Post · 13 June 2023
AI in the corporate strategy
This post was written in June 2023 and is published unchanged. Some examples have aged; the argument has not.
What is the significance of AI in corporate strategy? Can AI bring a competitive advantage?
A well-known Silicon Valley investor told me a few days ago: “I would no longer invest in a company that didn’t have an explicit AI strategy.” But this raises the question of what belongs in an AI strategy. I’ll explore that in this and other posts.
At the very beginning there is the pragmatic question: make or buy? Should you develop your own AI models, or can you buy ready-made models from the market? The answer is both.
In enterprise use, the specific issues are often so individual that transferring a model to another organisation shows little success. One example: the Canadian airline Westjet tried to adopt AI models from U.S. airlines. It soon turned out that the Canadian network, with many small airports and different climatic influences, has completely different characteristics from the U.S. network. Only with their own models do they now succeed at, for example, predictive capacity planning, which enables shorter processing times for baggage handling. From this we can deduce that there are three classes of AI model suitable for enterprise use: transferable models, custom models, and large models.
Transferable models are tailored to a very specific use case while being transferable to many other enterprises. For this, the use case must be generic enough that it is comparable across those companies. A very simple example is an AI for business intelligence that helps make data in a data warehouse easier for business users to query. Such models increase productivity, much like classic software, and therefore become indispensable.
Customised models are based on a company’s specific data. This can involve data that is not available to other companies or that requires special protection. With such models a company can gain a particular market advantage. Obtaining suitable data is one of the biggest hurdles in developing AI models today. Companies that recognise this and base their business strategy on it can build an unbeatable competitive advantage.
The third class are so-called foundation models. These are very large AI models trained by their developers on a massive amount of data. Large language models are the best known representatives of this class (for example ChatGPT); diffusion models — more rarely called large image models — are another (for example Midjourney), based on the generation or transformation of images. Because of the large amounts of data they have been trained on, they have, so to speak, “seen it all before”. They are able to learn from the examples seen during training and apply that experience to new data they have not yet seen. This allows them to achieve valuable results on new data with very little additional training — at least when the data is reasonably comparable to the training data. A large language model trained on huge amounts of text in English and German will produce good results processing English and German text. It is almost worthless when processing Chinese or Arabic. Foundation models are thus suitable for use in very different contexts with little additional adaptation. They have a very large knowledge base and receive their specific adaptation to the actual task inside the user company through fine-tuning or prompt engineering. We’ll look at exactly how that works in a separate post.
So what does this mean for our strategy? A good strategy takes advantage of opportunities in all three classes and applies each where it provides value. In practice, every company should buy transferable models, develop its own models, and use foundation models. But there is no unique selling point to be gained by doing so. This behaviour can only be the basis. A strategic competitive advantage can only be achieved by creating or adapting your own AI models on your own proprietary data. And for that, the handling of data must be completely rethought. Two examples.
OpenAI has created a really interesting product with ChatGPT. The large language model behind it is market leading, and the strategy of combining an open API, plugins and integration into Microsoft Azure is smart. But APIs and integration with a cloud provider are also offered by competitors, and plugins are easily copied. So what happens when the competition’s next model has a better architecture? It is a competition of the best minds. OpenAI’s strategy is therefore far more sophisticated than it looks: ChatGPT has an outstanding user interface and is directly and freely accessible to completely ordinary users. OpenAI thereby collects a huge amount of training data for the coming generation of AIs at low cost. By the current state of the art, this so-called reinforcement learning with human feedback is the best method for producing a very good AI model. AI researchers also refer to this as “the human in the loop”. Through interactions with human users the model learns and improves noticeably. With this tool, free to its users, OpenAI has secured a unique dataset that maps the feedback and ideas of more than a hundred million people. No competitor, not even Google or Facebook, can currently keep up.
Tesla is pursuing a similarly consistent strategy. Tesla is the market leader with its electric vehicles and, with the Model Y, leads the passenger car registration statistics across all drive types in Europe. At the same time Tesla has by far the highest margin of all mass manufacturers and the only global charging network. All signs of an extremely good strategy. But the main attraction of Tesla’s strategy lies elsewhere. Tesla has been equipping all its vehicles with the full hardware for autonomous driving for years, and regularly updates the software on all of them over the air. In every vehicle in the fleet — now more than four million worldwide — Autopilot runs constantly in the background, even when the driver is not actively engaging it. Autopilot constantly compares its own planned driving behaviour with the driver’s actual behaviour. As soon as the software detects important deviations, relevant data is saved and transmitted to Tesla at night. Tesla thus has feedback on its own software from a hundred billion actually driven kilometres, by day and night and in every weather situation. No other manufacturer has remotely comparable data. It is estimated that Waymo has less than one per cent of this amount, and other manufacturers are much further behind. In particular, no fleet of test vehicles, however large, has data on real accident situations, because test drivers must avoid accidents at all costs, if only for regulatory reasons. In practice, however, critical situations and accidents do occur, and Tesla receives detailed data on each.
What do companies do that don’t have their own data? There are basically three alternatives: they buy data from third parties or enter into partnerships for it; they build test environments where the data is collected in a controlled setting; or they build a simulation where the data is artificially generated. Often a combination is used. Simulation in particular places considerable demands on quality. You can observe this in autonomous driving: today’s simulators used by companies like Waymo or Tesla exceed the quality of video games by orders of magnitude — and that for a “player” who pays nothing at all for the experience.
So focus your attention on exactly one question: how can you gain a data advantage?