Speak at The Fifth Elephant 2026 Annual Conference

Describe your session in 2 paragraphs

Companies are struggling to get value from AI-for-data agents. They have invested considerable time and energy in onboarding such systems but found usage and adoption lacking. The reasons are manifold - business users lack confidence in the systems, providing the right context is a challenge, people simply don’t know what questions to ask. So rather than using the purpose-built product, teams continue to bombard the analytics teams with questions.

In this 30-min talk, I talk about four ideas that analytics leaders can implement to drive usage and adoption, and get real value out of their AI-for-data products. The material draws on conversations with a few vendor teams, a few enterprise deployments I’ve been close to over the last several months, and my own experience building Babbage Insight, where I ran into the same problems from the other side of the table.

Mention 1-2 takeaways from your session

1. Context is both ingested (from a semantic layer) and learned (from tonnes of unstructured data your company has). We need to leverage both to give agents the right information
2. Even after your product is “good enough”, driving adoption is more an art than science. Business people won’t use the product just because it is “right”. They need to know how to use it, where it fits in their workflows, etc.

Which audiences is your session going to beneficial for?

1. Analytics and product leadership trying to drive adoption of AI-for-data products within their companies.
2. Sales / solutioning / FDE teams of AI products, to know how they can unlock usage in their customers and drive usage

Speaker Bio

Karthik Shashidhar is a fractional data and AI consultant based in Bangalore. Prior to this, he was founder and CEO of Babbage Insight, and SVP of Data at Delhivery.

He has written about data, analytics, and AI on his Substack The Art of Data Science since 2022 (before that, on a personal blog since the mid-2000s). He writes a column for Mint and is the author of Between the Buyer and the Seller. He holds a BTech in Computer Science from IIT Madras and an MBA from IIM Bangalore

Outline / Session Plan

Opening (3 min). A real situation. Agent deployed, exec sponsor enthusiastic on day 1, usage chart looks flat by month 3. What’s the gap between “agent works” and “agent is part of how decisions get made”?

Frame (3 min). What “adoption” should mean for an AI-for-data agent, and what it shouldn’t. Two leading indicators that matter: query sophistication trending up over time, and analyst ad-hoc request volume trending down. Quick note on why raw query volume is a vanity metric.

Lever 1 - Context as exemplar curation (4.5 min). Show the way to the model. Existing analyst SQL, notebooks, dashboards, and recurring investigation paths as the richest context source - richer than a clean semantic layer alone. A concrete correction on the “use the query repository as the source of generated queries” pattern: history is best used for inferring common cuts, metric ambiguities, and analyst habits, while the actual query generation works better from a clean metric definition.

Lever 2 - Pair-drive, don’t train (4.5 min). Sit with two or three real users from a real team on their real questions. Diagnose failures live - is it a context gap, a product limit, or a phrasing issue? Each failure routes to a different fix. Why this builds trust faster than any training session, and why the goal is to shift from “you drive, they watch” to “they drive, you’re not in the room.”

Lever 3 - Trust through consistency and “I don’t know” (4.5 min). Build repeated-run consistency evals - the same question, run multiple times, should produce the same correct logic. An agent that refuses to answer when its context is thin beats an agent that confidently hallucinates. How to calibrate user expectation to what the tool actually has access to, so users can know when to trust an answer and when to escalate.

Lever 4 - Defend value economically (4.5 min). A working use case generating an outsized return can still be throttled because the cost meeting sees the token bill before it sees the value. The fix is to build the economic argument before the first cost review, not after. A concrete pattern: ~$600 of token spend, ~$100K of value created, throttled by a central AI committee because consumption looked high. What it would have taken to defend it.

Close (3 min). The boring data work is the sticky work. Forward-deployed mindset: do the first round of real analysis yourself, document as you go, turn it into reusable context, then let the tool climb from there. End with the line that captures the whole thing: the framework is the scaffold; the judgment is the structure.

Q&A (3 min).