Why AI Cannot Save an Enterprise That Doesn't Understand Its Data

# Why AI Cannot Save an Enterprise That Doesn’t Understand Its Data Source: [https://architectureintel.com/why-ai-cannot-save-an-enterprise-that-doesnt-understand-its-data-83613f209317?gi=472f1cf46403](https://architectureintel.com/why-ai-cannot-save-an-enterprise-that-doesnt-understand-its-data-83613f209317?gi=472f1cf46403) [](https://architectureintel.com/?source=post_page---byline--83613f209317-----------------------------------------) Press enter or click to view image in full size > This article was co\-authored with[Mustapha Fonsau](https://www.linkedin.com/in/mustapha-fonsau/), CIO at Talentys, creator of Arca Suite, and Sovereign Decision Intelligence & AI Strategist\. A board reviewing $37 million of investment exposure needs more than confidence that the figure is accurate\. It needs to understand what the exposure comprises, how the components connect to each other and whether those connections create dependencies that warrant action\. A report can account for every dollar and still leave all of it open\. The number is rarely the problem\. Organizations close that gap with people\. An analyst supplies the context the report leaves out\. An architect spots a dependency between two seemingly unrelated activities\. An executive remembers why an exception was approved four years ago\. Little of that reasoning survives the meeting, and AI does not remove the need for it\. It removes the pause in which it used to happen\. ## Thirteen Distinctions, Not Four The data–information–knowledge–wisdom model, usually attributed to Russell Ackoff, distinguishes between having data and exercising judgment, and most discussions start and stop there\. An enterprise needs something it can inspect decision by decision, and that takes finer distinctions: **Data → Context → Information → Composition → Relationships → Meaning → Knowledge → Understanding → Judgment → Decision → Action → Outcome → Learning** The reasoning can break down at any of these arrows\. This is not a sequence to be followed step by step, since experienced practitioners often identify the structural risk before detailing the individual components, but each distinction raises an operational question that would otherwise remain unaddressed\. ## When 1 \+ 1 \+ 1 Does Not Equal 3 Consider the number 37\. On its own, it is merely data\. Say that it represents millions of dollars of investment exposure: that adds context\. Add the reporting date, the currency and the scope, and it becomes usable information\. Its composition is three positions: $15 million with Counterparty A, $12 million with Counterparty B and $10 million with Counterparty C\. We can explain what the total contains and still know nothing about what could affect several positions simultaneously\. Suppose that A and B depend on the same critical supplier\. That single relationship ties $27 million, roughly 73 per cent of the total, to one point of failure\. The arithmetic has not changed, but the meaning has, because what appeared to be three separate positions holds a concentration\. Meaning, here, is what a relationship signifies once it is read through the organization’s own concepts and obligations\. Knowledge and understanding diverge at this point\. Knowledge is meaning backed by dated, sourced evidence and available to someone other than the person who discovered it\. Understanding is an explanation of why the concentration formed, whether through an acquisition, a supplier consolidation, or a series of unrelated commercial decisions made by people who never spoke to each other\. Knowledge tells you that the dependency exists\. Understanding tells you what to do about it\. Judgment rarely waits for the full causal picture\. Suppose nobody knows whether the supplier could absorb a disruption, and Counterparty A is strategically important: management reduces the position with B by $5 million and asks A for a contingency plan within 90 days\. The assumptions, the uncertainties and the conditions that would justify reopening the file are as important as the decision itself\. Then limits are adjusted, and a monitoring rule starts flagging exposure concentrated on a shared supplier\. ## Technically Flawless, Semantically Wrong Ontology is part of the structure that holds this together\. In Tom Gruber’s formulation, an ontology is “an explicit specification of a conceptualization”: what counts as a legal entity, how a contract creates an obligation, how several entities can depend on one supplier\. None of this starts from a blank page\. In finance, the EDM Council’s FIBO already formalizes legal entities, contracts and ownership, and BCBS 239 asks banks to establish integrated data taxonomies and consistent identifiers for legal entities, counterparties, customers and accounts\. Every enterprise already operates on such distinctions, whether or not they are formalized in an ontology\. Sales treats a customer as a commercial relationship, Finance as a billing account, Legal as a contracting entity\. All three are correct, and software designers recognize this as Eric Evans’s bounded contexts\. Forcing a single universal definition destroys the distinctions that people need in order to do their jobs\. The job of the enterprise, and of its architects, is to build the translation between those boundaries on purpose\. ## GetYounss’s stories in your inbox Join Medium for free to get updates from this writer\. Remember me for faster sign in When that translation is missing, nothing announces it\. An interface passes an identifier cleanly while the receiving application treats it as a different kind of thing: a booking entity where the receiving system expected a settlement entity, or a group treated as one of its subsidiaries\. Every technical check passes\. The failure surfaces cycles later, as a reconciliation that keeps needing manual adjustment and an exception queue growing where nobody looks\. The bill for semantic ambiguity is rarely a loss event\. It is a line that never disappears, large enough to irritate an executive committee and diffuse enough that nobody owns it\. ## AI Removes the Pause AI raises the cost of getting this wrong\. One assistant reads the exposure correctly and misses the shared supplier\. Another identifies the dependency and tests it against a policy that was replaced last quarter\. When a recommendation triggers a workflow, an interpretation becomes an action before anyone reviews it, and a fluent explanation proves nothing about which concepts, sources and rules were used\. A knowledge graph makes the relationships reachable; it does not ensure they are used correctly\. Facts need evidence and dates, policies need an owner and a scope, and permissions, validation and escalation have to live inside the workflow rather than in a governance document\. ## The Rule Nobody Rewrote Return to the $37 million\. Six months later, the reduction has been completed and the reported concentration is $22 million\. Then Counterparty C is acquired by a group whose operations run on the same supplier\. As the monitoring rule covers only direct supplier contracts, the dashboard does not change\. In reality, the dependency may now cover the full $32 million, and the condition recorded for reopening the decision has arguably been met without anyone noticing\. The number is correct again, and the interpretation is not\. Adjusting the limit would not solve the problem here\. The definition of dependency has to cover ownership and control, and the evidence gathered has to change accordingly, which is what Argyris and Schön termed double\-loop learning\. If the revision remains in a meeting note while the systems continue to apply the old rule, the organization has learned nothing it can act on\. Decision records help, provided they link evidence, assumptions and reasoning to what happened afterward\. ## The Practical Test Start with one decision that incurs real costs and keeps requiring manual interpretation or reconciliation\. Bring together the people responsible for it, the domain specialists and the teams who build and run the systems, and work through actual cases\. Decide which concepts, evidence and rules apply, then implement the result in the interfaces, the data structures and the controls\. Judge it by the work it improves\. So what is a corporation without an ontology? There is no such thing\. Every company operates on a conceptualization of what exists and how it connects; what most lack is an explicit specification of it\. For anyone about to give AI more autonomy, the test is simple\. Can you explain how your systems move from a record to an interpretation, from that interpretation to an action, and from the outcome back to a revised understanding? Wherever the answer is no, automation is carrying authority that nobody can supervise\. At that point, the question becomes a matter of leadership: who is running the company?