Let’s try a little thought experiment in technology-fiction: what would have happened if Larry Page and Sergey Brin’s PageRank, back in 1998, had failed to attract its first investors? Google didn’t invent the search engine: there were plenty of them before, from AltaVista, Excite, Lycos or others who were counting words on pages and selling their results page to the highest bidder, to catalog or portal concepts such as Yahoo! or AOL, that attempted to capture the users’ attention as much as they could to be able to serve them advertising, the more the better. Imagine if these concepts would have gone the distance, and deprived us of the web that resulted from Google’s easy and efficient way of searching …
What did Larry Page and Sergey Brin realize? They adapted and extended the citation index (or its consequence, the citation impact), a system that they knew worked for academic literature (remember, these guys were graduate students working on their Ph.D., and thus reading academic papers like there was no tomorrow!), and moved it into a new context: the web. They realized there was an underused signal, relationships among pages, and used it to generate a metric, weighting links according to the importance of the pages providing them, that not only allowed them to index the web, but also to define attention as a social construct: a page is important when other pages link to it with a certain term.
What’s today’s underused signal in corporate AI?
As I’ve said several times, corporate AI is completely different from individual AI: companies don’t need just answers to questions, what they need is persistent memory and state, feedback learning loops, embedded cybersecurity, logs that can be tracked and not manipulated, etc.
So far, companies are building these required scaffolding using twentieth-century tools such as Python, and AI agents are proving to be smart enough as to escape from these containments and do all sort of bad things — not because they are intrinsically evil (they cannot be), but because the goals and the guardrails are badly defined, or due to excessive permissions, prompt injection, or weak isolation. These companies are trying to define these properties by repeatedly assembling application by application, rather than supplied coherently by the platform.
However, companies are repeatedly making all sorts of decisions about suppliers, customers, inventory, pricing, service, etc. Those decisions are constantly generating evidence of what works and what not, provided that the company is smart enough to capture them. What if we could evaluate AI by whether the company learns and improves, beyond the quality of individual outputs? What if AI optimization becomes more important than having access to bigger models?
The competition moves from model capabilities to system performance
A company does not just need a very powerful AI model: it needs an architecture that coordinates capabilities toward its goals. Researchers writing on Berkeley’s AI research blog explicitly describe the shift toward compound AI systems, where system design can offer better returns than further model scaling.
Other experiments, such as FrugalGPT or DSPy are specific precedents for making model selection subordinate to results: FrugalGPT demonstrated that model cascades can match the strongest individual model with up to 98% lower cost on its evaluated tasks, and DSPy showed that optimized pipelines using smaller models can compete with approaches using much larger proprietary models.
What makes the next step consequential? The ability to learn across executions. If we manage to make each interaction inform subsequent decisions across the relevant population, we can achieve a substantial advantage. And world models add precisely this possibility: they can anticipate consequences before acting. Research about Dreamer successfully demonstrates this mechanism across diverse control tasks. And once you have that, you can use reinforcement learning or another optimization mechanism uses experience to progressively improve decisions.
Accumulated organizational learning becomes the true competitive advantage
While broadly available models available through commercial APIs or open-weight deployments can provide shared capabilities, company-specific experience can produce truly differentiated decisions, the ones that can really make the difference. Imagine, for instance, the renewal decisions you can make when you can learn from observed retention outcomes.
However, bear in mind that simple accumulation does not automatically guarantee an advantage: you have to process the information properly, but not “with an AI”… with “your AI”. Make sure you put the training in the correct layer, the one you can administer and control.
PageRank unlocked value in something the web was already producing: links. The next breakthrough in corporate AI could unlock value in something companies already produce every day: experience. If that experience becomes accumulated learning that the company controls and retains as models change, the competitive advantage starts to belong to the business itself. The company that stopped waiting for a smarter model might simply be the first to realize that its most valuable intelligence is the intelligence it learns to build — and keep.
My proposed buying criterion is based on the following question: can the company replace its model provider while retaining its operational history, learned policies, objectives and demonstrated gains? Can your company export and reuse that learning when changing platform providers? Do you own your corporate intelligence, or do you rent the whole thing? When you stop working with “vanilla AI” and you start working with “your own AI”, you control the learning asset, without requiring ownership of every underlying model.
And that’s when you can become a truly optimizable company. Basically, optimization, not a bigger model, is what you need.