One number, five definitions
Exposure, customer, default, liquidity, adjusted EBITDA. The most important numbers in a financial institution have several legitimate definitions. AI makes choosing the right one an engineering requirement.
By Penomic Research · Published · 6 minute read
Download the PDFThe meeting where the numbers do not match
Most senior leaders have sat in a meeting where risk, finance and the business present different figures for the same thing. Usually nobody is wrong. Each function uses a definition that is correct for its purpose: regulatory, accounting, internal management or client reporting.
Data teams have spent years reconciling these differences in warehouses and semantic layers. The reconciliation tells you that the numbers differ and by how much. It rarely tells a system which definition a particular question requires.
Why AI forces the issue
When a person asks a colleague for exposure to a borrower, the colleague asks what it is for. A model usually does not. It picks a plausible definition, often the one that appears most often in its context, and answers confidently.
For a natural-language interface to be safe in a financial institution, every important concept has to carry its context: which definition applies for which purpose, owner and jurisdiction, from which date, using which sources and calculations.
- Purpose: regulatory, accounting, risk management, client or board reporting.
- Scope: legal entity, business line, product and jurisdiction.
- Time: effective dates and the version in force at a given point.
- Lineage: the sources, transformations and adjustments involved.
- Ownership: who can change the definition and who must be told.
From reconciliation to resolution
The goal is not one golden definition. It is a system that resolves each question to the right definition for its purpose before any data is retrieved, and shows which one it used. That turns a reconciliation exercise into a governed capability that analysts, applications and agents can all rely on.
Where to start
List the ten numbers that cause the most disagreement at executive level. For each, document the legitimate definitions, their purposes and owners, and link them to the data that implements them. Then test a natural-language query layer against real executive questions and check that it chooses and cites the right definition each time.