AI Prototype to Production: What Makes Enterprise AI Apps Production-Ready

Watch the full session · 53 min

Vibe-coded prototypes are easy; production AI is not. Leaders from DataHub, Precisely, Snowflake and Fivetran explained what separates AI-enabled from enterprise-grade, where companies underinvest, how to govern agents and how much of a prototype survives the move to production.

Building an AI prototype has never been easier. Getting one into production at enterprise scale is still hard. That gap was the focus of "Prototype to Production: What it Takes to Build AI Enterprise Applications," a Data Science Connect expert panel webinar moderated by Brian Mink, the company's co-founder and president.

The panel brought together John Joyce, Co-Founder of DataHub; the VP of AI Labs at Precisely, who leads product innovation there; Arun Agarwal, Senior Product Marketing Manager, AI/ML at Snowflake; and Andrew Madson, Head of Developer Relations at Fivetran. The webinar was brought to you by Snowflake, SAP, Precisely, DataHub and Fivetran.

Why do AI prototypes stall before production?

Madson and Joyce both started with the same problem: no clear definition of success. Madson said teams should begin with a business objective, then work back to the AI capability and the data needed to meet it. Joyce said the problem definition also shapes the training and evaluation data, so a vague goal produces a data set that doesn't fit the use case.

Madson added that it's easy to "vibe code a prototype," but he recommended starting with a production-ready harness so teams don't have to rewrite everything to scale. Precisely's AI Labs leader said building the model is no longer the hard part. What breaks in production are cost at scale, latency, infrastructure limits, security and operational ownership, plus the gap between curated prototype data and messy production data. She cited a Precisely survey of 500 data, analytics and AI leaders, run with Drexel University's business school, which found that 70% say AI prototypes are aligned with the business but only 30% have measurable KPIs in place.

Agarwal said Snowflake customers are shifting from manual testing to automated validation, which he called "AI for AI": using LLMs to generate synthetic conversations, edge cases and failure scenarios, then scoring agents against them.

What makes an AI application enterprise-grade?

Joyce said an app that is merely AI-enabled "just means it runs on vibes": the output looks reasonable, so it ships. His first test for enterprise grade is whether anyone is clearly accountable for the feature and its quality.

Precisely's AI Labs leader described three principles: value, trust and control. Users have to see value every day before they adopt a tool. Trust comes from tracing inputs and outputs, explaining rationale and adding auditability. Control means model choice, guardrails, fallback paths and a human in the loop at key decision points.

Agarwal listed governed data, reliability, observability, cost control and a plan to scale from one user to thousands. "There's no AI strategy without a data strategy," he said. Madson said the model and the data must be treated as one system, each repeatable, observable and governed.

Where do enterprises underinvest when scaling AI?

Precisely's AI Labs leader pointed to the data foundation, including integration, quality, governance and runtime observability. Data will drift, she said, so teams should design for it from day one. Madson said decades of data anti-patterns, such as data lakes used as dumping grounds and fragmented silos, were tolerable when human analysts filled the gaps. Putting AI on top of that landscape, he said, means fixing ontologies, semantic layers and sources of truth.

Agarwal said the issue is often poor estimation rather than underinvestment. Teams overestimate conversation volume, turns and tokens, the project looks too expensive and gets shut down. "You don't optimize what you don't measure," he said, recommending tight usage tracking and scaling to clear milestones.

How should AI governance work in high-stakes environments?

"You can't govern what you can't see," Agarwal said. An agent can reach a correct-looking answer by a flawed path, picking the wrong tool or skipping a step. He described Snowflake's agent GPA framework, which checks an agent's goal, plan and actions.

Madson, who said he built AI systems at JPMorgan before agents arrived, said agent actions must be traceable and a human must remain responsible for answers. Too little data access invites hallucination, he said, while too much risks leakage. Joyce said "it's kind of like the wild west right now when it comes to MCP," with analysts and engineers connecting tools to internal systems without visibility. He expects central gateways to manage that access.

Precisely's AI Labs leader was blunt: "You need governance from day one. It cannot be an afterthought." Done early, she said, governance speeds teams up.

How much of a vibe-coded prototype survives?

At Precisely, she said, what survives is the model, the data and the evaluations. Agarwal said infrastructure, guardrails, error handling, data contracts and business logic tend to stay, while hardcoded flows change. Joyce said moving to production is usually an evolution rather than a rewrite, since "code is sort of becoming a commodity."

On balancing speed with stability, Madson said he has never seen a giant first project work, such as embedding all unstructured data for a general-purpose AI. Small projects with measurable ROI do. Agarwal said an internal Snowflake sales research agent took a couple of months to build and is now used daily by more than 6,000 people on its distribution team. DataHub moves features through alpha, private and public beta, then general availability.

For a first investment, Joyce recommended an evaluation data set, refreshed with anonymized real user conversations. Precisely's AI Labs leader advised finding what is unique to you. "Data is your moat," she said.

Keep up with enterprise AI in production

Data Science Connect has 24 free live webinars on its 2027 schedule across four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming sessions, or meet AI leaders in person at The AI Enterprise Conference on Tuesday, August 31, 2027, at Pier Sixty in New York City. Registration is coming soon.

Watch: Prototype to Production: What it Takes to Build AI Enterprise Applications on YouTube
Event: Webinars · More coverage: Webinars insights