The problem is rarely that the numbers are wrong. It is that each is correct for a different purpose, and nothing makes that context visible. That gap is what erodes trust today, and it is the single biggest barrier to trusting AI tomorrow.
One question in the quarterly review: “What was Q3 revenue?”
CRM · bookings
ERP · recognised
Board deck
Region · shipped
↓ each correct for a different purpose, until the context is made explicit
One governed, defined figure
Most large organisations can reproduce this scene on demand.
A review meeting opens, someone asks a simple question, and several teams answer with different figures for the same metric. The reflexive diagnosis is that someone’s data is wrong. It usually is not.
In most cases, every figure is correct. They differ because each was produced to answer a different question, under a different definition, at a different moment. The difficulty is that nothing in the organisation makes those differences visible, so a set of individually correct numbers appears to be a contradiction. That is the condition worth naming precisely: not messy data, but absent context. It is also why, in Precisely and Drexel University’s 2025 study, 67% of organisations reported they do not fully trust the data they use to decide, a figure that had risen sharply in a single year. Trust is not falling because data is getting worse. It is falling because data is fragmenting faster than its meaning is being agreed.
Consider a single word: revenue. Finance reports it on a recognition basis. Sales reports bookings. A regional unit may report the shipped value. Each is correct for the accountability it serves. The word customer fractures the same way: a signed account, a billed account, and a delivered account are three different populations, and each team is right to count the one it is measured on.
The instinct is to force a single definition and declare it the truth. This is the first mistake. An imposed definition does not resolve the disagreement; it makes one team’s number wrong for the purpose that team is answerable for, and pushes the reconciliation into a private spreadsheet where it can no longer be seen or governed.
The change that holds is smaller and more durable: make the context of each number explicit. Every figure that matters should carry its definition, its grain, its time basis, the scope it covers, the decision it serves, and the person who owns it. When that context is visible, two different numbers stop reading as a contradiction and start reading as two correct views of the same reality. The argument ends, not because everyone was forced onto one figure, but because everyone can see why the figures differ and which one applies to the decision in front of them.
This is also what a single source of truth genuinely is, and why it remains the right goal. It is not a single number that erases the others, and it is not a single warehouse into which everything drains. It is a governed place where the meaning and context of each number are defined, owned, and reconcilable. Pursue the physical store without the meaning, and the disagreements survive the migration intact.
Faced with fragmentation, most programmes consolidate: pull the sources into one platform, publish one figure, and retire the rest. Consolidation is not the error. But carried out without moving the context along with the data, it can leave governance weaker than before, for three reasons that are rarely stated plainly.
The differences between systems were themselves information. They told a reader why finance figures and sales figures were not the same. Collapse them into a single published number, and that derivation vanishes. What remains is one figure that no one can trace back, and an untraceable number is trusted less, not more, however firmly it is declared authoritative.
Moving data into one location is a plumbing achievement, not a governance one. Governance is the definitions, owners, and lineage that give a number meaning. Consolidation often moves the bytes and leaves the meaning behind, producing a tidier estate with no more agreement inside it than the one it replaced.
One pipeline feeding every report is also a single point of failure. When domain ownership is not preserved through the consolidation, no one remains accountable for whether a given number is correct for a given use. Central storage, diffuse responsibility, the worst of both.
The way through is to treat consolidation as an act of preserving context, not removing it. Three principles hold in practice. Surface lineage at the point of use, so every number is traceable to its source and transformations where it is read, not buried in a catalogue no one opens. Unify the data physically while keeping definitions and domain ownership federated: the finance team still owns revenue; the platform makes that definition discoverable and enforces it rather than overwriting it. The concrete instrument for this is a data contract: the domain that owns several commits, explicitly, to the definition, structure, and quality of what it publishes, so every team downstream, and every AI agent, consumes it on agreed terms rather than by assumption. And agree meaning before moving data, consolidating definitions is the governance act, moving storage is the plumbing that should follow it. Reverse that order, and the result is an expensive platform that still cannot answer a simple question the same way twice.
Everything above was survivable at human speed. A capable analyst reading two dashboards reconciles them silently, drawing on context held in their head. That quiet act of human reconciliation is precisely what an AI system cannot perform.
An agent asked for revenue does not weigh two definitions and choose the right one for the decision. It takes the definition it is handed, computes, and acts. The ambiguity a person used to absorb becomes an automated decision, made at scale, without the pause in which a human would have noticed the mismatch. This is why the semantic layer, the place where an organisation’s terms are defined once, stops being reporting hygiene and becomes the control surface for AI. It is where meaning is enforced before an agent acts on it. An assistant is only ever as reliable as the definitions beneath it; placed over a fragmented foundation, it does not resolve the fragmentation, it automates it.
Maturing a semantic layer is a specific piece of work, and it does not begin with the data model. A practical sequence:
01. Start from decisions, not tables.
Identify the small number of metrics that actually drive decisions and are most often disputed. Meaning has to be agreed there first; the rest can follow.
02. Define each metric with its full context.
Definition, grain, time basis, valid and invalid uses, and a named, accountable owner. A definition without an owner decays within a quarter.
03. Encode the definitions once, where both reporting and AI read from them.
So a dashboard and an agent compute the same metric identically, rather than each carrying its own quiet version of it.
04. Make definitions and lineage queryable.
A person or an agent should be able to ask not only what a number is, but what it means and where it came from, the difference between a system that can be trusted and one that merely answers.
05. Keep it living.
Version it, review it when the business changes, and re-test it against real and adversarial questions, ‘active customer by which definition?’ before any agent is permitted to act on it.
The hard part of this sequence is rarely the technology; it is the agreement itself. When Sales and Finance count a customer differently, neither team is being difficult; each definition is tied to the targets they are measured on and the decisions they answer for, so holding onto it is reasonable rather than obstructive. Overruling those differences from the outside tends only to push the reconciliation back into private spreadsheets. Treating them as a conversation between owners works better: a standing forum where the domains that produce and use a metric agree on its definition together, each metric given a named owner, with visible backing from leadership, so the agreement holds once the meeting ends. The aim is not to crown a winning number, but to reach a shared definition each team can still use for its own purpose, and to record why it was set that way, so the next team inherits the reasoning instead of reopening the debate.
On a Microsoft estate, this is the work of a unified data layer and a governed semantic model, OneLake and Fabric for the foundation, a shared semantic model for the definitions, Purview for ownership and lineage. The tooling matters, but it changes; the discipline of defining meaning once and carrying it everywhere is what endures.
A single source of truth is the right goal. Its mature form is a single source of agreed meaning and traceable lineage, not one figure that erases the context that made the others legitimate.
That is the foundation decisions can stand on, and the only foundation on which an AI agent can be trusted to act. For data and BI leaders weighing a consolidation programme, or preparing to place AI over an estate whose definitions have never been agreed upon, the meaning is where the work begins.
In comparison to the usual decision-making process, the change is evident at first sight.
| The Decision | Traditional Approach | Where This Points |
|---|---|---|
|
Data consolidation |
Centralise everything into one physical warehouse to force a single number. |
Federate ownership while centralising definitions; carry context and lineage through the move. |
|
Metric definition |
Set by IT or data engineering, from source-system tables. |
Set by the business domains, from the decisions they drive, with the context made explicit. |
|
AI readiness |
Point AI at the warehouse and trust it to infer what the numbers mean. |
Build the semantic layer as the control surface, so agents compute metrics exactly as the dashboards do. |
|
Trust in data |
Measured by data-quality scores: nulls, duplicates, completeness. |
Measured by traceability, explicit context, and agreed meaning. |
Expert Extend Capabilities
Expert Extend’s work on this problem is to make the context of every important number explicit, to preserve ownership and traceability through consolidation rather than lose them to it, and to build the semantic layer as the control surface that both reporting and AI depend on. On the Microsoft platform, with the governance carried through, not left behind in the estate it replaced.
– In one DACH insurance program, the reporting cycle transitioned from one to three months to same-day delivery.
Not through the introduction of new tools, but by first establishing definitions and ownership.
– In a post-migration operational overhaul for a hotel group, the result was a single, real-time overview that the operations team could rely on each morning, as the figures behind it finally conveyed a consistent message.
Navigation