# AI cannot optimize a company it cannot understand > Source: > Published: 2026-08-31 10:30:00+00:00 For many weeks now, I have been trying to describe in my *Fast Company* essays what I think is the future of [corporate AI](https://www.fastcompany.com/91587826/enterprise-ai-baconian-approach-business). Now imagine that someone, somewhere, comes up with a solution that fulfills the requirements I have been drafting here. Would your company be able to jump on its platform and access all the advantages we have been describing? To start optimizing a company with [artificial intelligence](https://www.fastcompany.com/section/artificial-intelligence), AI needs to understand what you want to optimize; AI cannot properly optimize what it cannot represent. The paradox is that companies are giving increasingly sophisticated and capable models access to the tools and workflows they use, but they have only a fragmentary representation of the map of the organization they are trying to optimize. Even if a large language model is extremely knowledgeable about management, [marketing](https://www.fastcompany.com/section/marketing), or logistics (after all, LLMs have “read” pretty much every book about it), [it usually knows nothing about ](https://www.fastcompany.com/91555415/real-reason-enterprise-ai-stuck)your specific customers, your dependencies, your bureaucratic processes, any possible exceptions, your risk tolerance, or the cascading implications of changing a process. We need to separate plain memory from what constitutes a real data model; while memory can recover what happened at a certain time, a data model can formally represent identities, relationships, permissions, constraints, and valid states. Nowadays, companies try to bridge this gap by providing as much context as they can. The problem is that context answers certain things such as the information the model should consider at a certain moment, but it does not answer something as relevant as how the organization works. In a previous article, we moved from memory toward models. Now, let’s move from context towards a model of organizational dynamics. The initial step we need to take is [to define an ontology](https://www.fastcompany.com/91574442/next-enterprise-ai-frontier-is-the-optimizable-company), which is the same as providing a company with nouns, verbs, and rules. Palantir has been trying to convince a lot of companies about this not only because it is the right way to proceed, but also because, in order to do so, Palantir [sends its forward deployed engineers (FDEs)](https://www.fastcompany.com/91544792/why-big-ai-companies-embedding-engineers-customers-what-does-that-mean). Building ontologies manually is expensive and labor intensive, which helps explain both the importance of FDEs and the economics of Palantir’s model. Despite Palantir trying to own the term “ontology,” what we are simply talking about (or maybe not so simple!) is representing real-world entities as objects, their relationships as links, and the company’s actions as verbs. Basically, an operational layer, or a sort of digital twin of the whole organization. It is an architecture that couples “nouns” with “verbs,” in a way that things such as factories, orders, workers, or customers are in the first group, and concepts such as launching promotions, adjusting prices, changing distributors, or executing a workflow are in the second. However, an ontology just tells you what’s in the company and what we can do with it. The next step is a world model, a term typically associated with robotics and physical AI, but that [I find extremely useful in corporate AI architectures](https://www.fastcompany.com/91483469/this-next-big-thing-corporate-ai), too. A world model goes further: It tries to learn the dynamics behind that representation, including what is likely to happen next. In the AI research field, world model means an internal representation capturing the dynamics of an environment, sufficiently well as to allow predicting consequences and supporting planning. [Drawing from an example used by IBM](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.ibm.com_think_news_world-2Dmodels-2Dnext-2Dfrontier-2Denterprise-2Dai&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=8AKVu-zTZzRFkUc-mqmon0H1psziDtj44MBL47ac3g92Iuk6OBgBunui8tbS_mZO&s=NP4kjjeTKddtfEefsnpVWZShDAw-BIVEHVsBOeLSwgg&e=), one thing is to describe why a machine fails, and another is to predict what will happen if a certain preventive maintenance is not taken right now. A company needs to be able to understand that, for instance, forcing customer service operators to reduce the duration of their phone calls tends to increase churn, not just that shorter support calls mean lower costs. To be reliable, [algorithmic decision-making needs causal reasoning](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nature.com_articles_s43588-2D025-2D00814-2D9&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=8AKVu-zTZzRFkUc-mqmon0H1psziDtj44MBL47ac3g92Iuk6OBgBunui8tbS_mZO&s=hy1TqvRm_yZruHHp55W4pVrwD04nOxdEjlTeE9agXgI&e=), given that decisions always involve cause-and-effect relationships and we need to align them with real-world objectives. Dave Wright, a veteran chief innovation officer at ServiceNow, sees [huge potential in the application of digital twins of entire companies](https://urldefense.proofpoint.com/v2/url?u=https-3A__www.wsj.com_cio-2Djournal_an-2Dinnovation-2Dveteran-2Don-2Dwhats-2Dnext-2Din-2Denterprise-2Dai-2D1e20ead2&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=8AKVu-zTZzRFkUc-mqmon0H1psziDtj44MBL47ac3g92Iuk6OBgBunui8tbS_mZO&s=k7w_hGjO1hofaH1Es_mYiFA4lgoErvL8HRUGI2mSy48&e=) for simulating scenarios, regulatory effects, etc. Once you have your company model, you can make it more and more valuable by feeding actions and outcomes back into it. Capturing decisions and their outcomes allows future decisions to be framed in the context of previous choices, and can feed retraining or fine-tuning. That way, with these loops, the company starts to own its own learning, as [Satya Nadella said](https://urldefense.proofpoint.com/v2/url?u=https-3A__x.com_satyanadella_status_2066182223213293753&d=DwMFaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE&m=8AKVu-zTZzRFkUc-mqmon0H1psziDtj44MBL47ac3g92Iuk6OBgBunui8tbS_mZO&s=p0XbiV3Donc7NkuBTj3ItXhML7-pOCGtO30jGikCB1I&e=). Don’t be scared. I know organizations are complex, dynamic, and politically loaded. A representation like this doesn’t have to be perfect, and the model, of course, will have to remain partial, open, revisable, and governed. The goal is not to build some omniscient oracle, but a representation that improves across time. The idea is to turn the company’s representation of itself into the durable asset. Two companies may use exactly the same frontier model, the same version of Claude, ChatGPT, Qwen, or any other LLM, but that model can increasingly become [just the engine operating the model](https://www.fastcompany.com/91590409/when-everyone-same-ai-what-makes-your-company-smarter) inside something much more valuable: the proprietary representation each company has built of itself, with its own context, rules, history, and accumulated learning. We need to move from the idea of [the optimizable company](https://www.fastcompany.com/91574442/next-enterprise-ai-frontier-is-the-optimizable-company) to a very interesting one: the understandable company, because before a company can become optimizable, it has to become understandable for the intelligence trying to optimize it. We spent decades trying to digitize companies in a way that software could record what they do. The next step will be to model them so that AI can understand what its actions mean. Soon, these types of tools will start to be readily available, and the question will become whether or not your company is prepared to apply them. Time to get ready.