Case studies and essays
Platform and team decisions, written up for anyone paying for a data platform or building a career on one. Case studies take one decision each, and the essays carry the patterns further. Start with the subject you are deciding about.
What actually happens when you put agents and LLMs to work inside a real data organization, and what the demos leave out.
1 case study · 5 essaysBrowse
Hiring, structure, capacity and the stakeholder work nobody puts on a roadmap. Building a function, not just a backlog.
2 case studies · 7 essaysBrowse
Executive partnership, strategy, and the difference between reporting on decisions and being present while they are made.
1 case study · 5 essaysBrowse
Quality, definitions, lineage and incidents. How to make reliability visible and rebuild confidence when something goes wrong.
2 case studies · 3 essaysBrowse
Platform decisions with the reasoning shown, including what they cost to run and the tradeoffs I accepted.
4 case studies · 2 essaysBrowse
One decision each, in the same six sections: situation, constraint, decision, the tradeoff I accepted, outcome, and what I would do differently. Observed changes are separated from the criteria I recommend, and published industry examples are credited to their sources.
Greenfield data function, analysts embedded across departments
Building a central data function from zero: org design, hiring, and centralizing scattered analysts without becoming the bottleneck.
Executive partnership across departments and at board level
What changes when data reaches the executive table: leading with the decision, showing uncertainty, and keeping the seat.
An AI assistant for self-service, on a governed data layer
Launching an AI analyst against a real warehouse: metrics as tools, a knowledge base of business context, and a feedback loop that became the eval set.
Data quality and reliability as an operating discipline
How trust in a data platform is lost, slowly or all at once, the four failure modes that break it, and how to run an incident well.
Marketing and payment attribution across teams that each kept a version
Attribution as a definition problem: four teams, four acquisition stories, and why named measures with owners beat a single blended number.
Modernizing a warehouse that had fallen behind, while reporting continued
Modernizing a warehouse that had fallen behind: migrate by the decisions reports serve, run two platforms deliberately, and retire what nobody would miss.
Warehouse consolidation across mergers, silos, and migrations
Consolidating warehouses after mergers, silos and migrations. Why definition reconciliation is the project and pipelines are support.
Ingestion rebuilds across heterogeneous sources and warehouses
Rebuilding data ingestion: choosing the integration approach, reconciling sources that disagree, and proving it by parallel run.
Platform cost governance and capacity decisions
What a data platform costs and who decides: cost attribution before optimization, and which inefficiencies to leave alone.
Central platform team federating ownership to domain analysts
Federating model ownership to analysts: what a central data team gives up, what it gets back, and the guardrails that make it safe.
October 2026
Moving a data team out of IT helps only when it changes what the team is measured on, who may decline a request and who pays for the engineering.
How to decide whether an AI data assistant should be widened, narrowed or retired: four numbers against the queue it replaced, and what retirement keeps.
When AI drafts the pipelines and models, review becomes the constraint: what changes in the queue, what a reviewer checks, and what it costs a data team.
Hiring for judgment in data: an interview built around one bounded decision, what to listen for, how to score it, and what the format costs.
The evidence of judgment in a data career is a record of decisions, not deliverables: four columns to keep, how to read them, and what the log costs.
Changing a shared metric: agree on its purpose and effective date, preserve comparable history, and keep earlier decisions explainable.
After a data error reaches a decision, fixing the number leaves work unfinished. Trace its use, correct the record, and return the choice to its owner.
When one senior analyst reviews everything: assign review capacity, develop another reviewer, and recognize the work that keeps decisions sound.
September 2026
When fresher data changes a decision: price the full operating promise, account for late records, and choose which workloads need the speed.
A data team whose roadmap is other people’s requests needs a remit, not a better scoring model: one sentence, used to decline one thing.
Dashboards show the number and stop. AI briefings add the reason and a next step. How to keep the numbers governed and the reasons checkable.
Developing judgment in a data team: delegate a bounded decision, review the reasoning, and help someone lead without quietly taking the work back.
A testable definition of data-driven, five rungs named for what happens when a number and an opinion disagree, and how decision habits connect to tooling.
A current data toolkit paired with the 2025 list so the drift shows. The core barely moved, and part of it is no longer data tooling at all.
Why sound analysis produces recommendations nobody acts on, and the rewrite that fixes it: name the decision, offer an alternative, state a condition.
Rolling out an AI data assistant in the team chat tool: why to start narrow, recruit people to break it, and what it exposes about the questions.
Self-service analytics fails on habits, not access. Four questions that separate answers from charts, and a worked example of a conversion drop.
Three passes at defining business intelligence, over five years. The definition held. The bottleneck moved from access to trust.
Handing dbt model ownership to analysts is a governance change, not a tooling one. Why the question is now authorship, and what the move cost.
Catalog, discovery, observability and governance are four different things. What changed when the thing querying the warehouse stopped being human.
Most dashboards end up ignored. Why that is the default outcome, what moves one from decoration to a decision, and what the fix costs.
Evaluating AI analysts against mature warehouses: ambiguous metrics, reviewed answers, access boundaries, and the evidence needed to expand a pilot.