The DAIS 2026 Speaker Spotlight is a series where we hand the mic to the speakers heading to Data + AI Summit and let them answer five short questions — in their own voice, no press-release polish.
Below, Mahavir Teraiya on what changes when agents go from one-off apps to entire fleets in production — and the lakehouse control plane adidas built to keep cost, governance, and ROI honest. Lightly edited for length — otherwise, the words are his.
“
A per-run trace tells you what happened on one call. It doesn't tell you what a fleet of agents actually costs, where it's leaking money, or what quietly breaks the minute someone tweaks a prompt.
— Mahavir Teraiya
The topic
What is your talk about, and who is it for?
For anyone running AI agents in production: this is the story of how adidas built an agent digital twin on Databricks, a lakehouse control plane that ties every agent, tool, prompt, retrieval, and model call back to governance, cost, and ROI.
Why this, why now
What's changed in the last 6–12 months that makes this topic urgent right now?
Something interesting just happened. Agentic AI quietly crossed from demo into real production. Spending is tracking toward $1.5T in 2025, with GenAI alone at $644B. And yet Gartner is predicting that more than 40% of agentic projects will be canceled by 2027 because costs balloon and controls can't keep up. The part that keeps me curious is this: a per-run trace tells you what happened on one call. It doesn't tell you what a fleet of agents actually costs, where it's leaking money, or what quietly breaks the minute someone tweaks a prompt. Closing that gap is the real gate to scale.
The personal stake
Why are you the person giving this talk?
I live inside this problem every day at Databricks with various customers. Adidas running 200+ serving endpoints, 300+ data and AI products, 6,000+ registered models, and over 600,000 pipeline runs, and at that scale every governance question gets harder and every cost surprise gets a lot bigger. We ended up building the digital twin because the familiar playbook of trace, alert, hope had quietly stopped working for us. So this isn't a slideware pattern. It's the system we actually rely on to keep agents shipping safely and on budget.
What you'll leave with
What will someone be able to do on Monday morning that they couldn't do before?
You'll walk out with a concrete Databricks pattern you can start wiring up on Monday. How to feed MLflow Tracing into a lakehouse control plane. How to roll up per-hop unit economics from Unity Catalog and system tables. How to spot cost leaks long before they show up on the invoice. And how to honestly attribute ROI to specific agent behaviors instead of waving your hands. Plus the eval and guardrail loop we lean on so a prompt or tool change doesn't quietly break production overnight.
The bigger picture
How does this fit into where Databricks — and data and AI more broadly — is heading?
What I find genuinely interesting right now is how agents are shifting from one-off apps into entire fleets, and the lakehouse is quietly becoming the natural place to govern them, cost them, and prove them out. Databricks has been pulling the pieces together (MLflow, Unity Catalog, system tables, AI Gateway) into something that's starting to look a lot like an agent control plane. The digital twin is what you build on top to make agentic AI both auditable and economically defensible. That's the direction the whole space is moving.
Part of the DAIS 2026 Speaker Spotlight series — more voices dropping in the weeks ahead. Got a DAIS speaker you'd love to hear from next? Mention them in the comments — we're always listening.
