Back to Blog

    Strategy

    OpenAI wins on customers, not on models

    David George (a16z) argues that OpenAI will not win because of the best models but because it awakens new behaviour and owns distribution. What that means for Swiss organisations on Microsoft 365.

    September 29, 20265 min read
    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    marcel.haas@cnext.ch
    In this article

    "Who has the best model?" is no longer the most important question for a business. Intelligence is available to everyone, at every price point. What matters is who awakens new behaviour in users and who gets it to them. That is the thesis of David George, General Partner at a16z, in his essay "OpenAI Understands Something Important and Rare", published on 28 September 2026. We think the analysis is right. And it has a direct consequence for every Swiss organisation deciding about AI agents right now.

    Cover image of the essay "OpenAI Understands Something Important and Rare" by David George

    Image: David George, a16z. Source: essay on X, 28 September 2026.

    What David George says

    George opens with Peter Drucker: the purpose of a business is to create a customer. OpenAI will not win because it has the best models, chips or product, even though a16z credits it with all three. OpenAI will win because it keeps creating new kinds of customers and follows the most durable distribution strategy.

    His argument in short. Anyone building a durable business at the AI frontier has four levers: awaken new behaviour, reach many people, price so that value sticks, and build high switching costs. Switching costs are gone, because models are swappable task by task. Price goes to whoever has scale anyway. That leaves the first two levers, and there George sees OpenAI ahead: in the simple, obvious-in-hindsight breakthroughs such as the chat interface, reasoning, tool calling and computer use, and in choosing a real platform over partnerships.

    Two observations from the essay strike us as especially important:

    1. 1Breadth beats focus. Improving a general capability such as "operating a computer" requires exposure to a very wide range of workflows. OpenAI has that breadth across consumers, prosumers and enterprises. It is expensive, it contradicts the textbook product organisation, and that is exactly why hardly anyone copies it.
    2. 2Agents have to write code. George describes what we learned over the past year as well: models are remarkably good at programming and remarkably unreliable when asked to call tools directly and carry out instructions in ordinary office work. Turn the task around ("write a program that does these steps") and they become dependable.

    George concludes that at OpenAI the causality runs backwards: the best models, chips and products follow from the best distribution, not the other way round. Whether that forecast holds, we do not know. The market share and chip figures in the essay are a16z's; we have not verified them.

    Why this matters to us

    CNEXT is a Bern agency for SharePoint, AI agents and applications, not a frontier lab. We do not build models, we build with them. Yet George describes the situation of our customers precisely, at a smaller scale.

    In an organisation of 30 to 1,000 people, intelligence has been available for two years. Microsoft 365 Copilot is licensed, ChatGPT is used privately, Copilot Studio is ready. And still, daily work in most departments has barely changed. The model is not the bottleneck. The bottleneck is that nobody has awakened the new behaviour: the case worker who lets an agent prepare the legal enquiry; the practice assistant who no longer types appointment requests; the project lead whose meeting is documented without anyone writing minutes.

    Organisations therefore face the same two levers as OpenAI, pointed inwards: awaken behaviour and secure distribution.

    What we take from it for you

    Do not bet on a model. If even a16z writes off switching costs at the model layer, no organisation should build agents that depend on one provider. In every agent we separate the workflow, the data access and the approvals from the model. Whether OpenAI, Anthropic or a local model via Ollama runs behind it is a configuration decision, not an architecture decision. How we solve that per case with Copilot Studio, Azure Foundry or a custom application is described on our AI agents page.

    Distribution, for you, means where the work happens. OpenAI fights for platform reach. You already have your platform: Teams, Outlook, SharePoint. An agent that needs its own interface loses the battle for habit against the one that answers in the department's channel. That is why we bring knowledge agents such as the Company Brain into the existing working environment first and talk about separate apps afterwards.

    New behaviour needs a first, small proof. George's account of OpenAI after the ChatGPT launch is a reminder that nobody there knew in advance which behaviour would catch on. We therefore work with one pilot per workflow: a recurring process, approved data, a testable result, an agreed test set with valid and invalid inputs. In the Agent Live Build, a first agent takes shape in two hours, and the business team carries it on afterwards.

    Let agents program instead of click. For some months we have built agents so that they generate and run small programs for their steps instead of calling tools one by one. The error rate drops, and the result can be checked and repeated. George's essay confirms that path. What it means for permissions and operations is in "Pace the Frontier".

    Limits

    The essay is the view of an investor with a stake in OpenAI. It describes a strategy, not a settled future. Market shares in agent traffic swing within weeks, and the platform question is open: whether organisations end up running their agents on OpenAI, on Microsoft or on their own harnesses remains to be seen. For your decision that is secondary. Whoever awakens behaviour and secures in-house distribution today benefits from whichever provider leads tomorrow.

    Next step

    Which recurring workflow should your first agent take over? Describe today's path, the systems involved and who decides today. I look forward to the conversation.

    Marcel Haas, CEO CNEXT

    Source: David George, "OpenAI Understands Something Important and Rare", X, 28 September 2026.

    OpenAIAgentic AIBest Practices
    Teilen:

    AI tools may have supported the creation of this article. CNEXT is responsible for the published content. Product features and prices can change; check the linked primary sources for current details.

    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    6x Microsoft Applied Skills

    Plan your first agent

    We define the workflow, data access and approvals for a first AI agent with you, independent of the model provider.