Sudharsana.
← Back to selected work

Professional work · Applied AI · technical discovery · delivery

Agents that know which customer they are talking about

Sales and Marketing agents built on the customer data the business already trusts.

Walmart Global Tech · applied AI

Ask the account

A seller asks a question. Watch which sources the agent reads, what it answers, and what happens when the records underneath are not joined.

A seller's question goes to an agent that reads behavior, intent, engagement and a shared customer identity, then answers with the account that crossed a signal threshold. Without the shared identity, one company's signals split across three records and none crosses the threshold.Which accounts show buying intent this week?BehaviorIntentEngagementShared identityAgentreads first,thenanswersAnswerNorthwind TradersSignals this week:pricing visits,a demo request,a reorder spike.Cites intent, behaviorIntent signals per accountthresholdNorthwind TradersContosoFabrikam

Grounded in what accounts did

The agent reads behavior and intent signals, joined through one shared customer identity, and answers with the account and the evidence behind it.

A seller can check every claim, because the answer names its signals.

A different question reads different sources

Re-engagement is about attention fading, so the agent leans on engagement and behavior instead of intent. The same identity sits underneath; the evidence on top changes.

The question decides which signals matter, not the other way round.

Without one identity, intent disappears

The same company arrives as three records. Each holds one signal, none crosses the threshold, and the agent confidently reports nothing. The model did not fail. The foundation underneath it did.

This is why the identity work sits underneath the agents.

Illustrative accounts and signals. The point is the join: every signal reaches the agent through one customer identity.

The question behind the question

A seller rarely needs more data. They need a next step: which account is warming up, who has gone quiet, what to say first.

I built Sales and Marketing agents that take those questions in plain language and answer them from the business’s own customer behavior, intent, and engagement data.

The decision: start from the identity

The tempting version hands raw events to a model and lets it sort them out. I went the other way. The signals are curated first, then joined through the shared customer identity in the next story.

That join is the difference between a useful answer and a confident miss. If one organization lives in three records, its intent splits three ways and no single record looks interesting. Play “Without a shared identity” above to watch the agent walk straight past the account that matters.

Built for the person using it

The data work sits underneath and the experience sits on top. The interface turns business context into something a seller can use in the middle of a working day.

Around the agents, I built an evaluation and telemetry loop for a customer-facing assistant, reading conversation patterns to decide what to improve next. An internal tool applied the same thinking earlier in the process: checking whether a team’s data is ready before anyone commits engineering time to an AI idea.

No new source of truth

These agents run on the same platform as the revenue reporting that surfaced more than $750M in order GMV. They did not need a new source of truth. They needed the one the business already had, in a shape a model could use.

The model was the easy part. Knowing the customer was the work.

Let’s talk ↗