← All field notes

Entity resolution · Oct 2026 · 10 min read

Why Your Metrics Disagree

It usually is not a BI problem. It is an unresolved model problem showing up as a reporting fight.

02

The meeting

Four teams, four correct numbers.

Sales enters the meeting with one revenue number. Finance has another. Operations has a coverage number that does not match either. Leadership asks which dashboard is wrong, and the meeting becomes a negotiation over exports. Everyone has evidence. Nobody has authority.

The usual diagnosis is tooling. Move the reports to Power BI. Rebuild the Looker model. Replace a legacy dashboard. Tighten the spreadsheet process. Those moves may improve access, performance, or presentation, but they rarely fix the root cause.

The disagreement usually starts before analytics. It starts when the business has not resolved what its core objects are, how they change over time, and which rule wins when two systems describe the same thing differently.

03

The real source

Metric disagreement is often entity disagreement.

An outlet is not just an outlet. In one system it is a billing account. In another it is a physical location. In a distributor feed it is a free-text name. In finance it may be a parent account. In field operations it may be split by route. If those records are not resolved into one governed entity, every metric built on top inherits the disagreement.

The same pattern shows up across domains. A dealer moves territory mid-quarter. A candidate has duplicate profiles with different skill histories. A lead is counted by call source in one system and by campaign in another. A service job is open operationally but closed financially. None of these are dashboard defects. They are rule defects.

This is why entity resolution is not data cleanup. Cleanup suggests a one-time exercise. Resolution is operating infrastructure. It is the layer that decides whether two records describe the same real-world thing, which attributes are authoritative, what time period a rule applies to, and what exception must be reviewed instead of silently averaged away.

04

The fix

Build the truth layer before the reporting layer.

The practical order is unfashionable because it is less visible at first. Name the entities. Define the relationships. Specify the temporal rules. Capture lineage. Make exceptions visible. Then build reporting, automation, and AI on top of the governed model.

In retail execution work, this means an audit event belongs to a store, territory, product, planogram, cycle, and evidence set. In dealer operations, a service job belongs to a vehicle, customer, dealer, warranty state, and parts flow. In marketing intelligence, a lead belongs to a location, channel, campaign, source event, and conversion state. The specifics differ. The principle does not.

Once the model exists, the dashboard has something stable to query. When a number changes, the system can tell you whether the change came from a new source file, a rule update, an entity merge, a late-arriving event, or a manual override. That is the difference between reporting and operational intelligence.

05

Use this

Five checks before blaming BI

01

Do the systems agree on the entity being counted?

A store, account, outlet, dealer, lead, or customer may mean different things in different systems.

02

Are temporal rules explicit?

Territory, ownership, status, and revenue recognition often depend on when the question is asked.

03

Can you trace a metric back to source events?

Without lineage, every disagreement becomes a manual investigation.

04

Where do exceptions live?

If corrections live in spreadsheets, the governed system is not the system of truth.

05

Who owns the definition?

A metric without an owner will eventually become several metrics with the same name.

Talk it through

Does your system look like this?

Fifteen minutes with an engineer, not sales. Bring the architecture, leave with a straight read on it.

Book a 15-min architecture call