Before AI, a Shared Definition of the Truth

The company’s biggest operational problem was downtime with no recorded cause. Lines stopped, and nobody could consistently say why. Much of the equipment was decades old and couldn’t be instrumented, so there was no sensor data to fall back on.

Leadership was under pressure to do something with AI, but there was nowhere to point it. The data lived in the ERP and in spreadsheets, with no shared model of what any of it meant.

Board reporting ran on a manually built spreadsheet. It was the number of record, but it depended on one person knowing where the data lived and how each figure was calculated.

We started where our thesis says to start: standard work before AI work. Rather than bolting an AI tool onto data nobody trusted, we built the foundation the AI would need.

We ran a 90-question scoping assessment across operations, finance, and IT. Then we mapped 57 ERP objects into a data crosswalk showing how every business entity related to every other. That map became the blueprint for the platform and for settling which definitions were correct.

We compared three paths to AI enablement (Snowflake Cortex, Microsoft Fabric, and direct model APIs) on total cost of ownership. We recommended against a custom web application in favor of platform-native capability. The smaller build was the right call for the client, and we said so.

Phase one put the core of the platform in production. The full build was scoped for later phases.

• A production ERP-to-Snowflake pipeline, with the highest-priority objects ingested first
• A Power BI board dashboard reconciled line by line against the manual report, so leadership could trust the new numbers before retiring the old ones
• A unified data model, so any AI tool works from one consistent definition of the business
• Snowflake Cortex provisioned with scripted access controls, a governed path to AI on the company’s own data. We also connected ChatGPT to the Snowflake data so teams could start asking questions right away, though we recommend keeping AI inside the platform, where the governance already lives
• A KPI analysis in a bowler structure, tying metrics to operational targets and showing which ones the data could support

• About $34K a year in labor handed back. A weekly date-change review that consumed 16 person-hours of meetings and spreadsheet work was automated, freeing roughly 830 hours a year.
• A board report leadership could trust. The dashboard was reconciled against the manual spreadsheet before anyone was asked to rely on it, so the number of record no longer lived in one person’s head.
• AI-ready data. The company’s data now sits in one governed platform with a shared data model, so any AI it adopts starts from consistent definitions.
• A sequenced path instead of a completeness project. The KPI analysis showed six general-ledger objects, about 13% of the remaining ingestion, would answer 27 of the 50 metrics the board reviews monthly.