Sýn

Sýn, Iceland's telco and media group, had accumulated three overlapping warehouses, roughly 1,200 stored procedures holding business logic, half a dozen ETL tools and three BI platforms. Each was justified when it arrived; none was ever retired.

It produced the reports it had to. It also produced 188 incidents a year, and conflicting numbers depending on which warehouse you asked. The data specialists spent their time on reconciliation, fixes and incident response, while the analysis the business actually wanted kept being deferred.

The programme turned on a single reversal: free the people from fixing data, and let them work with it. A Databricks lakehouse gave the foundation. What makes it different is what went in from the first day.

Solid foundations come before any AI. Every business source — advertising sales, billing and usage, CRM, the national registry, ERP and planning — now lands in one governed lakehouse with full lineage in Unity Catalog.

Nine business areas keep curated gold workspaces, modelled together with the business owners, alongside certified semantic models where each KPI is defined once — conformed dimensions, row-level security and Git versioning.

That last part carries more weight than it looks. Everything that consumes the platform — a dashboard, an analyst or an AI system — reads the same certified figures, and sees only what it is allowed to see.

Master data — product catalogues, reference dimensions, the values everything else joins against — sits in an awkward middle ground: too governed to edit directly, too operational for a ticket queue. At Sýn, updating a dimension meant spreadsheets, emails and engineering time.

So we built a governed write-back loop, driven from chat. Business users read and change master data conversationally in Claude Desktop, without raw lakehouse write access and without stepping around governance. Identity is checked at the door, privilege is minimal throughout, and nothing reaches the lakehouse on a model's say-so — only on a human's confirmation.

A custom skill parses uploads — xlsx, csv, pdf — maps values to the dimension schema and shows a confirmation table. An MCP server on Azure Container Apps exposes read and write over a stateless HTTP endpoint, with Entra ID validating every token through OAuth 2.1, Conditional Access and MFA, and Managed Identity connecting to storage and Databricks without stored secrets. Confirmed changes land as parquet in a dedicated master-data store — an immutable record — and Delta Live Tables validate, enrich and promote them through bronze, silver and gold to the golden record.

Because Databricks, Entra and the networking were already Sýn's, the net-new cost is a container app and a storage account: about €40 a month. The first dimension, the product catalogue, went from kick-off to production in six to eight weeks, and the pattern extends to any other dimension.

Consolidating the tools into one governed platform ends the arguments about whose number is right. The people who used to keep pipelines alive now sit with the business teams, contributing to the platform directly rather than adding to a backlog.

Folding legacy licences, parallel pilots and ageing infrastructure onto the lakehouse roughly halves the platform's run-rate — while adding the scalability, governance and AI capability that were not available before. Support through the Databricks partnership part-funded the programme, which lowered the up-front investment before the savings began.