. A horizontal SaaS tool that's 80% configured to your industry is annoying but survivable, your team fills the gaps manually. An agent that's 80% configured to your industry is dangerous. It will confidently take the wrong action on the 20% it doesn't understand, and now you have a malformed insurance claim, a misfiled SEC document, or a patient note that says the wrong thing. The cost of the gap is no longer "extra clicks." It's liability.

So the depth a vertical agent carries, knowing that in Delaware a particular indemnification clause means X, that this payer rejects prior-auths missing a specific modifier, that GAAP treats this lease one way and the IRS another, isn't a nice-to-have. It's the product. McKinsey's research on generative AI's economic potential makes a related point indirectly: the largest value pools cluster in specific functions like sales, software engineering, and customer operations, where deep workflow integration, not raw capability, determines whether the technology pays off (McKinsey's "The economic potential of generative AI").

Domain expertise is the last mile

There's a phrase worth borrowing from logistics: the last mile is where most of the cost and most of the failure live. For vertical agents, the last mile is domain expertise encoded as behavior, not just "the model knows about contracts" but "the agent follows the exact review checklist a senior associate would, flags the three clauses that actually get litigated, and routes anything ambiguous to a human." Horizontal platforms hand you a toolkit and tell you to build that last mile yourself. Most buyers can't, and don't want to.

The reliability math nobody puts on a slide

Here's the part that gets glossed over in pitch decks. Agentic workflows chain steps. If a single step is 95% reliable, a five-step chain is roughly 77% reliable (0.95^5). A ten-step chain drops to about 60%. For a workflow you're billing on a per-outcome basis, 60% reliability isn't a product, it's a refund machine.

Vertical agents win the reliability fight because narrowness lets you do things a horizontal platform structurally can't:

• Constrain the action space. A legal-review agent never needs to book a flight or query a CRM. Fewer possible actions means fewer ways to go wrong, and tighter guardrails.
• Hardcode the known edge cases. When you only do one job, you can enumerate the failure modes a domain expert already knows and handle each one explicitly.
• Validate against a ground truth. A vertical team can build evals from real, labeled industry data, past contracts, adjudicated claims, completed tax returns, because they have access to it. A horizontal platform has no single ground truth to test against.

This is why the conversation around agent reliability is really a conversation about scope. Anthropic's own guidance on building effective agents lands in the same place: the most robust systems favor narrow, well-defined tasks with tight feedback loops over sprawling open-ended autonomy (Anthropic's "Building effective agents"). Narrow scope isn't a limitation a vertical agent tolerates. It's the source of its reliability advantage.

Economics: who actually captures the value

Pricing is where the vertical advantage turns into money. A horizontal platform almost always prices on consumption, seats, tokens, tasks, compute. That's a thin, commoditized layer, and it gets squeezed every time model costs fall. A vertical agent can price on outcome, because it knows what the outcome is worth in its industry.

Consider the spread. A prior-authorization that a health system would otherwise pay a staffer 30 minutes to process has a known, defensible dollar value. A contract review that replaces two hours of associate time bills against a known hourly rate. The vertical agent can capture a slice of that value, say, a fixed fee per completed prior-auth or per reviewed contract, that has nothing to do with how many tokens it burned. This is the per-outcome pricing model that the broader GaaS category keeps circling back to, and verticals are where it actually works, because only a vertical knows the value of a single unit of output.

Andreessen Horowitz has argued that AI is enabling startups to attack the services economy directly, going after the labor budget, not just the software budget (a16z's "AI's economics will reshape the services business"). That reframing matters here. Horizontal platforms compete for a share of the IT software line item. Vertical agents compete for a share of a much larger pool: the cost of the humans currently doing the work. A 5% take of a payroll line dwarfs a generous take of a software line. That's the structural reason vertical economics look better.

The moat question: where defensibility really lives

If the model is a commodity and any team can call the same API, what stops a competitor, or the horizontal platform itself, from copying a vertical agent? Three things, in rough order of durability.

Proprietary workflow data. Every contract reviewed, every claim adjudicated, every close completed generates labeled data that makes the next run better. A horizontal platform sees a thin sliver of a thousand workflows; a vertical agent sees the full depth of one. Over time that depth compounds into evals, fine-tunes, and edge-case coverage a newcomer can't shortcut. We treat this as the central moat in the cluster, and it deserves its own deep dive.

Depth of integration. A vertical agent that's wired into the system of record, the EHR, the practice-management suite, the GL, is expensive to rip out. The deeper the integration, the higher the switching cost, and switching cost is a moat that has nothing to do with model quality.

System-of-record position. The strongest vertical agents don't just plug into the system of record; over time they become one. Once the agent is where the work lives and the data accumulates, displacing it means displacing the workflow itself.

None of these come from the model. All of them come from being narrow enough to go deep. That's the throughline.

When vertical agents lose

Now the honest part, because "vertical always wins" is a lazy take and buyers can smell it.

Vertical agents lose under four conditions:

1. The market is too small. Depth is expensive to build. If the total addressable spend in your niche can't support the cost of encoding all that domain expertise and maintaining it as regulations change, the vertical play starves. A horizontal platform with a thin configuration layer may be the only economically viable way to serve a fragmented long tail of small markets.

2. The workflow is genuinely generic. Some tasks really are the same everywhere, scheduling a meeting, summarizing a document, drafting a generic email. There's no proprietary last mile to own, so a horizontal agent does it just as well and amortizes its R&D across every industry. Building a "vertical" wrapper around a commodity task is how startups die.

3. A horizontal incumbent owns the system of record. This is the scary one. If Microsoft, Salesforce, or ServiceNow already sits on the system of record your vertical agent needs, they can ship a "good enough" agent natively and bundle it for free. Your superior depth may not survive being a paid add-on next to a free default. The vertical agent that gets eaten by a horizontal platform usually gets eaten here.

4. The frontier model absorbs the depth. As base models get better at reasoning over long context and tool use, some of the "depth" a vertical agent hardcodes today becomes something the model just does tomorrow. The defensible verticals are the ones where the moat is data and integration, not clever prompting, because prompting gets commoditized on the next model release.

The pattern under all four: vertical loses when there's no durable, proprietary last mile to own. Narrowness without depth is just a smaller horizontal platform, and that's the worst of both worlds.

A decision framework for builders and buyers

If you're building, ask three questions before committing to a vertical wedge. Is the workflow high-stakes and idiosyncratic enough that "80% right" is unacceptable? Is there proprietary data or a system-of-record position you can accumulate that a horizontal player can't easily replicate? Is the value of a single completed outcome large and legible enough to price against? Three yeses is a strong vertical thesis. Two or fewer, and you should think hard about whether you're really building a feature that a horizontal platform, or a frontier model, will absorb.

If you're buying, invert it. For your highest-stakes, most-regulated, most-specific workflows, prefer the vertical agent that speaks your industry's language and will sign up for outcome-based terms. For generic, low-stakes, cross-functional tasks, a horizontal platform is cheaper and more flexible, and you lose nothing by going broad. Most organizations will run a portfolio, a horizontal platform for the long tail, a handful of deep vertical agents for the workflows that actually matter. That mixed posture isn't indecision. It's the correct answer.

Insights Most People Overlook

• "Vertical" is a claim about the buyer, not the model. Two products can call the identical API and one is a vertical agent while the other is a horizontal platform, the difference is who they're sold to and how much of the buyer's specific workflow they own. Founders who think the moat is in the model are optimizing the commodity layer.

• Horizontal platforms are often the best go-to-market for verticals, temporarily. Plenty of successful vertical agents started life on a horizontal framework, then peeled off once they had enough proprietary data and integration to justify owning the stack. Starting horizontal to find a wedge, then going vertical to defend it, is an underrated sequence.

• The most dangerous competitor isn't another startup, it's the model provider. Every frontier release quietly eats a layer of "depth" that vertical agents used to hardcode. The verticals that survive are the ones whose moat is proprietary data and deep integration, things a better model doesn't replicate. If your edge is a clever system prompt, you're renting your moat from a lab that can revoke it next quarter.

• Per-outcome pricing is a vertical-only privilege. Horizontal platforms literally cannot price on outcome because they don't know what your outcome is worth. The ability to bill per resolved ticket, per reviewed contract, or per processed claim is downstream of being narrow enough to know the unit economics of one industry. Pricing power is a symptom of verticality, not a separate strategy.

• The services-to-software flip changes who your competitor is. When a vertical agent attacks a labor budget instead of a software budget, the incumbent it's displacing is a staffing firm, a BPO, or an in-house team, not another SaaS vendor. That reframes everything: pricing, sales motion, and the size of the prize.

Frequently Asked Questions

Is a vertical agent just a horizontal platform with a niche prompt? No, and assuming so is the most common founder mistake. A real vertical agent owns proprietary workflow data, deep system-of-record integrations, encoded domain rules, and outcome-based pricing. A niche prompt on a generic platform has none of those and gets commoditized the moment a competitor, or the base model, catches up.

Can a horizontal platform ever be the right buy for a regulated industry? Sometimes, for the generic slices, internal document search, meeting summaries, first-draft emails. But for the regulated core (adjudication, clinical documentation, filing), the audit trail and edge-case handling a vertical agent provides usually outweighs horizontal flexibility. Match the tool to the stakes of each workflow, not to a single org-wide standard.

What happens to vertical agents as frontier models get smarter? The commodity layer (reasoning, language, basic tool use) keeps improving and lifts everyone equally. The defensible part, proprietary data, integrations, system-of-record position, doesn't get absorbed by a smarter model. So better models actually raise the floor for good vertical agents while wiping out the ones whose only edge was prompt cleverness.

How small is too small for a vertical play? There's no universal number, but the test is whether the total spend in the niche can fund the ongoing cost of encoding and maintaining domain depth as rules change. Regulated industries shift constantly, and that maintenance is real. If a thin horizontal configuration could serve the niche acceptably, the market is probably too small or too generic for a defensible vertical.

Why can vertical agents charge per outcome when horizontal platforms charge per token? Because pricing power follows knowledge of value. A vertical agent knows a completed prior-auth saves 30 minutes of staff time and can price against that. A horizontal platform doesn't know what any given task is worth to you, so it falls back to metering consumption, the thinnest, most commoditized way to charge.

Do vertical agents have to build their own models? Almost never. The winning move is to use frontier models as a commodity input and invest the differentiation budget in data, integrations, evals, and workflow depth. Building a proprietary model is expensive and rarely the moat; owning the proprietary workflow data is.

Conclusion

Vertical agents beat horizontal platforms because agentic value lives in the last mile, the specific, high-stakes, regulated workflow of a single industry, and that last mile can't be abstracted into a general platform without discarding the exact thing buyers pay for. Depth buys reliability through a constrained action space, captures more value through outcome-based pricing, and compounds into moats built on proprietary data, deep integration, and system-of-record position. None of that comes from the model; all of it comes from being narrow enough to go deep.

But narrow is not automatically a winner. When the market is too small, the workflow is genuinely generic, a horizontal incumbent already owns the system of record, or the frontier model absorbs the depth, the vertical thesis collapses, and what's left is just a smaller, weaker horizontal platform. The builders and buyers who win the agentic AI-as-a-service era won't pick a side dogmatically. They'll match depth to stakes: vertical agents where the work is specific and the cost of being wrong is high, horizontal platforms where it isn't. Everything else in this cluster, the moats, the pricing, the regulated-industry playbooks, is a more detailed answer to that same question.

References

• McKinsey & Company, The economic potential of generative AI: The next productivity frontier
• Anthropic, Building effective agents
• Andreessen Horowitz (a16z), AI's economics will reshape the services business