Many SaaS company boards are currently panicking because their multi-million dollar generative AI pilot hasn't changed the business. In my experience, the real wins in enterprise technology are almost always unglamorous.

A recent analysis by Andreessen Horowitz looked at where actual budgets are going rather than subjective surveys. The data showed that production-grade adoption is concentrated in highly structured, verifiable areas. Software engineering leads the pack, followed by customer service and database search.

This is what they call a quiet accumulation of edge. Instead of overnight revolutions replacing entire departments, we are seeing a slow, steady reduction of operational friction in specific workflows. The companies winning this transition are not the ones chasing flashy keynotes. They are the ones focusing on their data foundations and system architecture.

We saw this pattern clearly when we modernized the global production infrastructure at Unity. We had a large legacy footprint running workloads at scale, and the industry pressure was to slap a conversational assistant on top of the existing systems. *Side note: we did that too, at scale, but that is a story about the unit economics of an Enterprise AI platform for another day.

Instead, we focused on rebuilding the substrate itself to be AI-ready. This meant cleaning up databases and simplifying access. We also had to restructure how data moved. We built self-healing systems that could spot infrastructure problems and fix them before they caused downtime. It allowed us to scale our operational capacity without expanding our team size or buying expensive new tools. We measured success by the significant reduction in incident response times and system outages (shout-out to the brilliant Unity Platform and PRE teams)!

If you want to move past the pilot phase, you have to stop treating AI as an application and start treating it as infrastructure. This means shifting your focus from the user interface to the data pipeline. You can test your own AI pilots with a simple question: if you remove the shiny front-end interface (whether it's a chatbot, a dashboard, or a portal) does it still provide value to the team? If the value lies solely in how the information is presented rather than what the system does with it, you've built a visualization tool, not an AI engine.

When assessing ROI, stop looking at superficial metrics like user engagement or number of prompts run (or worse, number of tokens consumed as a sign of "adoption"!). Instead, look at velocity and capability. Are your teams shipping code faster? Is your customer support resolving complex issues without escalation? Most importantly, look at the ratio of operational overhead to transaction volume. True AI ROI shows up as an ability to scale your operations without a corresponding increase in headcount or tooling spend.

Meaningful productivity gains do not come from summarizing technical logs or searching internal documentation. The real impact is when you authorize systems to take action. At Unity, we used AI to detect anomalies, isolate failing nodes, and redirect traffic automatically. It was smarter, needed, more precise, and accomplished without expensive retooling. 

If you are an executive sitting in the weekly Friday demo, ask yourself a simple question: are you applauding a neat interface that looks cool on a screen, or are you looking at a system that actually cleans up the underlying database running your business?