AI Platform Wars Are Moving Into Workflows Where Billions Are At Stake

TOPSHOT - Humanoid robots are displayed at the robotics pilot testing and validation platform of Beijing Innovation Center of Humanoid Robotics (X-Humanoid) in Beijing on March 20, 2026. (Photo by ADEK BERRY / AFP via Getty Images)

AFP via Getty Images

Despite AI infrastructure investment that topped hundreds of billions in 2025, most professionals still interact with AI by leaving their work, opening a chatbot, pasting context, and manually reintegrating the response. The integration gap is now the central strategic battleground of the next AI cycle. The companies that close it may capture platform-level economics and those that do not may likely be compressed into infrastructure roles with shrinking margins. Even Patrick Collison published on X recently that he is looking for a collaborative LLM workflow tool that combines file/context management, prompt orchestration, coding agents, and versioned build outputs; essentially GNU Autotools meets Notion for AI pipelines.

Andrew Gershfeld, GP at Flint Capital, has been among the more rigorous voices in articulating why this shift matters structurally for investors. In a recent thesis, he draws a direct parallel to the PC era: Intel powered the revolution, but Microsoft and Apple captured the durable economics by owning the environment through which users accessed computation. His argument is that leading AI developers face the same bifurcation today; remain a model provider and face inevitable margin compression as ecosystems absorb AI as a native feature, or secure control of a workflow surface and qualify for platform valuations.

Venture and growth investors who spent 2024 and 2025 pricing AI companies on model capability benchmarks may be systematically mispricing the next wave. The companies with the highest long-term defensibility may not be the ones with the best reasoning scores. They may be the ones with the deepest workflow lock-in.

The incumbents with the clearest structural advantage are already visible. Google anchors the professional workday through Docs, Gmail, and Calendar, and has embedded Gemini into those surfaces with increasing depth. Microsoft controls enterprise workflows through Windows, Office, and Teams, and has spent two years weaving Copilot into the daily operational layer of millions of organizations. Meta owns the messaging infrastructure where billions coordinate decisions, plans, and intent. Apple controls the hardware entry point for nearly every consumer interaction with computation. These companies do not need to win on model performance. They need to make intelligence disappear into surfaces people already inhabit.

OpenAI's memo on enterprise strategy acknowledged the pressure directly, emphasizing platform ownership, workflow integration, and multi-product ecosystem deployment as competitive necessities rather than growth options. That language reflects an awareness that model capability alone, however impressive, does not guarantee the distribution or contextual depth required to justify platform valuations.

The vertical AI companies are making the same argument from a different entry point. Cursor's approach did not build a better code assistant. It embedded AI into the coding environment itself, giving the model continuous access to the full context of a codebase, not just the snippet a developer chose to paste. Harvey is taking the same approach in legal infrastructure, building AI systems that operate with comprehensive awareness of case documents, firm-specific standards, and jurisdictional constraints. Notion recently restructured itself as a hub for AI agents rather than a document editor with AI bolted on. The common move is context depth over general capability.

Leo AI, building AI systems for engineering workflows, illustrates why this matters at the operational level. Founder Maor Farid has argued that in high-stakes engineering environments, generic AI is insufficient not because of reasoning limitations but because engineering decisions depend on manufacturing constraints, supplier logic, proprietary documentation, and company-specific standards that no general model has seen. The AI becomes exponentially more useful once embedded in the operational environment rather than accessed as a separate tool.

Gershfeld's Flint Capital colleague Sergey Gribov has been investing in this thesis from what he calls the “unsexy infrastructure angle”. Portfolio companies that Gribov backed, including Viewz (AI-native platform for finance and accounting), Vayu (AI-powered billing orchestration), and RiskFront (agentic AI for risk and compliance operations, recently acquired by K2 Integrity) are each built on the premise that the operational categories nobody romanticizes; compliance workflows, billing logic, governance infrastructure. These sectors are precisely where AI embeds most durably. They are also the categories where switching costs and logistics/friction involved are highest.

“With such startups as Viewz, Leo AI, and RiskFront, the bet is the same: AI that lives inside the workflow, not next to it,” says Gribov. “A chatbot you visit is a feature. A system that holds your operational context — your ledgers, your engineering constraints, your risk logic — is infrastructure. Features get commoditized; infrastructure gets renewed.”

The finance and accounting sector is producing supporting evidence for this thesis at scale. Investors are pouring capital into what was previously considered accounting's most static infrastructure layer. The pattern holds across industries: AI becomes defensible when it is not adjacent to workflows but constitutive of them. As Gribov puts it, the next AI winners won’t be the companies that ask users to leave their workflow and come to the model, but the ones where the model disappears into the workflow itself.

Gershfeld frames the end state as a semantic operating layer, an environment in which AI continuously understands the relationships between conversations, documents, schedules, and collaborative threads rather than responding to discrete prompts. A team discussing a product launch in a chat does not want an assistant that summarizes the conversation on request. It wants a system that recognizes a project forming, links dates to calendars, assigns tasks to owners, and maintains structured state across the team's operational surfaces.

The companies that build that layer will control the surface where intent becomes executable. In technology markets, that has historically been the position that captures the most durable economic value.

For investors pricing AI companies in 2026, the relevant question is which companies are accumulating the contextual depth, workflow integration, and embedded user habits that make switching genuinely costly. The big tech earnings cycle will increasingly reveal which incumbents are succeeding at this conversion, and which AI-native challengers are building workflow surfaces that could rival it. The companies that cannot credibly answer the distribution question may find that technical superiority is a shrinking competitive moat as intelligence becomes ambient infrastructure rather than a destination anyone consciously visits.