Over 40% of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls (Gartner, 2025). In most cases the limit isn’t the model: it’s that the company asks agents to decide before making explicit the definitions, objectives and constraints those decisions rest on.

Ask three AI assistants what last month’s margin was, and you may get three different numbers: 18.4%, 21.1%, 16.9%. None of them did the math wrong. One simply sums ERP invoices, one uses CRM orders, one reconstructs from inventory. Same data, three implicit definitions of the word “margin”.

This is the distance between AI-ready data and AI-ready decisions, and it widens with agents: McKinsey (2025) finds that fewer than one company in ten takes agents beyond the pilot, while eight in ten cite data as the main obstacle to scaling. Making data AI-ready is necessary; turning it into decisions takes two distinct layers: a semantic layer and a decision layer.

Why agents give three answers to the same data

Agents access the same data, but they don’t know the same rules. The word “margin” is written nowhere in the systems: it has to be reconstructed each time from tables, joins and implicit conventions. This is where three agents, or three sessions of the same agent, diverge.

The issue isn’t data quality, but the grain and semantics with which it’s read. A table called orders doesn’t say whether revenue is gross or net of returns, whether it’s recognized at order or shipment date, whether the currency is already converted, which time zone closes the “month”. Faced with this ambiguity, the model picks a plausible join and a plausible convention, and does so differently on each run. McKinsey (2025) puts it plainly: without a shared semantic foundation, “agents may act on incomplete or conflicting interpretations of identical data”.

A more powerful model doesn’t close this gap: it makes it more insidious, because it produces more fluent, more confident answers on top of the same ambiguities. In a multi-agent system the effect multiplies: each agent propagates its own definition downstream, and coherence between decisions is lost before it even reaches action.

What a semantic layer is

A semantic layer is a catalog of approved business definitions built on top of the data that already exists. Each metric is declared once (what it means, which tables it rests on, with which joins, at which time grain and which filters) and from there it becomes the single version people and agents read.

In practice it separates two planes that today get confused. Measures (gross margin, net revenue, average ticket) and dimensions (time, customer, product, region) are defined in the layer; the query that computes them is generated deterministically from that definition. The agent no longer writes the SQL: it selects measures and dimensions by name, for example “gross margin by region, last 4 quarters”, and the layer always emits the same SQL, with the same join paths and the same grain. Less reconstruction means fewer metric hallucinations.

This also moves governance to the right place. With a protocol like MCP (Model Context Protocol), agents discover and call governed metrics by name instead of querying raw tables. Permissions (role, region, tenant) enter query generation, not downstream: an agent cannot build a query that returns rows it shouldn’t see. One definition, one control point, one place to update it: change the rule once and every agent inherits it.

Why a semantic layer is necessary but not sufficient

With a semantic layer, the margin question has a single answer: 19.6% gross margin, calculated on ERP invoices, returns excluded by definition. The figure is now correct, shared and traceable. But knowing the margin doesn’t yet say what to do to protect it.

A shared definition removes ambiguity from the data. It doesn’t touch ambiguity in the decision: buy raw materials now or wait, raise the price or hold it, move production or not. These choices depend on objectives, levers, constraints and thresholds that don’t exist in the semantic layer. Understandable data is not yet a governable decision: that’s why you need a second layer, built on the same definitions.

What a decision layer is

The decision layer is the level that connects approved metrics to objectives, levers, constraints and models: it’s where the company makes explicit how it decides. It rests on the semantic layer, because every KPI calls an approved metric, so there are no divergent definitions between what is measured and what is decided on.

It holds four elements. The economic objectives and KPIs to protect or maximize, for instance a period margin above a threshold. The controllable levers (price, quantity, timing, allocation), the variables the company can actually act on. The constraints and thresholds: cash, capacity, coverage, minimum margin, beyond which an alternative is not admissible. And the models that tie it together: predictive (what will happen), causal (the effect of a lever on the outcome) and optimization (the best combination under the constraints).

On this basis the agent invents neither objectives nor constraints: it builds admissible scenarios, evaluates them against the same KPIs, and returns them with a probability of realization, a confidence interval and an expected impact on margin, not a point figure passed off as certainty. It proposes the alternative with the best expected value and cites the metrics, model and constraints used. This is the discipline Gartner (2025) classifies as “transformational” under the name Decision Intelligence: modeling decisions as assets, to compare and improve them over time.

How a decision’s lifecycle is governed

Definitions and decisions follow the same three-state lifecycle: proposed, approved, in drift. The agent writes the draft: a metric, a threshold, a new driver in the KPI tree. A business owner promotes it: from that moment it’s the approved version everyone reads. When the source table changes, or observed results drift consistently from expected ones, the definition goes into drift and back to review. Editing an approved definition returns it to “proposed”.

Drift is not a manual check but a continuous measure: the decision taken is compared with the observed outcome, and the model is updated. Every decision carries its governance metadata (owner, objective, perimeter, maximum amount, log) and lives within that perimeter. The agent proposes; beyond the amount threshold or outside the perimeter, a person decides.

This is the principle “the AI proposes, the human responds”. Trust doesn’t come from the AI explaining an answer after the fact, but from every decision being traceable before it’s made and its impact measured afterwards. An agent, in effect, operates like a new hire: with objectives, a perimeter and a record of what it has decided.

How it fits onto existing systems

The two-layer architecture doesn’t ask you to rebuild the data stack. It reads from the sources the company already has (ERP, CRM, data warehouse) and from the external data that matters for the decision, such as raw-material prices, exchange rates and demand, passing them through the semantic layer. The two layers are added on top; they don’t replace anything.

That’s why the starting point isn’t a multi-year program but a circumscribed decision: a purchase, a price, a stock level. You define the metrics you need, model that decision, measure its value, and from there extend to other decisions that share the same definitions. It’s an enterprise-grade approach accessible to the mid-market and to industrial SMEs, too.

The context pushes in this direction: Gartner (2025) expects that by the end of 2026, 40% of enterprise applications will embed task-specific agents, up from less than 5% in 2025. The question is no longer whether agents will make decisions, but on which definitions and under what governance.

KEY TAKEAWAYS

• Over 40% of agentic AI projects will be canceled by 2027 (Gartner, 2025): the limit is often not the model, but decisions left implicit.
• Agents diverge on the same data because they share the tables, not the definitions, joins and time grain.
• A semantic layer defines each metric once and generates deterministic queries: the data becomes correct and readable by people and agents.
• A semantic layer is necessary but not sufficient: a decision layer adds objectives, levers, constraints and models, and makes the decision modelable, comparable and measurable.
• Trust comes from traceability: the agent proposes within a perimeter, the owner approves, the outcome is measured and the model is updated.

FAQ

What is a semantic layer?

A semantic layer is a catalog of approved business definitions built on top of existing company data: it sets once what each metric means, which tables it rests on and how sources connect, so that people and AI agents read the same definition instead of rebuilding it on every query.

What’s the difference between a semantic layer and a decision layer?

A semantic layer defines what the data means and makes it understandable; a decision layer connects those metrics to objectives, levers, constraints and models to determine which decision to make. The first makes the data correct, the second makes the decision modelable, comparable and measurable.

Why do AI agents give different answers to the same question?

Because they access the same data but not the same definitions. Without a shared semantic foundation, each agent resolves ambiguities (joins, returns, periods, currencies) plausibly and differently every session. A more powerful model doesn’t remove the ambiguity: it hides it under more confident answers.

What is Decision Intelligence?

Decision Intelligence is the discipline that models decisions as assets, connecting data, models and constraints to make every choice explicit, traceable and improvable over time. Gartner classifies it as a transformational technology in its 2025 AI Hype Cycle.

Do you need to rewrite company systems to adopt a decision layer?

No. A decision layer fits onto existing systems (ERP, CRM, data warehouse and external data), reading them through the semantic layer. You start from a circumscribed decision, measure its value and scale, without rebuilding the data infrastructure.

How do you maintain trust in decisions made with AI?

By making every decision traceable: the agent proposes within a defined perimeter, an owner approves, and the outcome is compared with what was expected. What matters isn’t that the AI explains an answer, but that the decision is governed before and measured after.

WhAI: the two-layer architecture

This architecture is what WhAI, Vedrai’s Decision Intelligence platform, is built on. The semantic layer defines metrics once and makes them readable by people and agents; the decision layer connects them to objectives, levers, constraints and models, so that every recommendation arrives with its source, its confidence interval and its expected impact on margin. The agent proposes, the owner approves within a defined perimeter, and every decision stays traceable and comparable against the observed result. It’s the shift from AI-ready data to AI-ready decisions: not a faster answer, but a decision the company can govern, defend and improve over time.