All work

03 · Agentic AI · Tata CLiQ

TARA: designing a shopping assistant with no reference pattern to copy

The safe call was a search bar wearing a chat costume. I pushed for an agent that asks before it answers, because the real cost in the funnel wasn’t missing search. It was irrelevant results from under-specified queries.

Context

No precedent, and a moving target

Tata CLiQ wanted an AI-powered shopping assistant. There was no internal precedent and no competitor pattern worth copying closely.

The underlying model behaviour was also still being defined by engineering while design work was underway, so we were designing for a system whose capabilities were changing under us.

The decision

Chat wrapper or conversational agent?

The fork

Should TARA behave like a search bar with a chat wrapper, or like a genuine conversational agent that can ask clarifying questions, hold context, and narrow intent over multiple turns?

Option A

A search bar with a chat wrapper

Safe, familiar, low-risk.

Option B · my call

A genuine conversational agent

Asks clarifying questions, holds context, narrows intent over several turns. Higher-risk, higher-ceiling.

Why this, not that

The problem wasn’t search. It was vague queries.

The biggest cost on CLiQ’s discovery funnel wasn’t a lack of search. It was irrelevant results from under-specified queries.

A chat wrapper doesn’t fix that. A system that asks one good clarifying question does. So I pushed for the conversational-agent direction, not the chat-wrapper direction.

How we got there

Building a UX vocabulary that didn’t exist yet

Agentic AI in commerce doesn’t have an established UX vocabulary yet, so we had to build one as we went.

I mentored 3 designers through research specific to agentic AI: mapping what users actually needed against what they typed. That gap is exactly what an agent has to close.

What happened

The numbers while TARA was live

  • ~2× cart additions and PV/V
  • 1.7× session length
  • A measurable drop in irrelevant-search abandonment

Where it stands now. TARA was paused after launch. That was a cost and roadmap call from the business, not a reflection on adoption or the design. The metrics above are what it delivered while live, and the decisions documented here are the strongest evidence of how I work under ambiguity, whatever the feature’s current status.

What I’d argue

Most “AI features” bolted onto e-commerce today are search bars wearing a chat costume. The real design work in agentic commerce is deciding what the system is allowed to not know yet, and how it asks for the rest. That’s a product decision as much as an interaction one.

Back to the start · 01 Systems for behaviour

The nudges framework: one system instead of a dozen one-off banners