The Margin LabsThe Margin Labs

Case study

Unified KPI Command Center

Led by NunoData
  • 6 sources unified
  • 4 entities consolidated
  • Hours to on-demand
  • One-request snapshot

Challenge

Key metrics lived in six spreadsheets owned by different people on different cadences. Consolidated reporting was a manual weekly grind with no on-demand view.

What was built

A KPI intelligence layer: specialised skills pulling live data from all six sources, covering supply chain, finance, service and growth across four trading entities, orchestrated into one view.

The story

How it ran.

The leadership team had a reporting rhythm that worked when the group was smaller and broke under growth. Six spreadsheets covered supply chain, finance, customer service and growth. Each one had a different owner, a different cadence and a different definition of the headline numbers. Friday afternoons were spent reconciling, and by Monday the reconciliations were already stale.

Before any build, we ran a trust audit on the underlying numbers. For each KPI we asked the same three questions: who owns the source, how is it calculated and how often does it change. About a quarter of the metrics had no single owner. Another set were calculated two different ways depending on who was asked. We resolved those disagreements before we wrote any code, because a dashboard built on contested data has a half-life of weeks.

The unified-view design came next. Rather than collapsing everything into a single warehouse, we built a thin orchestration layer that pulled live from each source on request. The advantage: the source teams kept ownership of their data and their cadence. The reporting layer never went stale because it never had its own copy.

Each source got a specialised skill: one for the ERP, one for the eCommerce stack, one for the service desk, one for the warehouse system, one for the ad accounts and one for the finance ledger. Each skill knew how to translate the source's idea of an entity into the unified view's idea of one. The four trading entities rolled up cleanly because the entity normalisation lived in the skill layer, not in human heads.

Orchestration was where the leverage compounded. A single request from the leadership channel triggered the relevant skills in parallel, waited for the slowest, and assembled the snapshot. The same plumbing powered the weekly review, the monthly board pack and the ad-hoc question. The grind disappeared because nothing was being rebuilt; everything was being assembled.

Six months in, the most-used surface was the simplest: a one-line request from the leadership channel that returned the snapshot in under a minute. The dashboard the team had originally asked for got built and barely used. The lesson was clear: on-demand beats scheduled when the cost of asking is low enough.

Methodology

The sequence we ran.

  1. 1

    Source mapping: enumerate every source, its owner, its cadence and its definition of each entity.

  2. 2

    Trust audit: resolve calculation disagreements before building anything.

  3. 3

    Unified-view design: define the entities and metrics the leadership team actually decides on.

  4. 4

    Orchestrated pull: build thin skills per source; assemble the snapshot on request.

Architecture

What sat behind it.

Sources

  • ERP general ledger
  • Shopify storefront
  • Gorgias service desk
  • Warehouse management system
  • Meta and Google ad accounts
  • Finance ledger and consolidation

Integration

  • One skill per source with native API access
  • Entity normalisation across four trading entities
  • Provenance and freshness on every record

Orchestration

  • Parallel skill invocation with timeout handling
  • Snapshot assembly with reconciliation checks
  • Cache layer for repeat requests inside short windows

Surface

  • Slack one-request snapshot
  • Weekly review template
  • Board-pack assembly view

Lessons

What we would carry forward.

  • Source-of-truth discipline beats clever pipelines. Resolve definitions before you write a line of code.
  • On-demand beats scheduled when the cost of asking is near zero.
  • Keep ownership in the source. The unified view stays trusted because it never holds its own copy.
  • Entity normalisation belongs in the skill layer, not in human heads.

Know exactly where AI pays before you commit a budget.

Start with the free AI Profit Roadmap. We map the highest-leverage automation and systems work in your business, with no obligation.