Data leaders from AtScale, Snowflake, Denodo, Fivetran and Revefi shared how executives can separate AI hype from real opportunities. They offered criteria for picking use cases, advice for CIOs under board pressure, a framework for build-versus-buy decisions and examples of projects that delivered measurable returns.
Key takeaways
- Apply the litmus test. Mariani asks whether a project would still happen if AI weren't in its name.
- Check readiness and adoption, not just data. Bohan said transformational AI disrupts people and processes, so pick projects users will adopt.
- Start where trust already exists. Natural language queries on a vetted semantic layer, or a data team's request backlog.
- Stay flexible on build vs. buy. Panelists warned against the sunk cost fallacy and building your own LLM.
- Improve an existing process first. A baseline makes ROI easier to prove to the CFO.
"If AI wasn't in the name of the project, would I do it?" That is the litmus test Dave Mariani, CTO and co-founder of AtScale, offered data executives in the Data Science Connect webinar "From AI to ROI: A Data Executive's Guide to Impactful Enterprise Use Cases." Five panelists discussed how to pick AI use cases, handle board pressure, decide whether to build or buy, and prove a return.
Brian Mink, DSC co-founder and president, moderated the panel with Mariani; Nikolai Scholz, Senior Product Manager at Snowflake; Kevin Bohan, Director of Product Marketing at Denodo; Shiva Mogili, Director of Product Management at Fivetran; and Girish Bhat, SVP at Revefi.
How can executives tell real AI opportunities from hype?
Every panelist started with the business problem. "If the use cases are not clear and it cannot be tied to a measurable business outcome, it's probably hype," Bhat said. If a task is just boring and manual, he added, plain automation may be enough.
Mariani said boards are pressuring technology leaders to show AI progress, which can lead to force-fitting AI where another tool is better. Bohan added a step he said is often overlooked: assessing whether the organization is ready and whether users will adopt the change. Transformational AI is also disruptive to people and processes, he said.
What criteria should guide AI investment decisions?
Mariani asked three questions: Does it solve a real business problem? Is the data in the shape and quality needed? Can it scale? For analytics use cases, he said, a semantic layer supplies the business context generative AI needs.
Bohan described a "success blueprint": define measurable outcomes, get stakeholder buy-in early, record baseline costs and timelines, and track results on a value dashboard. He also recommended lining up several projects so a first win builds momentum. Scholz suggested putting on a product manager hat, weighing trade-offs and urgency. If the value is large enough, he said, teams will find a way to get the data in shape.
Mogili contrasted big-bang transformations, which need detailed up-front analysis, with small experiments that can be measured and dropped if stakeholders don't find value. He prefers the experimental approach.
What should a CIO do under pressure to "do AI"?
Responding to an audience question about boards pushing for AI without a clear use case, Mariani said many AtScale customers start with natural language queries on top of a semantic layer they already trust for BI. In his experience, users skip dashboard-style questions and simply ask for insights, and the LLM runs many queries to surface patterns they hadn't seen. Because guardrails already exist, he said, it is a way to start without "unleashing a beast."
Bohan suggested looking at the data team's backlog. Teams that wait two or three weeks for data requests are good candidates for natural language self-service, which frees data teams for strategic work. Bhat recommended picking acute pain points and breaking them into tasks, while Mogili and Scholz suggested gathering ideas from business champions and enabling employees to use AI day to day.
Should enterprises build or buy AI solutions?
AI features embedded in individual applications work well for problems inside that application, Bohan said, but most enterprise use cases span multiple systems and need integrated data and context. Scholz said the decision is a spectrum, shaped by requirements, data gravity and whether a business is niche enough to justify building.
Mogili urged agility and warned against the sunk cost fallacy. He described a large professional services customer that built its own retrieval chat application, found that ChatGPT with custom knowledge sources gave better results, and scrapped the internal tool.
Bhat laid out three routes: do it yourself, use a platform provider, or choose an independent vendor that works across platforms and models. "Please do not build an LLM or a frontier model on your own unless you have deep pockets," he said. Mariani added that companies should keep models and tools interchangeable. "You should actually architect for having no idea where things are going to go," he said.
What AI use cases have delivered measurable ROI?
Mogili described a Fivetran customer, a large Australian bank, that first centralized its data, then consolidated years of support interactions to automate common customer questions. It set metrics for resolution time, response time and chat NPS, rolled out in phases, and saw about a 50% improvement in some key measures, he said.
Bhat said one large Revefi customer saves more than 25% on data platform spend using predictive AI to analyze workloads, queries and usage. Scholz described New American Funding, a mortgage lender that moved document extraction from a vendor and manual work onto Snowflake, reaching production after about a month of testing and a month of implementation. Internally, Snowflake's audit team uses an agent to review order forms for non-standard terms.
Projects that improve an existing process are easier to justify because they have a baseline, Mariani and Scholz agreed. For new, greenfield capabilities, Mariani said, usage is a reasonable proxy for value.
The panel's closing advice
- Bohan: Think strategically but act incrementally, one business problem at a time.
- Scholz: "No AI strategy without data strategy," including unstructured data.
- Mariani: Start with a trusted data foundation and, for analytics, a semantic layer.
- Bhat: Make reliable data and data cost management your first AI investment.
- Mogili: "Treat this as a startup," with limited resources and pressure to prove value.
Join the next panel
Data Science Connect's 2027 schedule includes 24 free live webinars in 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 data and AI leaders in person at The AI Enterprise Conference on August 31, 2027, at Pier Sixty in New York City.
Frequently asked questions
How do you tell if an AI use case is hype?
Panelists said a use case that isn't clear or can't be tied to a measurable business outcome is probably hype; Mariani asks whether the project would happen if AI weren't in its name.
Should companies build or buy AI solutions?
The panel said it depends on requirements, data gravity and niche needs. App-embedded AI works for single-system problems, but most use cases span systems, and teams should drop custom builds when better tools appear.
What is a good first AI use case for enterprises?
Mariani recommended natural language queries on a trusted semantic layer, while Bohan suggested giving backlogged business teams self-service access to data through natural language.
How do you measure ROI on a greenfield AI project?
Mariani said that for new capabilities without a baseline, usage and adoption are a reasonable proxy for value.
Watch: From AI to ROI: A Data Executive’s Guide to Impactful Enterprise Use Cases on YouTube
Event: Webinars · More coverage: Webinars insights
