
Underneath almost every product built in 2024-2025 sits the same handful of foundation models. On top of them are the model builders’ own app layers: ChatGPT, Claude, Perplexity, and Gemini, with projects, tools, memory, and agents. End users never actually touch a base model. When a startup calls the OpenAI API to build a copilot, and OpenAI uses the same API to build its own assistant, they’re both just clients. But one of them also owns the model, the roadmap, the pricing, and the default app on millions of desktops.
That structural advantage will compound. Model builders see aggregate usage patterns across millions of users, can ship features directly into their flagship apps, and don’t pay anyone else’s margin. As their app layers grow more capable, a lot of the generic “copilot” use cases that wrappers rely on will simply be absorbed into the default chat products.
The B2C AI gold rush could hit its first real sorting in 2026. The explosion of consumer apps wrapped around foundation models, note-taking copilots, chat companions, slide generators, writing tools, sit in the same place in the stack: they broker access between users and a small handful of underlying models. Unless a consumer app is doing something genuinely novel, it is competing with the model builders’ own app layers. A lot of perfectly decent AI apps will discover they were never the product. They were the middleman.
“Unless you're a B2C app that's doing something truly novel, I think you're under more threat because the model builders can just add more and more stuff, and they've got so much funding, they can keep doing it. But enterprise is so complex that they same doesn’t apply” – Caspar Eliot
Enterprise is a very different story. You’re not just tweaking a chat box; you’re trying to drop AI into SAP instances from 2009, half-documented APIs, regional compliance rules, union agreements, three different CRMs, and a reporting calendar the CFO will not move. You’re dealing with brittle legacy systems, messy permissions, approvals that run through six people, and workflows that span email, tickets, spreadsheets, and mainframes. That complexity is exactly why there is still plenty of room for specialized builders on the enterprise side. To get AI to actually work in a bank, a telco, or a manufacturer, you need deep integrations into systems of record, serious data plumbing, change management so people don’t revolt, and domain logic that reflects real policy and edge cases.
2025 was the year every company stood up basic agents: summarize the meeting, autosend recap emails, do pre-read research. They look impressive in demos and pilots. But they messed up in production: returning records for the wrong John Smith, misreading the escalation path, or pulling stale pricing from the wrong system.
In 2026, the serious work is wiring improvement loops into those workflows: eval suites tied to specific processes, and agents that update their retrieval, prompts, and guardrails when they’re corrected. Integration into SAP, CRMs, and policies remains hard, but priority shifts to enterprise agents and evals that get more reliable over time in a given workflow. The leap from GPT-2.5 to GPT-5 happened at the consumer level; 2026 is when that kind of compounding improvement finally starts to show up inside the enterprise.