One version of the truth: what it costs and what it is worth

What the technology changed and what it did not

Microsoft Fabric puts every analytics workload over one logical lake, with data held in open table format so each workload reads the same copy rather than maintaining its own extract. Power BI reads it directly through Direct Lake without an import step, removing the refresh window that made reporting stale before anyone read it.

These are genuine architectural improvements and they solve the problem institutions least often have. The problem institutions most often have is that finance, risk and the front office each define margin differently, and each is defensible.

A single copy of the data does not settle that. It makes the disagreement more visible, which is useful, and it does not resolve it.

Where the value actually is

We would put it in three places, in order.

First, traceability. In a regulated institution the ability to trace a published figure back through every transformation to the originating record converts an internal report into something a board committee can act on. That capability is architectural and it is difficult to retrofit.

Second, decision latency. Not query speed — decision speed. A figure that arrives continuously rather than monthly changes which decisions are possible, and the value is entirely in what somebody does differently as a result.

Third, and least, analyst productivity. It is real, it is the easiest to model, and it is the argument most likely to be disputed by anybody who has seen a previous programme fail to deliver it.

• Lineage from published figure to source transaction, inspectable by a third party
• Reconciliation to the general ledger built into the pipeline rather than performed afterwards
• Row-level security sourced from the directory so entitlement is evidenced, not asserted
• Definitions held once, in the model, so every report inherits them
• Decision latency measured in days rather than reporting cycles

The cost most programmes underestimate

Not engineering. Agreement.

Every institution we have worked with has underestimated the elapsed time required to get finance, risk and the business to sign up to one set of definitions. It is unglamorous, it involves people whose incentives differ, and it cannot be delegated to the data team without producing a platform the business disputes.

The second underestimate is capacity cost. Consumption-based compute is efficient when designed well and expensive when not, and the feedback arrives on an invoice a month later. An inefficient notebook or an over-refreshed model costs money in a way a perpetual licence never did.