Analytics that builds the platform: say goodbye to Ticketing 2.0 with agents in the git repo

The problem: the false promise of perfect data and the Ticketing 2.0 trap.

With the rise of conversational analytics and modern semantic layers, a received idea is taking hold: the data team would only have to provide perfectly modelled data, and the business would only have to consume it.

In practice this is a bureaucratised illusion:
- eighty percent of exploratory value hides in raw data that has not been modelled yet;
- confining business teams to a frozen catalogue reinvents the historical bottleneck, Ticketing 2.0, with Jira tickets to add the smallest column to the semantic layer;
- to work around that heaviness during deep dives, teams create shadow data through CSV exports and isolated transformations, killing knowledge capitalisation.

The solution: shifting analysis left into the monorepo.

The arrival of engineering AI agents such as Claude Code in VSCode collapses the technical barrier of git and dbt for profiles with a technical veneer. Rather than isolating engineering and analysis, the modern 2026 architecture rests on merging those two worlds: bringing analysts directly into the platform's dbt repository.

What the audience will learn. In this talk based on concrete field reports, from organisations managing more than one hundred and fifty analysts to recent deployments such as Optic 2000, we cover:
1. The Ticketing 2.0 anti-pattern: why a watertight separation between curators and analysts destroys serendipity and analytical momentum.
2. The deep dive and pipe workflow: how an analyst assisted by an AI agent can navigate raw sources, surface new dimensions and update the underlying dbt models during the exploration itself, so that the analysis builds the platform.
3. Operational practice: building specialised AI skills to generate auditable analyses in Markdown, update the data portal documentation and permanently eliminate phantom logic.