AI leaders from Snowflake, Revefi and The Coca-Cola Company say most AI pilots stall for organizational reasons, not technical ones. They shared a three-part test for deciding what to scale, the infrastructure that holds up in production, and a 70-20-10 rule for where to spend effort.
Key takeaways
- Most pilots stall for systemic reasons: unclear goals, siloed teams and no business owner, more than weak models.
- Test every project on three things: business impact, operational feasibility and sustainability in production.
- Adoption depends on simplicity. The most scalable AI is invisible to the people using it.
- Plan for availability and cost. Uptime and consumption-based pricing matter as much as model quality.
- Follow the 70-20-10 rule: 70% of effort on people and process, 20% on technology and data, 10% on models.
By some estimates, 42% to 95% of AI pilots never make it past the pilot stage. Moderator Felipe Archila called it "pilot purgatory." In the Data Science Connect webinar "From AI to ROI: Turning Pilots Into Scalable Impact," three practitioners discussed how enterprises get out of it.
Archila, Director of Digital Workplace Analytics at The Coca-Cola Company, moderated the discussion with Girish Bhat, SVP at Revefi, and Santiago Giraldo, Senior Director of Product Marketing for AI at Snowflake. Panelists noted that their views were their own and not their employers'.
Why do AI pilots fail to reach production?
The causes are "not technology-centric," Bhat said. "It's more systemic failures." He described organizations running many disconnected pilots, chasing each new tool out of fear of missing out, without clear goals, roles or business owners.
Giraldo pointed to trust. Executives who aren't technical won't act on AI output unless they can see how a conclusion was reached and where its data came from. Transparency and traceability, he said, are what get them to use it.
How should enterprises decide which AI projects to scale?
Bhat uses a three-part test:
- Business impact: Does the project drive revenue, cut costs, reduce risk or raise productivity? Say which.
- Operational feasibility: Are the right data, tools and people in place? Most companies, he said, don't need to build their own LLMs.
- Sustainability: Will it work outside a lab built on curated data, and can it be monitored, governed and kept within budget? If the pilot's data was a one-off that can't be reproduced in production, he advised not scaling the project.
Giraldo advised treating AI as a portfolio: start small, expect some projects to fail, and scale what works. He added adoption as a fourth test. "The best AI is the AI that everybody can use," he said, "but people don't have to be technical to understand how it works."
He pointed to ride-share apps, which run dozens of models behind a single button, and to Raven, Snowflake's internal sales agent, which gives account executives verified answers about customer accounts on demand.
What infrastructure does scaled AI require?
Bhat started with availability. Model providers have outages, and production AI has to run reliably in real time. He also listed reliable data pipelines, data quality and observability, MLOps for deployment and drift, and budget discipline under consumption-based pricing. Without that discipline, he warned, "token maxing" can become a runaway cost.
On choosing tools in a fast-moving market, Giraldo recommended platforms that connect systems rather than lock teams in. That means support for many foundation models, open formats such as Apache Iceberg, and interoperability across vendors, while keeping security, governance and semantic context in one place.
How do you align teams around AI?
Giraldo said a shared mission matters more than any meeting cadence. Bhat recommended embedding AI architects and engineers directly in functional teams such as marketing and sales, and moving away from 12- to 18-month planning cycles and committee-driven decisions.
He also cited a 70-20-10 rule: 70% of AI scaling effort on people and processes, 20% on technology and data, and 10% on models. Many organizations, he said, have reversed it, putting about 90% of their focus on the LLM.
Archila compared the problem to the shift to remote work. Teams that simply moved in-person routines into digital tools got frustrated. Using AI to automate a bad process, he said, carries the same risk.
How are AI roles changing?
In the near term, Bhat said, roles are expanding. Data engineers are taking on AI work, chief data officers are adding AI to their remit, and product managers are adapting to how fast prototypes can now be built. Over time he expects flatter teams and redesigned processes.
Giraldo said he built an agent in under 30 minutes that simulates a panel of AI buyers to score marketing content before it's published. He said it has measurably improved how that content performs.
The panel's one-line advice
- Giraldo: "Experiment and evolve."
- Bhat: Start with targeted use cases and clear outcomes, and build in cost optimization, observability and governance from day one. "Design it in. Don't bolt it on when things fail."
- Archila: "Start with the problem and then bring your people along."
Join the next conversation
"From AI to ROI" returns as part of the AI × Enterprise Panel Series on Wednesday, April 21, 2027, one of 24 free Data Science Connect webinars scheduled for 2027. See the full webinar schedule to register for upcoming sessions, or meet enterprise AI leaders in person at The AI Enterprise Conference in New York City on August 31, 2027.
Frequently asked questions
What percentage of AI pilots fail?
Depending on the study, 42% to 95% of AI pilots never move beyond the pilot stage, according to figures cited by moderator Felipe Archila.
What is the 70-20-10 rule for AI?
A rule of thumb cited by Girish Bhat: put 70% of AI scaling effort into people and processes, 20% into technology and data, and 10% into models and algorithms.
How do you choose which AI pilots to scale?
Evaluate business impact, operational feasibility and sustainability in production, and check whether non-technical users can actually adopt the tool.
What is pilot purgatory?
A term for AI projects that show promise in proofs of concept but never reach production or deliver ROI.
Watch: From AI to ROI: Turning Pilots Into Scalable Impact (Webinar Replay) on YouTube
Event: Webinars · More coverage: Webinars insights
