AI's ceiling often gets described through a few extra benchmark points. A slightly better score on MMLU or GSM8K. A cleverer chain-of-thought or the latest MCP. A shiny new reasoning mode that feels almost human.
BI's ceiling is "talk to your data", "agentic analytics", "insights that write themselves", "your dashboard replaced by a chat window".
But the floor is where most of us actually live.
AI's floor is the stubborn hallucination rate still holding back production use. The opaque model updates that break your workflow overnight. The memory limits that forget context halfway through a task. The cost per token that quietly adds up before value is proven.
BI's floor is old, simple, but very real:
- System stability
- Support quality
- Write-back or override
- Cell-level calculation and commentary
- Pixel-perfect paginated reports
- Version control
And I mean features backed by testing, clear ownership, support and operational controls. A useful prototype is a starting point; running an enterprise service is a different commitment. Some of these needs were addressed by earlier products, and I haven't found one product that covers every base.
The ceiling attracts the headlines and Wall Street's money. It probably will change how millions of self-employed people and small businesses operate.
But in enterprise land, I think the floor deserves at least as much attention as the ceiling. Reliable everyday use is where I'd start.
In my own conversations, interest in new AI capabilities has not always translated into sustained use. Some features are still hard to rely on in production. I have more than once tried to present some AI-generated "wisdom" to real people with real wisdom, only to find it was something they already knew, or not that helpful because the answer missed what actually happens in their work.
Meanwhile, people are trying to get simple, reliable things done. Excel spreadsheets remain part of that work.
AI-assisted development gives us another way to build stronger floors. I still see teams spending too many hours getting a report ready for a decision. A more capable model can help, but it does not remove the need for reliable data, clear ownership and a process people can use.
If we can help more people use AI and BI reliably through sturdy, everyday "floor" features, we have a better chance of turning technical progress into useful business results. That's a bet I'd put a few bottles of good wine on.
Am I asking too many questions? Or asking too much?
First posted on LinkedIn. Revised for this site in November 2026.