AI Graduates from "Chat": In 2026, Agents Will Get the Job Done
The 2026 AI story will be told not as an improvement in performance itself, but as the year that bridges the gap between capability and adoption. OpenAI CFO Sarah Friar and investor Vinod Khosla emphasized that a turning point has arrived, pushing AI from a "tool that answers via chat" to an infrastructure that completes tasks and delivers results (OpenAI Podcast Ep. 12).
1. 2026 is the year "agents get the job done"
In their view, while 2025 was the maturation of "vibe coding," 2026 is the year that agents (especially multi-agents) begin to produce visible results.
1-1. Enterprise: Automating "daily operations" like ERP, contracts, and accounting
They cited as an example a future where multi-agents handle full tasks such as "ERP operations," "daily reconciliation," "accrual of accounts payable," and "contract tracking." What is important here is not mere RPA, but a transition to "task workers" that handle the entire sequence from situational understanding to judgment, execution, and verification.
1-2. Consumer: Travel planning is a "multi-agent problem that is a hassle for humans"
Travel planning that spans "food preferences," "restaurant reservations," "flight tickets," and "personal calendars" was cited as an area where multi-agents will truly excel. As AI shifts from something you "call upon" to something that "blends into the environment," such hassles will be reduced all at once.
2. We have handed over the "keys to a Ferrari," but we haven't mastered driving it yet
Friar uses a striking analogy for the current state of AI. "We have handed people massive intelligence—the keys to a Ferrari. But most people are just now getting out onto the public road for the first time." What is important is not just the intelligence of the model, but the path (UI/workflow/integration) that allows anyone to reach a result.
In this context, the theme for 2026 is from "question answering" to "outcomes."
Finishing travel planning "all the way to booking"
Organizing a "second opinion" after receiving a doctor's explanation
Creating a "meal plan" for a child with diabetes
—the focus will be on translating these into "results that improve life."
3. A map of the bubble theory: Look at "API calls," not stock prices
Khosla criticizes the stance of measuring a "bubble" by stock prices, drawing a parallel to the dot-com era, and states that "reality" lies in usage volume (API calls, traffic). In other words, we should look at "how much it is being called," not "prices swayed by fear and greed."
The constraint appearing in the background here is a shortage of compute. OpenAI states that "computing resources are the scarcest thing in AI," and that demand is constrained not by "time" but by "supply (compute)."
4. The source of growth is the correlation between "power x revenue" = proof of infrastructure
Friar describes the structure of business growth as "infrastructure" using quite specific numbers. In an OpenAI public article, it is explained that computing resources (GW) and ARR are expanding at almost the same slope,
Compute: 2023 0.2GW → 2024 0.6GW → 2025 approx. 1.9GW
ARR: 2023 $2 billion → 2024 $6 billion → 2025 over $20 billion
—this correspondence is presented.
This logic suggests that "AI is not a service like Netflix that hits a ceiling based on disposable time, but rather an infrastructure whose utility can multiply like electricity."
5. Healthcare is "the most important but hardest to regulate"—which is precisely why its adoption will be symbolic.
Healthcare is a high-risk area, and Khosla mentions regulations (such as the FDA) and resistance from vested interests. On the other hand, usage is already massive. OpenAI has been introduced as having "230 million people asking health and wellness questions every week," and this has been reported repeatedly in the context of related products (ChatGPT Health).
Furthermore, AI adoption is progressing on the medical front as well, with an AMA survey reporting that "66% of physicians use some form of AI in their work" (it is safer to read this as the usage rate of AI in general, rather than "ChatGPT itself").
6. The winning strategy for startups: Stack "data x permissions x operations" on top of the model.
To the question, "Isn't there no room left?" their answer is clear.Models are general-purpose, while global operations are specific. The winning strategy is to
data dormant within companies (much of which is behind firewalls)
permission design (who can access what)
complex operational workflows (approvals, audits, compliance)
in areas that hold these, and build a "system for execution." In other words, it is not just about being "smart," but about creating a mechanism that can take action.