The Fifth Elephant
The Fifth Elephant

The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

All submissions

Sanket Patil

@sanketvpatil

Agents act on what they are given: data platforms and data quality in an agentic world

Submitted Oct 11, 2026

🎤 I am submitting: To speak I have a submission for: A talk - 30-40 mins

Submission type
Talk/session proposal – 30–40 mins

Description

Most current writing on agentic AI asks which decisions can be automated, and to what extent. It says little about what the agent acts upon. In retail, an agent that recommends a price change, flags an assortment gap or rewrites product content depends on a chain it does not supply: data gathered from many sources, normalised into stable representations of products, sellers and categories, and enriched by models that classify, extract and match. When a person makes such a decision, much of the error in that chain is absorbed at the point of decision: an analyst with domain understanding questions an implausible price, sets aside a stale record, or notices that two products being compared are not the same. When an agent makes the decision, that judgement is absent unless it has been designed into the system, and the error is acted upon at volume. The case for a robust data platform, reliable AI capabilities and explicit data quality mechanisms therefore becomes stronger as decisions move to agents.

Drawing on retail decisioning across pricing, assortment and content, built largely on public data crawled from the web, the talk describes what the agent needs underneath it. It needs representations stable enough to reason over, with the customer’s own rules made explicit. It needs decisions that carry their evidence, so that a domain expert can reconstruct and dispute any of them. It needs data quality represented as a state – trusted, suspect, quarantined, corrected – with agents permitted to act only on trusted data. Quality must be measured from the perspective of the consumer of the data, by a function independent of the one that produces it. Finally, it needs a correction loop, in which each human correction refines the rules, the shared knowledge and the evaluation set, so that the same error does not recur.

Takeaways

  • Removing the human from the decision raises the bar on the data platform beneath it.
  • The models are increasingly available to everyone. The data platform, the quality states, the record of corrections and the evaluation sets are not, and they determine how far a class of decisions can safely be handed to an agent.

Audience
Data and ML engineers, data platform leads, and product leaders building agentic systems on enterprise data. No retail background is assumed.

Bio
Sanket Patil builds data and AI products at DataWeave, which provides retail data and analytics – pricing, assortment and digital shelf – built largely on public data from e-commerce websites. He drives innovation and AI strategy. He holds a PhD in distributed systems from IIIT Bangalore, where he also teaches a course on Network Science.

All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures