The AI saddle point

We’ve all felt it working alongside our favorite coding agent, AI business analyst, or copilot; your talent and brainpower amplified to move at the speed of imagination. Want a new plot showcasing win rate in North America? Sure thing. A new react page outlining a competitive analysis coupled with internal metrics? On the double. A scientific research paper distilled to the level of a five year old and in the voice of a pirate? Y’arr me little matey!

In these moments, you feel as though the only thing limiting you is your own creativity and the current load balancing of your favorite model. New doors have been opened, software can be created, and the breadth of the internet is at your fingertips. Personally, I have delved into data analysis for my team and my organization. I want to understand as much of my company's data as I can get my hands on.

However, this power comes with risk. It’s never been easier to spin up your own database and make a prediction, create a comprehensive data analysis, or trust a bunch of weights and biases with your API credentials. However, without expertise in these areas, you’re only as good as your next token.

In mathematics, a saddle point is a point of inflection. In one direction, an area of stability. In another, a drop off bounding into the unknown. I believe this is where we stand with AI. AI makes you 10x better in the things you know, and 10x more dangerous in the things you don’t. Models are incentivized to get you to an answer, leveraging any available data sources, tooling, or code language to do so. You are still the one guiding them towards the right answer.

Let’s step back to consider the evolution of models. In the early chatGPT days, models were machines which took the internet data they were trained on to provide next-token prediction to your heart’s content. They were even factually correct sometimes! Since that initial release, a number of improvements have transformed these models to leverage tools to serve as effective logical problem-solvers. The time-horizon for independent agent activities is increasing, and this same horizon extending into days was named the big capability increase for models in the next 12 months by Jeff Dean, Chief Scientist at Google Deepmind. It’s what has spurred the rapid adoption of Claude Code and other AI-assisted software development tools.

Now, back to your personal AI companion. Let’s imagine that same value at exponential scale and without your intervention. Each employee has double digit agents working 24 hours a day, based on triggers, timers, or anything in between. My 10x saddle point becomes 100x or even 1,000x, as agents continue to increase automated decisions and drive nontrivial business impact.

This is where the foundation matters. As I think through my own use case and any other data-driven initiative, the quality of the underlying data becomes paramount. You need to be sure that the data supporting all of these decisions is high quality and well maintained. You need confidence in data lineage, well-defined data products, and clear connections between physical and logical concepts. Inconsistent or low-quality data creates brittleness, a foundation that will crack under the load of an agentic world, making your agents confidently dangerous.

In conversations with Collibra customers, I’ve found that the ones who invest in a strong data governance foundation are already seeing dividends in the value of their AI initiatives. Trust in the truth of data spurs value streams that didn’t exist previously. I’ll even be fortunate enough to speak with Ram Ramanathan, Senior Director of Data Governance at Macy’s on July 9th to learn about their journey (feel free to register here). Data governance has become an accelerator of AI driven value, and the acceleration will only get higher as agents become more capable.

The good news is that it is not too late to invest in your data foundation. Collibra accelerates this journey with a robust and performant data context layer, and employs fortunate people like me who get to be along for the ride alongside you. Want to see how Collibra can accelerate your AI journey? Request a demo here.

Note: While I love and appreciate AI, it was not used to write this blog, only edit my typos and grammar mistakes.

• 

  Eric Warner

  Eric Warner

  Director, AI Engineering

  Collibra