For much of the past period, the primary anxiety among most AI application entrepreneurs has been: will large models eat up all applications?

This year, OpenAI and Anthropic have continuously launched vertical suites covering law, design, and office productivity. Microsoft, leveraging its Office ecosystem, has bundled Copilot to steadily penetrate enterprises.

An increasing number of people now believe that as long as foundational models keep evolving, today’s AI applications are merely waiting to be “built-in” by official platforms.

Yet just recently, Arvind Jain—founder of Glean, a $7B enterprise AI unicorn, and previously builder of the public cybersecurity giant Rubrik—appeared on the top-tier venture capital podcast 20VC, delivering a sharp judgment:

OpenAI and Anthropic cannot win in real application-layer enterprises. The reason is simple: general-purpose models can handle “standardized production,” but they can never replace an application company’s deep understanding of vertical business domains.

More importantly, cutting-edge closed-source large models themselves are losing appeal. Currently, open-source models already fully meet demand in over 90% of enterprise scenarios.

Below is the full interview transcript:

Enterprises Are Increasingly “Untrusting” of Frontier Models

In early July, Palantir CEO Alex Karp made a striking claim on CNBC: large enterprises’ distrust toward frontier model vendors has reached unprecedented levels.

The root cause lies in the long chain between the “capabilities” demonstrated in labs and the actual “value” delivered in enterprise operations.

Currently, many enterprises deploy AI in a highly direct way: purchase model services and integrate them into internal systems—documents, emails, CRM, tickets, code repositories, and various business platforms—via MCP and similar protocols. Management imagines that if the model sees enough data, it will automatically understand the company and begin generating value.

Reality differs drastically.

A large enterprise may operate hundreds or even thousands of disparate systems. The same customer might be labeled “account” in sales, “entity” in finance, and mapped to multiple projects in delivery.

Permissions, versioning, business definitions, and implicit rules layer upon one another. While the model can read this information, it doesn’t know which piece is most relevant to the current task—or how the company historically completed similar work.

Thus, it resorts to a brute-force-like context assembly: first retrieving vast amounts of material, then attempting to infer relationships, finally generating an answer that looks plausible.

This leads to two immediate consequences: first, slowness, as significant time is spent searching for raw materials; second, high cost, as massive token consumption occurs during ineffective retrieval and context stitching. Enterprises end up buying a smart brain without providing memory, maps, or workbooks.

Therefore, the real bottleneck for enterprise AI adoption is rarely model capability—it’s context engineering.

Here, “context” goes beyond simply stuffing a few documents into prompts. It encompasses the enterprise’s knowledge architecture, data access policies, terminology relationships, business workflows, and those uncodified, tacit insights embedded in employee experience.

For example, when a task is executed for the first time, the enterprise may document ten standard steps. Over subsequent years, employees continuously reorder, bypass exceptions, and develop shortcuts. Much of what truly drives efficiency remains undocumented—existing only in human judgment. In the future, as intelligent agents take on tasks, these experiences will gradually crystallize within the agents themselves.

This raises a deeper issue than “who owns the data”: whose does the organizational wisdom accumulated by AI over long-term work belong to—the enterprise or the vendor providing the intelligent agent?

Enterprises are no longer just worried about technical dependency—they fear operational dependence. If core work is performed by external agents that accumulate compounding organizational knowledge, switching suppliers ceases to be a mere software replacement and becomes akin to replacing the company’s “external brain.”

Thus, what enterprises truly need to control is not just the model and data, but also the experience AI generates during work.

90% of Enterprise AI Scenarios Don’t Need the Strongest Model

If context determines whether AI can be used, then cost determines whether it can be scaled.

One of the most notable insights from Arvind Jain’s interview is this: today, 90%, or even more, enterprise use cases no longer require the strongest models.

The implication behind this figure is clear: enterprises will only pay for results that are “good enough” and cost-controlled.

Whether it’s an internal meeting minutes or a follow-up sales email, there’s no need to invoke the world’s most powerful model. As long as quality meets business requirements, cheaper models naturally prevail.

This signals a familiar pattern in the software industry: scarce capabilities become standardized supply, and technological advantages turn into interchangeable commodities.

Glean’s approach is particularly illustrative. Instead of forcing customers to bet on a single model, it automatically selects the appropriate model based on task complexity: simpler tasks go to lower-cost open-source models, while complex ones trigger stronger closed-source models.

In this architecture, the relationship between application companies and model providers shifts. The fiercer the competition among model vendors and the faster the open-source community advances, the cheaper and richer the raw materials become for application companies. The real danger isn’t continued model upgrades—but rather, application companies failing to create other sources of value beyond merely calling models.

This is why “wrapper” is such an easily abused term.

Calling the same model doesn’t equate to offering the same product. Just as all e-commerce sites run on similar cloud infrastructure, what truly defines user experience is supply chain, fulfillment, recommendation engines, customer service, and organizational efficiency. Similarly, differentiation in enterprise AI comes from permission systems, context quality, workflow orchestration, result evaluation, and delivery capability.

Naturally, open-source substitution isn’t without resistance.

Arvind Jain noted that some enterprises still harbor psychological and compliance concerns regarding Chinese open-source models—even when models can be deployed locally and data never leaves the organization, decision-makers remain wary of potential security and reputational risks.

But this reflects a typical challenge in technology diffusion: the “first adopter problem.” Once leading enterprises start using these models and set a precedent, concerns will shift from “can we use it?” to “how can we use it more securely?”

Long-term, the enterprise model market is unlikely to be dominated by a single player—it will likely evolve into a multi-model ecosystem. Profit margins at the model layer will be compressed by competition, while value increasingly migrates to the application layer responsible for selecting, connecting, and managing models.

Microsoft’s Entry Advantage Is Gone

Model commoditization also undermines one of the most effective competitive weapons in traditional enterprise software: bundling.

Microsoft’s strength has never been just its products—it’s its ability to embed new features into enterprise-purchased software suites. For a large enterprise, getting a “70-point” Copilot for free is often far more convenient than purchasing a standalone “90-point” product. The latter requires additional spending, plus security reviews, legal approvals, vendor management, and employee training.

Arvind Jain candidly admitted that Microsoft is a more realistic competitor than frontier model vendors. Many clients reject Glean because “we’ve already bought Microsoft.”

However, AI’s pay-per-task pricing model is creating a crack in the bundling logic.

Traditional SaaS charges per seat—once a suite is purchased, enterprises tend to encourage maximum usage. AI, however, is closer to pay-per-task: regardless of bundling, executing any task consumes compute resources. Enterprises can simultaneously enable multiple tools, let employees choose the best fit, and pay only for actual usage.

This means bundling still holds procurement and distribution advantages—but it no longer guarantees final lock-in. Future competition isn’t just about “who gets on the vendor list,” but also “who actually gets assigned the task.”

This is a subtle yet critical shift: the entry advantage of the software era is being re-priced by task completion quality in the AI era.

AI Cost Reduction Is a False Proposition

Another frequently confused aspect of enterprise AI is return on investment (ROI).

In customer service scenarios, ROI is easy to calculate: a support agent once handled 10 tickets daily; after AI integration, they handle 12—efficiency gains are immediately visible. But in R&D, management, and knowledge work, calculation is far more complex.

Taking Glean as an example, nearly 100% of initial code generation is now AI-driven—no manual first draft is written. Yet product delivery speed hasn’t exploded.

Why? Because writing code is just the shortest link in the chain. When code volume increases, code review, architectural design, cross-team coordination, testing, and deployment still proceed at the original pace.

When AI makes code generation extremely fast, the bottleneck shifts to code review. If human review is skipped in pursuit of speed, enterprises risk ending up with unmaintainable, incomprehensible codebases years later. Thus, Glean maintains strict human review, sacrificing short-term efficiency for long-term maintainability.

This reveals a fundamental truth: AI doesn’t automatically eliminate process bottlenecks—it merely pushes them downstream.

Thus, many AI applications currently don’t look economically viable, and pure cost reduction remains elusive in the short term. Long-term, there are only two paths to genuine AI cost reduction:

First, wait for costs to fall.

History shows technology always gets cheaper. Open-source models have already driven inference prices down to one-tenth of proprietary models. OpenAI’s price cuts are inevitable. Once token costs drop to one-tenth—or even lower—what was once a $1M agent becomes an exceptionally compelling investment.

Second, use AI to amplify individual productivity—not just through headcount reduction.

Imagine two competing firms in the same space, both equipped with equally advanced AI tools. Company A opts to shrink its team, doing the same work with fewer people. Company B maintains team size, using AI to scale output tenfold and deliver superior products.

Over time, Company B will utterly outperform Company A.

Because a company’s ultimate competitiveness isn’t just cost—it’s output. Enterprise leaders understand AI capabilities in two ways: “the ability to do more with one person” versus “the ability to do one thing with fewer people.” If you choose the latter, and your competitor chooses the former, you’ll be quickly left behind.

AI won’t eliminate jobs—but it will eliminate roles focused solely on repetitive execution.

Future roles will be fully hybridized: a technically savvy product manager can directly use AI to generate prototypes; a salesperson can simultaneously handle solution demos and negotiations, no longer needing a solutions engineer trailing behind.

Pure execution roles like basic data analysts, recruiter sourcers, and report generators will gradually vanish—but cross-functional, composite talent will become essential.

Integration into Core Workflows Is the Key to AI Applications

Finally, returning to the entrepreneur’s original anxiety: what should application companies do when model providers enter the space directly?

OpenAI is rolling out vertical suites; Anthropic is building industry-specific solutions—from design to law to finance—appearing intent on eating away at application-layer companies’ livelihoods, one by one.

Arvind Jain’s verdict: no need to panic. These players operate on fundamentally different competitive dimensions.

Take design: Claude introduced design assistance features, seemingly threatening Figma’s business. In reality, designers still rely primarily on Figma. Claude enables non-designers—product managers, marketers—to generate simple assets independently.

In other words, Claude functions more like market expansion—bringing tool users into the fold who previously didn’t use tools—but deep work remains firmly within Figma.

Even deeper, model vendors fail to grasp the true inner workings of enterprises.

General-purpose models currently interface with enterprise systems via standardized protocols like MCP. They connect crudely, perform brute-force data retrieval, and attempt to complete tasks. But this model suffers from two fatal flaws:

First, extreme inefficiency. Vast token consumption occurs merely in “finding things”—the AI must first determine what information is needed, where it resides, and how to stitch it together. Before any real work begins, most of the budget is already burned.

Second, complete ignorance of context. Internal permissions, historical decision logic, proprietary jargon, and implicit rules are entirely unknown to general-purpose models. Asking an external model to understand “why this supplier was flagged as high-risk last year” leaves it clueless about where to start.

Glean’s positioning serves as a direct contrast. They spent years integrating hundreds of internal enterprise systems—Salesforce, ServiceNow, Google Drive, Slack, Notion, dozens or even hundreds of distinct data sources—building a permission-based organizational knowledge graph.

When employees search for information, results returned are the most relevant ones within their access scope. This moat is built over time and depth of embedding.

Thus, what application companies should genuinely worry about is whether their product is deeply embedded in core workflows.

If value is merely a chat interface, model upgrades could wipe it out overnight. But if a product captures enterprise context, permissions, processes, and long-term learning, model upgrades actually amplify its value.

This is also why Glean’s founder has shifted his operating philosophy. He once emphasized restraint, capital efficiency, and ROI on every dollar spent. Now, he worries that excessive caution may cause him to miss a rare window for market capture.

To him, enterprise demand for AI has been awakened. Not entering customers’ core workflows today may mean future entry costs are ten times higher.

Yet “acceleration” doesn’t mean abandoning commercial discipline. Products still need to be priced, still need to deliver measurable value, and marketing investments must yield returns. Capital can help companies seize time—but it cannot indefinitely mask the fact that a product lacks real value.

This is perhaps the most authentic contradiction in the enterprise AI market: on one hand, the window of opportunity demands bold investment; on the other, model costs, customer ROI, and business models demand restraint.

Finding balance between the two is the true test for this generation of AI entrepreneurs.

Source: Silicon Observation Pro, Author: Yuan Yuan

Source: Silicon Observation Pro