AI on the Shop Floor

This presentation is password protected.

Incorrect password
AI in Store Operations
1 / 19
Use arrow keys or scroll
Thought leadership · June 2026

AI on the shop floor

Agentic AI in store operations: strategy, implementation and evidence from Kmart Australia.

Fabio Oliveira
Head of Innovation and Design, Kmart Australia
The starting point

Retail AI looks at customers. The frontline is the underserved frontier.

Context

Where this evidence comes from

300+
Kmart stores across Australia and NZ, part of the Wesfarmers group
~55,000
Store team members, the user base every frontline tool must serve
28
People across Innovation, Innovation Experience Design and Digital Stores
Strategy

Five principles behind the approach

  1. Centralise innovation, embed it in operations

    Decentralised innovation does not survive large incumbents. A small, skilled, dedicated team working alongside operations does.

  2. Separate explored from validated

    What people say and what models predict is exploration. Only what is observed in real store conditions counts as validation.

  3. Innovation designs and validates, operations executes

    Once a solution is validated, it transitions to the operating business. The handover is planned from day one, not at the end.

  4. Fixed time, variable scope

    Six-week bets with defined gates. If confidence drops below the threshold mid-cycle, the bet is stopped, not stretched.

  5. Build vertically, not horizontally

    One thing working end to end in a real store beats five things half-built in a lab.

Portfolio

The store operations AI portfolio

Assistant

Team Member Assistant

Voice-first AI assistant answering product, policy and process questions on the devices team members already carry.

Agentic

Auto PDT

Kmart's first agentic implementation: an LLM agent that operates existing store applications by voice.

Agentic

Sidekick

Agentic face-to-face shopping assistant connecting a consumer LLM to product search and in-store navigation.

Decision support

Zone Manager's Best Friend

AI decision support for zone managers, researched and shaped with the managers themselves.

Process + AI

One Way Flow + CV

Replenishment flow redesign with computer vision layered on top, measured store by store.

The next sections go deep on the two frontline implementations: the assistant and the agent.

01 · Team Member Assistant

One assistant, on the device they already carry

It began with a question: what if a casual team member on their third shift could be as productive as someone with three years on the floor?

The Team Member Assistant (TMA) is a voice-first AI assistant for store team members, running on the Zebra handheld devices already used for daily store work.

It answers the questions that dominate a shift: is this in stock, where is it, what does it cost, what is the process for this, what is the policy on that.

  • The problem

    Information lives in many systems. The team member, often with a customer waiting, has to stop and navigate to find it.

  • The bet

    If an assistant answers in seconds, by voice, without interrupting the physical work, it gets used. If it gets used, it changes service speed.

01 · Team Member Assistant

How it was built

01 · Team Member Assistant

The architecture at a glance

Experience layer

Custom React Native front end: voice capture, push to talk, image carousels, headset and device integration. This is where adoption is won or lost.

Owned

Intelligence layer

Enterprise LLM platform (Gemini) for retrieval and reasoning. Cloud-agnostic by design: the provider can change without rebuilding the experience.

Swappable

Data layer

Real-time availability from the data warehouse, refreshed every two hours, plus policy and process content restructured for retrieval.

Live

The split is deliberate: the experience is the moat with users, while the model underneath is a commodity that keeps improving.

01 · Team Member Assistant

The usability principles that made it stick

01 · Team Member Assistant

The evidence

100+
Team members using the tool daily, including night-fill casuals
1,000+
Questions asked through the system, and climbing
100%
Of usage came through voice input, none typed
98%
Of questions were product-related: availability, price, location
245
Questions from 26 users in a single weekend after fleet-device deployment
4s
Answer time for common questions, down from 30 seconds at launch

Single-store pilot at Kmart Broadmeadows, deliberately run past the novelty window (3+ weeks) so adoption could be separated from curiosity. Handled policy questions and filtered inappropriate queries without incident.

01 · Team Member Assistant

What the pilot taught us

02 · Auto PDT

The first agentic implementation

Auto PDT is a voice layer that operates the existing applications on Kmart's handheld terminals. The team member speaks; an LLM agent does the tapping.

The agent reads the live screen through Android's accessibility tree, decides the next action, performs it, re-reads the screen and repeats until the task is done. No changes to any existing app or backend system.

  • Why agentic rather than scripted

    Hard-coded screen maps break every time an app updates. The agent adapts to whatever is on screen: if a button moves, it finds the new path.

  • Why it matters

    It turns a six-tap stock check into a spoken sentence, on hardware the business already owns, without a single integration dependency.

02 · Auto PDT

Anatomy of the agent

01

Voice listener

Hardware button trigger, so no false activations on a noisy floor. On-device speech-to-text.

02

Command interpreter

An LLM resolves natural speech into a structured intent from a bounded action set.

Agentic loop
03

Screen reader

Reads live app state through the accessibility tree: fields, buttons, visible text.

04

Action engine

Read, decide, act, re-read, until the task is done or it stops gracefully.

05

Overlay UI

Minimal floating layer: listening state, confirmation, result. Collapses in seconds.

Typically 3-5 LLM calls and one to two seconds of navigation per task. The interpreter works on intent, not exact words, so partial recognition still works. No changes to the apps being operated.

02 · Auto PDT

Trust is the gating factor, so it is designed in

A team member burned once by a wrong markdown will never use the system again. Graceful failure is the single most important behaviour.
02 · Auto PDT

The economics of agentic, and the discipline around it

$0
Hardware spend. Existing devices, existing apps, existing LLM APIs
<1¢
Per voice command, at 3-5 small LLM calls per action
<4s / 90%
Cycle gates: end-to-end latency and accuracy over 50+ controlled runs
03 · The wider portfolio

Beyond the floor: the same playbook elsewhere

04 · Strategy in practice

Remove waste, eliminate variability, then digitise

01

Remove waste

Some processes should be eliminated, not automated. The simplification work cut whole activities before anyone wrote a line of code.

02

Eliminate variability

Digitising a broken manual process does not fix it; it just makes the variability visible faster. One Way Flow standardised the physical work and delivered 41.4% on its own.

03

Digitise, then make it agentic

AI lands best on a process that is already simple and stable. The assistant, the agent and the vision layer all compound from that base.

There is a limit to how far digital can go in a manual process. Done in the wrong order, digital only amplifies the noise.
05 · Where this goes

From answering, to operating, to orchestrating

Proven in store

Assistant

Answers on demand. The Team Member Assistant compresses information retrieval to a spoken question.

In build, trust-gated

Agent

Operates the tools. Auto PDT performs the work inside existing systems instead of just answering questions.

The trajectory

Orchestration

Humans direct intent. A manager directing dozens of store-level agents; attention becomes the scarce resource and interfaces shift from operation to direction.

The frontline voice tools are the first rung. The design rules they proved, voice first, bounded autonomy, graceful failure, trust earned in phases, are the same rules the whole ladder will need.
Closing

What I would take into any retail operation

  1. Start where the hands are full

    The frontline is the underserved frontier, and the fastest place to prove value.

  2. Evidence from real stores beats demos

    Pilot in production conditions, past the novelty window, and let usage rewrite the roadmap.

  3. Design trust before capability

    Bounded actions, graceful failure and confirmation on writes are what make agents deployable.

  4. Voice in, text out, co-created on the floor

    Usability principles discovered with users are the adoption strategy.

  5. Sequence it

    Simplify the process, then digitise it, then make it agentic. In that order.

Fabio Oliveira · June 2026
fabio.so · Questions and discussion welcome