Balin Michael

I've been thinking about a significant question lately: Why do so many enterprise AI demos look impressive, yet struggle in real business scenarios? After months of building legal AI at Snowflake, I've come to believe the answer comes down to context. For instance, we developed a tool to assist attorneys in reviewing contract redlines. The initial version functioned like many AI tools, analyzing customer changes against our playbook and suggesting responses. Technically, it was effective. However, attorneys lacked trust in it, and rightfully so. An experienced attorney doesn't evaluate a clause in isolation. Before addressing a liability cap, they consider various factors: - How substantial is this deal? - Is this a strategic customer? - What have we agreed to in previous negotiations? - What's standard for this industry? - Where are we in the deal cycle? This context is instinctive for a human reviewer, but the model lacks it. Although the model can reason effectively, it doesn't understand your customers, history, or business nuances. Present it with the same clause twice, and it will likely produce the same response, regardless of whether it's a $50K renewal or a $5M strategic opportunity. What ultimately made the difference wasn't changing the model. It was providing the model access to critical business information: deal size, industry, prior agreements, historical concessions, and other signals that attorneys naturally consider when making decisions. Once AI could access more of what the attorney perceived, the recommendations became significantly more valuable. I believe the primary challenge in enterprise AI today isn't the intelligence of the model. Many organizations already have access to powerful models. The real challenge is ensuring the right business context is available at the right moment, while maintaining governance, permissions, and data boundaries. Much of this context already exists but is often scattered across systems, hidden behind access controls, and difficult for AI systems to access and use effectively. Organizations that successfully address this will gain a substantial advantage—not because they possess a superior model, but because they have a more effective method of linking institutional knowledge to the model. This is more of an architecture problem than a model problem. I've been documenting the architecture we implemented to tackle this issue and will share more insights in a couple of weeks. I'd love to hear how others are approaching the gap between AI demos and AI in production. #EnterpriseAI #AIGovernance #LegalTech #AppliedAI #ContextEngineering

18

1 Comment

Data modeling is pretty much at the heart of analytics engineering. I recently came across Joe Reis’s Practical Data Modeling substack (and ended up subscribing), and it’s quickly become one of my go-to reads. What I really like is how Joe takes topics that feel “settled” and gives them a fresh spin, or clears up confusion that still floats around. His recent post on Medallion Architecture is a great example. It’s free and well worth a read. “The Medallion architecture is no more a data model than a parking lot is a type of car.” 😄 I’ve seen the same mix-up in practice. Medallion (Bronze → Silver → Gold) is really about lifecycle stages for refining data. Data modeling, on the other hand, is about defining entities, relationships, and grains across conceptual, logical, and physical layers. They work together, but they’re not the same thing. Too often, teams use “Gold” as shorthand for a certain model, but as Joe points out, Gold doesn’t prescribe any particular approach. It’s simply about making data consumable. That distinction matters, because when teams conflate Medallion with modeling, they can lose sight of real challenges, like aligning on consistent definitions, grains, and structures so the business can trust the data and act on it with confidence. I’ve dropped the link to Joe’s post in the comments. Highly recommended. #DataModeling #AnalyticsEngineering #MedallionArchitecture

45

3 Comments

I’m still thinking about something I heard earlier this week at Snowflake’s Data for Breakfast in NYC. Ian McComber from Ramp mentioned that, at least for now, their willingness to spend on AI tokens so their teams can experiment is “basically infinite.” That idea stuck with me. The only real way to develop intuition about AI is by using it constantly. Trying tools, building things, breaking things, and seeing what actually works. It makes you wonder: How many organizations are slowing themselves down by trying to optimize costs before they optimize learning? Right now, the scarcest resource might not be tokens. It might be experimentation.

10

1 Comment

Last week, I wrote a post poking fun at people using AI as their analysts. As always, a few folks in the comments told me I just didn’t know how to prompt properly. ...And, they're right. They (intentionally) weren’t great prompts. But let me ask you this: How often does a stakeholder question come with all the detail, background, and nuance you need to get the answer right the first time? If we struggle to get that context, how many stakeholders do you think will provide it to their AI analyst?

225

67 Comments