EP110 Agentic AI for Product Management: Building in an Autonomous Enterprise

Eshaan Jain gave me the example that made agentic AI stop sounding clever and start sounding expensive. An agent negotiating a discount on your behalf pushes past the point of sense and sells a product for $5 that cost $10 to buy. Nobody had told it there was a floor. Eshaan spent nearly five years at Amazon on a $40 billion annual contract portfolio, and now advises T-Mobile, so he has watched automation fail with real money attached. **What you'll hear** \* The risk register Eshaan keeps, and which decisions he lets an agent make with no human at all. \* Why a discount pushed from 10 to 18 percent stops the agent, pulls in a person, and rolls back if nobody answers. \* How Amazon taught a machine learning model to read vendor contracts, and the 95 percent accuracy it reached. \* The chain of specialised agents he would build across quote to cash, from catalog lookup to contract review. \* Why the AI programmes still running in 2028 will be the ones measuring carbon, not the ones with the biggest models. "So if you procure the product for, let's say, $10 and the agent is now selling the product for $5, that is a big no no. So that is a critical risk for me." — Eshaan Jain, \[00:26:42] "The AI projects and the programs that are going to last 2027 and will eventually move into 2028 will not be the ones that will have the most powerful models. They will be the ones who have clear visibility on both the cost and the carbon emissions for each of their inferences." — Eshaan Jain, \[00:32:19] **Chapters** \* 00:00 Why agentic AI lands in the back office first, not the chat box \* 03:00 What drives him: the book Agentic Enterprise and an IEEE paper on carbon-aware inference \* 05:44 What agentic AI actually changes about a product manager's job \* 07:30 The one thing AI has not taken over: prioritisation still needs human judgment \* 08:51 Quote to cash explained, and where a chain of agents would sit inside it \* 10:02 Inside a $40 billion portfolio: teaching a model to read Amazon's vendor contracts \* 13:44 What large language models do with contracts that older models could not \* 15:58 The specialised agents he would build for pricing, carts and contract review \* 19:02 Two scaling lessons from Amazon: never underestimate the data, then define accuracy \* 21:11 What he had to unlearn moving from in-house at Amazon to the T-Mobile account \* 23:25 In the boardroom: why product managers now have to think like risk officers \* 24:58 The risk register: severity tiers, and which calls an agent may make alone \* 25:53 Past 10 percent: human approval, the rollback window, and the $10 product sold for $5 \* 27:36 What evaluation and multi-level approval look like before an agent reaches production \* 29:09 The token optimisers' manifesto: the inference cost nobody budgets for \* 30:29 Power, water and pollution: why communities are pushing back on new data centres \* 31:27 Dynamic routing: stop sending every email to a frontier model \* 33:26 One change to make next week: inventory every workflow, then pilot one Eshaan Jain is a senior product manager at Mphasis Silverline on the T-Mobile account, after nearly five years as a senior technical product manager at Amazon and three years as a Salesforce consultant in PwC's digital risk practice. He is also publicity officer for the British Computer Society's open source specialist group. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.productcoalition.com/subscribe