Kill Your Data Silos, Not Your Data Team:
Lessons from the Frontlines of AI Product Development
Special thanks to John Farrall’s excellent Alternative Data Weekly, which surfaced many of the ideas explored in this post. If you’re not already reading it, you should be.
There’s a brewing tension in the world of data and AI-between scale and specificity, automation and understanding, centralization and embedded intelligence. This week, we explored four insightful articles that, while seemingly disparate, converge around a powerful message: the way we think about data infrastructure, AI agents, and organizational design needs to shift if we want to move from hype to impact.
Here’s what we took away—and what it means for how Storm King Analytics is evolving our approach to building smarter, leaner decision systems.
1. You Don’t Need a Data Strategy (At Least Not the Way You Think)
Based on Your AI Product Doesn’t Need a Data Strategy by Jaser Bratzadeh
In what feels like a counterintuitive take, Bratzadeh challenges the long-held assumption that every AI initiative must begin with a robust, organization-wide data strategy. Instead, he argues, successful AI products are often born out of a deep understanding of specific workflows and pain points, not from a generalized effort to “get our data in order.”
This is not an argument against data quality or governance. Rather, it’s a reframing: treat data strategy as a means, not an end. Start with the product, then reverse engineer the data needs. Build the scaffolding around the use case, not the other way around.
For teams like ours, working with military and development clients, this aligns with what we’ve seen in practice: tools that meet users where they are, messy spreadsheets and all, can outperform the cleanest data warehouse if they deliver timely and relevant insights.
2. Kill Your Central Data Team? Only If You Replace It With Embedded Intelligence
Inspired by Kill Your Data Team by Sven Balnojan
Balnojan delivers a sharp critique of centralized data teams, suggesting they often become bottlenecks, detached from the operational problems product teams are trying to solve. His proposed alternative? Embed data capabilities directly into cross-functional product teams.
This echoes a growing trend we’re seeing with clients: data fluency is no longer a specialist skill. Whether it’s a product manager, a training NCO, or a policy analyst, the ability to ask the right questions of data and act on the answers is now a core competency.
But here’s where it gets tricky: the real unlock comes not just from embedding analysts, but from empowering translators-people who can speak both “data” and “domain.” These are the bridge-builders who turn theoretical models into practical decisions and who can reframe business goals in a way that engineers and modelers understand.
At Storm King, we’ve come to see these translator roles as some of the most valuable players in the room. They help avoid both overengineering and underutilization. They understand that sometimes the right answer isn’t “more AI”-it’s a better question, better framed.
3. MCP: A Simpler Way to Talk About a More Complex World
From MCP for Dummies by Aetheron Lab
The concept of Model Context Protocol (MCP) may sound like science fiction, but it provides a framework for addressing real-world problems that involve distributed decisions, imperfect information, and complex interdependencies.
Think: disaster response, military operations, multi-actor economic development. In these settings, no single agent has the full picture, but coordination is still essential. MCP creates a structure in which AI agents, human users, and evolving context interact continuously.
For us, MCP provides a new lens through which to frame the decision-support tools we’re building. It shifts the question from “What’s the best answer?” to “What’s the best coordinated set of actions across agents, given what each knows?”
As our platform Vumbua evolves, we’re exploring MCP principles to guide everything from how alerts are prioritized to how foresight models are developed in different operational contexts.
4. AI Agents Tried to Run a Fake Company—It Didn't Go Well
Based on Professors Staffed a Fake Company Entirely With AI Agents by Joe Wilkins
In an experiment at Carnegie Mellon University, researchers created a fake software startup, TheAgentCompany, and staffed it entirely with AI agents from OpenAI, Google, Anthropic, Meta, and Amazon. Each bot was assigned a role, from software engineer to HR manager, and tasked with day-to-day operations like writing performance reviews and navigating file systems.
The results? Disastrous. The best-performing agent completed just 24% of its tasks; others fared far worse. Agents hallucinated coworkers, renamed users to bypass chat restrictions, and made decisions devoid of common sense.
We’re not building self-replicating AI (yet), but we are testing modular agent stacks that specialize in ingestion, enrichment, and scenario modeling. Each “agent” has a bounded task and plugs into the larger system like a relay in a complex circuit.
This makes our architecture more explainable, more flexible, and more aligned with the way humans actually make decisions in messy, real-world settings.
Unifying Thread: From Data-Centric to Decision-Centric
If we had to summarize the common thread across these four perspectives, it’s this: we’re witnessing a shift from data-centricity to decision-centricity. The questions are no longer “Do we have all the data?” or “Is our lakehouse optimized?” but:
Can we answer meaningful questions faster?
Can we coordinate intelligently across silos?
Can we embed decision support where it matters most?
And crucially: Do we have people who can translate between abstraction and action?
At Storm King Analytics, we’re betting on tools, models, and people that respect the messiness of human systems, elevate domain context over abstraction, and aim not to automate decisions, but to augment decision makers.
That’s the future we’re building toward.
Liked this post? Subscribe to Alternative Data Weekly by John Farrall for more curated insight at the intersection of analytics, markets, and emerging tech.
Want to collaborate with us? Drop us a line. Whether you're wrestling with disjointed spreadsheets, designing agent workflows, or trying to scale decision intelligence across your org, we’d love to help you discover what’s possible.
Thanks for the shout out!