What BI means, revisited
I defined business intelligence in 2021, revisited it in 2024, and I’m revisiting it again. What changed, what didn’t, and which of my own definitions aged worst.
Rewritten from a post on my previous blog, first published in March 2021.
I wrote a definition of business intelligence in 2021. I rewrote it in 2024. Reading both back now, the interesting part isn’t how the field changed. It’s which of the two posts aged worse, and it wasn’t the one I expected.
2021
Here’s roughly what I said then.
Business intelligence is a collection of elements that turn data into useful information quickly enough to act on, so decisions rest on facts rather than instinct. There is no standard BI solution. What you build depends on business requirements, data volume, user count, reporting needs, latency tolerance, and what you can support.
Then I laid out the layers.
Sources. Audit everything you have before choosing how to integrate it. Types, volumes, duplicates, nulls, inconsistencies. What you find determines the tooling, not the other way around.
ETL. The pump at the middle of the system. The design questions are how many sources, how much data, how often, what transformation, and what your retention policy says.
Warehouse. Where operational data meets. Either data marts first and then a central repository, or everything into the warehouse and functional models built off it. Either way, design it for reporting, not for the shape your sources happen to be in. Sources can be normalized; the warehouse should flatten for performance and simplicity.
Star schema over snowflake. Fewer joins, easier to work with, at the cost of some redundancy. Snowflake normalizes and saves storage and makes every query a join exercise. I recommended star and I’d still recommend star.
Data models. Built on top of the warehouse, carrying KPIs, calculated measures, and pre-aggregation. Tabular for simplicity, multidimensional for the advanced cases.
Reports and analysis. Paginated reports for the pixel-perfect printable case. Self-service for users building their own off a prepared dataset. Interactive analysis for slicing. Dashboards for live KPIs with drill-down. Data mining where you have enough history to predict.
AI and ML got one short paragraph at the end. Models invoked during transformation to serve a business need. That was the whole treatment.
2024
Three years later I wrote it again, and the tone had changed completely. Cloud ate everything. ELT overtook ETL. dbt, Fivetran, and Airbyte replaced hand-built pipelines. Snowflake, Databricks, BigQuery, Redshift. Lakehouses as the hybrid. Reverse ETL pushing data back into Salesforce. Generative AI writing SQL, described as a junior developer who doesn’t complain. Governance as the party police. Data catalogs as Google Maps for your data.
I closed with a line about BI being messy but now with more AI confetti.
It’s a funnier post. It’s also the one that aged worse, and it took me a while to work out why.
The 2021 post was about structure. What the layers are, what decisions each one forces, what tradeoff you’re accepting when you pick star over snowflake. Almost none of that expired. The 2024 post was mostly an inventory of what was popular that year, and inventories rot. Half the tool names have shifted position, the AI paragraph reads as breathless, and the actual argument, if I’m honest, was “here is a list of things that exist now.”
Worth remembering the next time something feels urgent enough to write about. The structural post survives the hype cycle. The hype post does not.
2026
So what is BI now?
The 2021 definition still works. Turning data into useful information quickly enough to act on, so decisions rest on facts rather than instinct. Nothing about that sentence needed updating in five years, which is either a sign it was a good definition or a sign it was vague enough to survive anything. I’ll let you decide.
What changed is the question underneath it.
For twenty years the BI problem was framed as access. People couldn’t get to the data, so we built dashboards. They couldn’t get what they needed from the dashboards, so we built self-service. Self-service mostly didn’t work, and the standard explanation was that the interface was still too hard, so every few years someone shipped an easier interface.
Then the interface problem actually got solved. Anyone can ask a question in plain language now and get a well-formed answer back. And the thing that became obvious, at least to me, is that the interface was the visible problem but not the load-bearing one. Trust was underneath it the whole time, and a friction-free interface just got people to the trust problem faster.
A dashboard nobody uses is often not a usability failure. It’s someone deciding, with good reason, that they don’t know where the number came from or whether it’s current. The interface was a proxy for that, and removing it just exposed the real problem faster.
Which is why I’d now put governance at the front of the definition rather than the back. In 2021 I gave AI one paragraph at the end. In 2024 I gave it jokes. The version I’d write today is shorter and less fun:
The binding constraint on BI is trust: whether someone will act on a number without checking it first. That isn’t the whole field. Analysis, discovery, the actual finding of something worth knowing, none of that goes away. But it’s the constraint that decides whether any of the rest gets used, and everything else, the pipelines, the warehouse, the models, the dashboards, the agent, is machinery downstream of it.
The reason this matters more now than it did is a small thing about how people read. A dashboard carries its own credibility signals. You can see the filter, the date range, the last refresh, the shape of the chart against what you expected. Users have spent twenty years learning to squint at those.
An answer in a sentence carries none of that. It arrives finished. Whatever went into it, correct definitions or a join nobody sanity-checked, the output reads the same, and the squinting has nowhere to land.
So the order of work inverted. The unglamorous end, governed data with real business context attached, stopped being the compliance chapter and became the prerequisite. And the piece most data teams still don’t have: evaluations on what the system says, reviewed and fed back, so the loop tightens instead of drifting. Without that you have a demo. With it you have something people will act on without double-checking, for the uses that have earned it, which was the goal in 2021 and is still the goal now.
The definition didn’t change. The bottleneck moved.
If I write this again in 2029, my guess is the tools in the 2024 post will look quaint, the structure in the 2021 post will still mostly hold, and I’ll have been wrong about something specific in this section. That’s roughly the pattern so far.