by Ward Rushton
AI has become a constant presence in executive conversations. Every week brings a new headline, a new vendor pitch, or a new internal proposal promising to “unlock value” with generative models. Yet for many leadership teams, enthusiasm quickly gives way to uncertainty.
Everyone agrees AI matters. Far fewer people can explain what a reliable, ROI-positive AI step actually looks like inside their own organization.
The gap usually isn’t ambition or data. It’s execution. More specifically, it’s the difference between experimenting with AI and productionalizing it in a way the business can trust.
The architecture patterns described below come from conversations with Geoff Clark, an expert software architect who has spent years designing and scaling Snowflake-based data and AI systems, and who has seen firsthand what makes promising AI initiatives succeed or stall.
To make that difference concrete, let’s follow a realistic use case from its genesis through a proof of concept, the growing pains that inevitably follow, and finally into a production-ready architecture.
The Spark
Most organizations don’t lack customer feedback. They drown in it.
Complaint emails land in shared inboxes. Call center conversations are recorded and transcribed. Frontline service teams jot down bullet-point notes after in-person interactions. Individually, each data source tells part of a story. Collectively, they represent a rich but fragmented view of customer sentiment that is difficult to analyze at scale.
At some point, someone asks the obvious question: Why aren’t we using AI for this?
In our conversation, Geoff pointed out when discussing early AI initiatives, this moment is usually driven by intuition rather than architecture. Leaders see value locked in unstructured data and assume generative AI can unlock it. They are right, but only partially. The challenge is not generating output, it is building a system that can support real decisions.
A Fast and Impressive Proof of Concept
In the early phase, speed matters more than perfection. The goal is to make the idea tangible and prove that real business value exists.
Geoff typically advises teams to start by keeping the proof of concept entirely within the Snowflake ecosystem. Data already lives there (or can be accessed through Snowflake stages that reference external storage such as S3 without unnecessary duplication). By staying close to the data, teams avoid an entire class of security, governance, and data movement issues that often slow early AI initiatives.
For the interface, a Streamlit app in Snowflake provides a lightweight way to stand up a working application quickly. With a small amount of Python and some boilerplate templates, teams can build a simple UI that allows users to select time ranges, filter by customer segment, or ask high-level questions. Because the application runs inside Snowflake and authenticates through Snowflake, it feels like a normal web app to the user, while remaining governed behind the scenes.
Behind that interface, Snowflake Cortex functions can be used to generate sentiment scores and thematic tags from emails, transcripts, and notes. The outputs are immediately useful. Leaders can see sentiment distributions, emerging complaint themes, and representative examples drawn directly from their own data.
At this stage, the enthusiasm is justified! The proof of concept demonstrates that generative AI can turn unstructured customer feedback into actionable signals. It also does so without introducing new infrastructure or brittle integrations, which Geoff emphasizes is critical when the goal is to learn quickly without creating future clean-up work.
Success Creates Pressure
The moment a proof of concept resonates with stakeholders, expectations begin to shift.
Suddenly, more people want access. Regional leaders ask for views tailored to their territories. Product teams want to correlate sentiment with releases. Executives start asking for regular summaries rather than one-off demos. What felt like a lightweight experiment begins to resemble something the business might depend on.
This is where cracks appear, not because the AI is flawed, but because the architecture was never meant to carry production weight. Geoff notes that this is a common inflection point. Manual steps that were acceptable in a demo become risks. Edge cases start to matter. Questions about access control, consistency, and trust move from background concerns to boardroom topics.
It is tempting to respond by tweaking prompts or swapping models. In practice, the underlying issue is architectural. A proof of concept proves possibility. A production system must prove reliability.
Mature teams pause here and reassess what they are actually trying to build.
What started as “analyze customer sentiment” evolves into something more operational. Leaders want to understand not just whether sentiment is changing, but why. They want confidence that trends are grounded in real customer interactions. They want to move quickly without sacrificing accuracy.
Geoff frames this shift as moving from an AI output to a business capability. Once that mental shift happens, the requirements become clearer. The system needs stronger controls. It needs predictable behavior. It needs to scale in both data volume and organizational reliance.
Growing into production
In Snowflake, the transition from experimentation to production does not require starting over. It does require changing how the application is deployed and governed.
Geoff frequently recommends moving from Streamlit-based prototypes to Snowpark Container Services when production requirements emerge. Snowpark Container Services allows teams to deploy containerized applications directly within Snowflake, providing fine-grained control over application logic, workflows, and behavior while keeping the system close to the data.
This approach solves several problems at once. It enables more robust handling of edge cases. It supports predictable scaling. It also allows organizations to continue leveraging Snowflake’s role-based access controls, security model, and governance features without rebuilding them elsewhere.
On the AI side, capabilities are composed deliberately. Cortex Search handles retrieval-augmented generation over unstructured customer communications, managing the complexity of chunking and retrieval behind the scenes. Cortex Analyst enables natural language interaction with structured sentiment tables by using a semantic layer that captures how the data fits together.
The result is a system that supports both exploration and explanation. Leaders can ask questions in plain language, understand trends, and inspect supporting examples without relying on bespoke analysis each time.
From insight to action
Once trust is established, expectations rise again, this time productively.
Executives no longer want isolated answers. They want outcomes. Regular summaries. Early warnings. Narratives that connect numbers to lived customer experience.
This is where Geoff points to agentic workflows as a natural next step. Agents act as an orchestration layer, breaking complex requests into smaller tasks, calling the appropriate underlying capabilities, and synthesizing a result. For customer sentiment, that might mean identifying meaningful shifts, pulling representative examples, and producing an executive summary ready for distribution.
At this point, the AI system is no longer an experiment. It is embedded in how the organization understands and responds to its customers.
Even with a Snowflake-native approach, moving from idea to production involves tradeoffs. Speed gives way to structure. Abstraction yields to control. Exploration evolves into standardization.
Geoff emphasizes that these tradeoffs are not failures. They are signals of maturity. When teams recognize them early, they can design systems that evolve naturally rather than break under pressure.
The takeaway
AI does not create value when it’s an impressive parlor trick. It creates value when people rely on it.
Organizations that succeed with generative AI are rarely the ones chasing the flashiest demos. They are the ones that understand the journey from idea to proof of concept to production, and who design their architecture with that journey in mind.
For customer sentiment analysis and similar use cases, Snowflake offers a path that supports that evolution. Start fast. Learn quickly. Then grow into something the business can depend on.
As architects like Geoff Clark have seen repeatedly, that is where AI stops being an experiment and starts becoming an advantage.
Need help bringing your AI ideas to life? Reach out to evolv.
Ward Rushton is a senior technology consultant at evolv, with over five years of experience supporting financial services and automotive organizations. He specializes in data-informed product strategy, modernizing decisioning platforms, and leading cross-functional teams to deliver measurable business outcomes. As a trusted client lead on complex engagements, Ward is known for translating technical complexity into clear, actionable direction. Based in Dallas, he brings a pragmatic, hands-on approach to enterprise technology work.
