Deep Dive | Agentic Economy #09: The Vertical AI Playbook — Finding Real Money Beyond the LLM Price War

Consumer AI subscriptions are caught in a squeeze: token costs keep climbing while users resist paying more. That tension makes the model fragile. The more durable bet in AI commercialization points somewhere else entirely, toward high-value users, deep workflow lock-in, and pricing tied to real business outcomes. That's the thesis vertical AI is built on.

In Issue 9 of Agentic Economy, we break down what Harvey, Farther, and Adyen are actually doing, and what it tells us about how vertical platforms can build real competitive advantage as foundation models become commodities. We also get into the harder questions: what happens when token subsidies dry up, and what does it mean that OpenAI and Anthropic are now sending engineers directly into enterprise accounts.

—————————————————————————————————————————

Every technology, once it becomes ubiquitous, stops commanding a premium. Foundation models are no exception. As the marginal returns of parameter competition diminish, what was once a scarce intelligence is rapidly becoming commoditized public infrastructure. The result is that lightweight applications that simply call third-party APIs without deep scene embedding are watching their business models collapse.

But commoditization is never the end of the story. Every time an emerging technology achieves mass adoption, value migrates away from those who own the technology toward those who can actually deploy it.

That dynamic is what is driving the rapid rise of vertical AI platforms today.

A vertical AI platform is an AI application layer that goes deep into a specific industry, tightly wraps the capabilities of general-purpose foundation models, and reengineers business processes around particular workflows. By building proprietary evaluation systems (Evals) and multi-agent architectures, these platforms are reducing foundation models to interchangeable compute components, locking the industry's most critical workflow assets firmly inside their own systems. The core thesis is to eliminate friction from complex professional processes and turn that work into a system asset that compounds over time.

To understand why this category exists, one fact has to be clear: enterprises and professionals do not pay for a model's parameter count. They pay for its ability to embed deeply into internal workflows, close data loops, and drive real revenue.

That is why high-billing lawyers, wealth advisors serving high-net-worth clients, and large merchants with massive transaction volumes are becoming the most contested targets for the next generation of vertical AI platforms. These users control budgets, carry compliance obligations, and are oriented entirely toward measurable business outcomes. Saving a thousand-dollar-an-hour attorney ten hours, or helping a wealth advisor grow AUM and improve after-tax returns, these are outcomes that can be directly quantified.This precise alignment with high-value productive users is the economic foundation that makes vertical AI viable.

Two directions are emerging.

The first is using AI to reengineer professional workflows, dramatically compressing the operational costs that once required large institutions to absorb. In legal and wealth management, high-threshold work like compliance, risk control, and professional delivery is being systematically handled through technology platforms, letting professionals accomplish more with fewer resources.

The second is rebuilding transaction infrastructure, reimagining the connection between merchants and AI agents. In Agentic Commerce, the intent interception and interaction layer is controlled by AI labs, but actual transaction conversion still happens inside the merchant's own infrastructure. Adyen Agentic acts as a universal translator, helping merchants integrate once and participate across AI shopping platforms without rebuilding their systems for every new protocol.

All three cases enter from different angles, but they are all doing the same thing: taking core capabilities that were previously impossible to standardize, and sedimenting them into systems as durable, reusable assets. Harvey sediments legal judgment and domain knowledge; Farther sediments advisor client relationships and tax optimization capabilities; Adyen sediments merchant product data, protocol compatibility, and settlement capabilities.

This is precisely what Microsoft CEO Satya Nadella calls Token Capital: the long-term value of AI comes not from executing individual tasks, but from structurally retaining human judgment, knowledge, and workflows inside a system, forming a self-iterating asset through continuous use.

This issue of Agentic Economy, Issue 9, examines the practices of legal platform Harvey, wealth management system Farther, and payments giant Adyen, exploring how vertical platforms build competitive advantage by reorganizing around high-value users in a world of commoditizing foundation models. It also examines the real constraints these platforms face, from the compute pressure of the Token Apocalypse to the direct competition from model vendors' field engineering teams.

$190M ARR, $468M in Compute Bills: Harvey's Unsustainable Scale Game

Harvey is one of the highest-valued and fastest-growing examples in the current vertical AI wave, and no company better illustrates both the promise and the strain of this model.

This legal platform, which owns no foundation model of its own, grew ARR from $100 million to $190 million in just five months (August 2025 to January 2026) through deep customization of law firm workflows, reaching a valuation of $11 billion. The lesson is clear: vertical platforms do not need to compete in the foundation model arms race. A genuine understanding of industry tasks and the ability to reshape how high-value users work is enough to build a formidable commercial engine.

But behind those impressive numbers sits an ever-expanding compute bill.

According to public disclosures, Harvey's monthly token usage has grown from roughly 1 trillion to 12 to 13 trillion. At $3 per million tokens, the annualized theoretical inference cost reaches $468 million. Even with vendor discounts and techniques like Prompt Caching temporarily keeping the real bill lower, that dependency is the risk: the moment subsidies narrow, costs snap back immediately. Under this kind of financial pressure, ARR growth struggles to convert into real cash flow, and the platform faces a constant threat of being punished by its own scale.

This reflects a cost paradox that AI-native applications cannot escape: the more popular the product, the higher the inference costs. Traditional SaaS carries near-zero marginal cost, but in a setting like legal work, with long context windows and high inference density, every complex task consumes real compute. Building proprietary models has therefore shifted from a technical option to an operational necessity driven by cost.

Harvey is now pursuing a post-training strategy for its proprietary models, working closely with Applied Compute to fine-tune open-source base models such as GLM-5.1 on legal domain tasks. According to the latest technical disclosures from both companies, the post-trained model achieved a rubric pass rate of 0.913 on Harvey's Legal Agent Benchmark (LAB), up from 0.853, surpassing GPT-5.5 xhigh and approaching Opus 4.8 Max.

Cost compression has been equally significant. By switching to GPT-5 Mini as the grader and batching multiple evaluation criteria per call, evaluation costs dropped by 40 to 100 times. This means Harvey can now iterate its evaluation loop continuously at a fraction of the cost, and the proprietary evaluation system itself has become a compounding competitive asset.

More notable is what changed beneath the performance numbers. Output completeness, numerical precision, document grounding, and hallucination suppression all showed measurable improvement. During training, the model made fewer tool calls, but each one was more precise, and total token consumption fell accordingly. In other words, the model learned how to work effectively within a specific tool environment, and that behavioral pattern, accumulated across thousands of legal tasks, is far harder for outsiders to replicate than the model weights themselves.

Harvey's case shows that the competitive foundation for vertical AI platforms is shifting deeper. Workflow design and client relationships matter, of course, but the ability to post-train open-source models, build proprietary evaluation systems, and optimize multi-agent architectures for inference cost are becoming the new sources of differentiation.

Farther: Breaking the Hold of Traditional Wirehouses on Advisors

If Harvey compressed the delivery costs inside large professional service firms, wealth management platform Farther shows how to help top talent escape the gravitational pull of traditional institutions.

Farther is a technology-first RIA platform that recruits wealth advisors leaving firms like Morgan Stanley, Merrill Lynch, UBS, and Goldman Sachs. Inside the traditional full-service wirehouse model, advisors typically face low payout ratios and heavy mid- and back-office administrative burdens. Rather than selling software to external clients, Farther's approach is to recruit advisors directly onto its platform, replacing the capabilities once monopolized by large institutions with a single integrated system. Along with higher payouts, it bundles tax-loss harvesting, direct indexing, private market access, compliance review, and document management. According to the company's own data, the tax intelligence layer alone can deliver a 1 to 3 percent improvement in after-tax investment returns for clients.

The market has responded. In May 2026, Farther raised $150 million in a Series D led by General Atlantic, becoming a unicorn with recruited assets surpassing $23 billion, including a team poached from Goldman Sachs Private Wealth managing over $1.5 billion in client assets. The steady flow of advisors choosing to leave shows that the hold traditional wirehouses once had is loosening, and going independent is no longer a fringe choice.

Harvey focuses on improving professional delivery efficiency inside law firms. Farther builds an independent platform so advisors no longer need to rely on large wirehouses to run a full practice. The two enter from different angles, but both are redefining how professional services get produced. On Farther's platform, sophisticated tools like direct indexing and private market access that were once reserved exclusively for ultra-high-net-worth clients at large institution desks are now available to independent advisors for their own client base, significantly expanding what a solo practice can offer.

Traditional SaaS only handles the shallow end of process automation, things like record-keeping, storage, and basic workflows, and cannot absorb the complex coordination and execution that professional services require. Multi-agent AI systems are naturally suited to take on the gray zone between administrative execution and non-standardized judgment calls: compliance review, personalized document drafting, asset allocation recommendations. Work that once required an entire back-office team is being absorbed by the system. Administrative burden falls, and advisors can redirect their time toward the decisions that genuinely require human judgment.

The Underestimated Merchant Side: Closing the Loop in Agentic Commerce

Discussion around Agentic Commerce continues to run hot, but attention has been almost entirely captured by the consumer side, with most of the conversation focused on how AI assistants search for products, compare prices, and place orders on behalf of users. The reality on the merchant side is considerably more sobering.

Walmart's conversion rate on its AI-native checkout (Instant Checkout) is currently only one third of its traditional click-through model, and the share of merchants that have fully integrated Shopify's AI checkout system remains limited in 2026. Between AI activating demand and actually completing a transaction, there is a clear and persistent gap.

The reason is that agent-led transactions are a full systems engineering problem. Understanding user intent is just the first step. Converting that intent into revenue requires inventory verification, tax calculation, fraud risk controls, fulfillment, and payment settlement across the entire chain, and all of those capabilities are still locked inside merchants' own local systems. Meanwhile, multiple agent payment protocols, including UCP, ACP, AP2, Agent Pay, and Visa Tokenization, coexist without being mutually compatible. Merchants have neither the incentive nor the bandwidth to adapt to each one individually.

Adyen has responded to this with Adyen Agentic, a suite of modular APIs that covers different stages of the transaction chain.

• Agentic Feed distributes a merchant's product catalog, pricing, and real-time inventory data to major AI platforms in a standardized format.

• Agentic Cart connects the merchant's existing checkout, tax, fulfillment, and order management systems to conversational commerce platforms.

• Agentic Payments handles identity verification, network risk controls, and multi-channel settlement for agent-led transactions.

Merchants integrate once. Adyen then translates that single integration across different AI agent platforms and protocols, so merchants never have to rebuild their systems when the landscape shifts.

In the Agentic Commerce ecosystem, AI labs and front-end interfaces may capture user intent, but the actual value conversion, transaction completion, and revenue settlement still depend heavily on the merchant's own infrastructure. Compared to the intensely competitive front-end entry points, the merchant-side integration layer has a far better chance of becoming stable, billable foundational infrastructure.

The Hidden Pressure on Vertical Platforms: Model Vendors Moving In and the Token Cost Crunch

As the market for low-cost general tools becomes saturated, the commercial logic of sustaining large model platforms through low-priced subscriptions is showing its fragility. When capabilities like summarizing web pages or drafting emails can be replicated by cheaper alternatives at any moment, vertical platforms have no choice but to move toward clients who genuinely care about business outcomes. But the further they push into high-value industries, the more complex the competitive environment becomes.

One source of pressure comes from model vendors actively expanding their business boundaries. OpenAI and Anthropic are no longer content to be API wholesalers. They are sending field development engineering teams directly into core client sites. In April 2026, OpenAI partnered with Customers Bank, which holds $26 billion in assets, with engineers embedded on-site to build agents for loan approvals and account opening using local data. Anthropic partnered with financial IT giant FIS, embedding its FDE team inside FIS systems to develop anti-money laundering tools, reaching deep into banking operations through FIS's extensive network of bank clients.

This on-site collaboration model shows that model vendors are using infrastructure channels to directly learn and replicate the internal workflows of high-barrier industries.

The other pressure is an unsustainable Token pricing logic. Most frontier foundation model Tokens are currently being sold at subsidized, below-cost prices. As enterprise-grade multi-agent architectures drive high-frequency usage, the moment those subsidies narrow, vertical platforms that depend entirely on external frontier APIs will find their compute bills impossible to sustain.

This pressure will intensify as inference demand climbs. When hundreds of agents run around the clock in the background at high frequency, demand for compute will grow exponentially, while the physical constraints of supply chains anchored to manufacturing cycles as long as those of ASML lithography machines mean the underlying hardware cannot scale fast enough to keep up. For the vast majority of everyday business tasks, running everything through frontier models is simply a severe misallocation of resources.

This is precisely why Harvey has had to partner with Applied Compute to build dedicated test sets, proprietary evaluation systems, and human annotation pipelines. Vertical platforms are not just building products. They are engaged in demanding cost engineering: precisely calculating the Token consumption of every task, determining which intermediate steps can be routed to lower-cost fine-tuned open-source models, which critical decisions must call the flagship model, and where human review should step in.

Relying on a polished workflow interface layered over a model is no longer enough to sustain lasting competitive advantage. Taking backend cost engineering to its limit is the defining challenge for any vertical AI platform that wants to survive long term.

Where the Real Value Hides

When large language models become as unremarkable as tap water, the application layer's real value gravitates toward the two ends that actually matter: the very top and the very bottom of the chain.

The good news is that scarcity didn't go anywhere. At the top, you still have things no algorithm can touch: client relationships, messy judgment calls, and hard-won know-how that lives in people's heads, not in training data. At the bottom, you have merchant networks carrying product data, compliance requirements, and payment rails. What vertical platforms actually do, when they're doing it right, is take that expertise from high-value players and turn it into something that compounds over time.

This also means the competitive playbook is getting a lot more down-to-earth. The "just keep scaling" narrative that carried software companies for a decade is running into a wall. Compute isn't free, and physical supply chains don't bend to a good pitch deck.

Going forward, staying alive in the application layer is really about playing a smart arbitrage game between cost and performance. As model prices stabilize and compute stays tight, platforms need to prove they can actually engineer their way to the right trade-offs, not just throw money at the problem.

Sure, the big model labs have more GPUs and bigger teams. But for nimble vertical platforms and specialized professionals, the edge that's genuinely hard to copy is this: turning years of domain expertise into system-level assets that no foundation model provider can just replicate from the outside. Skip the fight for generic traffic. Focus on closing real business loops for the people who actually create value. That's the playbook for vertical AI that lasts.