Self-serve failed because
you handed over tables
Nearly every organisation has tried self-serve reporting and concluded that people do not want it. What they actually declined was a database with a chart builder on top.
What the first attempt looked like
Access was granted, a tool was rolled out, training was run, and a handful of enthusiasts built things. Everyone else opened it once, saw a list of tables with names like fact_order_line_v2, and went back to asking the analyst.
The conclusion drawn was that the business is not data literate. The more accurate conclusion is that the interface required the user to already know the data model, which is a specialist skill nobody had agreed to acquire.
The gap between a table and a question
A person wanting to know whether returns are up in the north does not want a table. They want returns, and a region, and a comparison to last month. The work between those two things is joins, filters, grain, and knowing which of three date columns is the one that matters.
Handing that work to the person asking the question is the failure. It is not made easier by a nicer chart picker.
What changes with governed metrics
When returns is a defined metric, and region and month are defined dimensions, the interface can offer those directly. The user picks things they recognise from their own job, and the platform resolves the joins and the grain underneath.
The specialist is still required, but once per metric rather than once per question. That is the whole difference, and it is the difference between a queue and a self-serve system.
Where the specialist still belongs
New metrics, genuinely novel analysis, anything that requires deciding what a number should mean. That work is real and does not go away.
What goes away is the specialist rebuilding a variation of the same regional breakdown for the fourth time this quarter.
Common questions
Do people actually use this, or does it become another unused rollout? Adoption tracks whether the first thing someone tries returns a sensible answer. If it does, they come back. Most failed rollouts lost their users in the first five minutes, on the table list.
What about people who want a real query language? They should have one. Analysts working directly against the model is a legitimate and important path. Self-serve is for everyone who was never going to write SQL and should not have to.