Why AI alone is not enough for critical decisions

Ask three AI assistants what last month’s margin was, on the same data, and you can get three plausible, different answers: 18.4%, 21.1%, 16.9%. It is not a model problem: the definition of “margin”, the rule for what to include, the logic for how you choose are written nowhere. They live in people’s heads and in spreadsheets.

However powerful, an AI can only work with what it finds: if decision knowledge is not formalized, there is nothing to reason over reliably. That is why, before talking about models, it is worth looking at three layers a company must prepare.

First layer: formalize and document decision knowledge

Most critical decisions rest on tacit knowledge: experience, unwritten rules, implicit criteria that managers apply without spelling them out. Knowledge-management research has long framed this as the distinction between tacit and explicit knowledge, and the point is well known: what stays tacit is not transferable, neither to a new person nor to a system.

The first task, then, is not technological but cognitive: making explicit how you decide. What information enters a choice, which constraints and objectives govern it, which alternatives are considered and by what criterion you choose. Standards exist for this, such as the Decision Model and Notation (DMN), which describe a decision’s logic in a form readable by people and machines. It is slow, unglamorous work, but it is the ground everything else stands on.

Second layer: the semantic layer, defining the numbers once

With the logic formalized, you need data that means the same thing to everyone. This is where the semantic layer (or metrics layer) comes in, a now-established concept in data engineering: a layer where metrics, dimensions and relationships are defined once and hold for every tool and every user. “Margin”, “active customer”, “order” stop being interpretations and become single definitions.

For AI the benefit is twofold: it reasons on business concepts instead of tables, and it has an explicit scope that reduces hallucinations, because every answer is anchored to a shared definition and the source it comes from.

Third layer: the decision layer, from numbers to choices

Knowing what margin is does not yet say what to do. The third layer, the decision layer, connects metrics to objectives, constraints and the levers you act on: it is the territory of what analysts call decision intelligence. Here you make explicit the trade-offs (margin versus volume, cash versus growth), the thresholds beyond which a situation demands action, the alternative scenarios with probability and impact.

It is the layer that turns a dashboard into a decision: not “here are the numbers”, but “given these conditions, these are the options, with these expected effects”.

“ AI does not create the knowledge you decide with: it uses it. If it is not formalized, there is nothing to use. ”

Only now does agentic AI come in

With these three layers in order, agentic AI can make the leap that matters: no longer automating tasks, but supporting critical decisions. An agent that rests on formalized knowledge, shared definitions and explicit decision logic can simulate scenarios, flag when a threshold is crossed and propose an action with its source cited, leaving the choice to whoever is accountable.

It is also what the data suggests. In Oliver Wyman’s 2026 survey of 200 senior executives, only 4% cite access to the best model as a lasting advantage, versus 74% who point to reengineered processes and 64% to proprietary data and context. The advantage is not the model: it is the knowledge you let it use.

Where to start

You do not need a multi-year program to begin. The most concrete way is to pick a single critical, recurring decision, say when to review prices or when to replenish stock, and document it fully: the information that feeds it, the constraints, the criteria, the alternatives. From there you define the few metrics that matter, make the logic explicit, and only then put AI to support that decision, with a person who approves.

It is a path that starts from knowledge, not technology, and it is exactly this order that separates an AI that reduces uncertainty from one that adds noise.

KEY TAKEAWAYS

• Before using AI and agentic AI for critical decisions, three layers come first: formalized decision knowledge, a semantic layer, a decision layer.
• Much decision knowledge is tacit: it must be made explicit (information, constraints, criteria, alternatives). Standards like DMN help formalize it.
• The semantic layer defines the numbers once: AI reasons on business concepts and anchors every answer to a definition, reducing hallucinations.
• The decision layer connects metrics to objectives, constraints, trade-offs and thresholds: it turns a dashboard into a decision (decision intelligence).
• Only after these layers can agentic AI support critical decisions. In the data, only 4% see the model as a lasting advantage, versus 74% processes and 64% proprietary data and context.

FREQUENTLY ASKED QUESTIONS (FAQ)

What does a company need before using AI for critical decisions?

Three things, in order: formalize and document how it decides (decision knowledge), define metrics in a shared way (semantic layer), and make the choice logic explicit with objectives and constraints (decision layer). Only then does AI have something reliable to work with.

Why is formalizing decision knowledge so important?

Because most of it is tacit and not transferable. What is not explicit cannot be used or explained by an AI system, which would end up reinventing criteria and definitions every session.

What is the semantic layer and what does it have to do with AI?

A layer where metrics and definitions are set once and hold for every tool and agent. It gives AI clear business concepts and an explicit scope, reducing errors and hallucinations.

When should you introduce agentic AI into decisions?

After preparing the three layers, starting from a single critical, recurring decision, with the agent proposing and a person approving.