What does it take to turn AI experiments into measurable results in retail and supply chain operations?

What does it take to turn AI experiments into measurable results in retail and supply chain operations?

In this interview Darius Vasefi speaks with Aaron Millstone, SVP of Customer Strategy and Applied AI at Manhattan Associates, about how AI agents are being deployed, how teams measure their impact, and where human oversight remains essential.

Aaron explains Manhattan’s approach—a “probabilistic brain and a deterministic spine”—combining AI reasoning with defined workflows for reliable execution. He shares a warehouse example showing how faster problem resolution can reduce overtime, improve on-time shipping and lower expedited freight costs.

Key topics:

From AI Experiments to Business Results

A conversation with Aaron Millstone, Manhattan Associates

In this episode of Retail Tech Podcast, Darius Vasefi speaks with Aaron Millstone about putting AI to work across retail and supply chain operations. Drawing on more than 30 years at Accenture, Oracle, NTT DATA and Manhattan Associates, Aaron discusses the technology, implementation work and organizational changes needed to produce measurable results.

Moving from experimentation to measurable value

The conversation opens with the shift from measuring AI adoption through token consumption to asking what the investment actually delivers. Aaron distinguishes improvements in individual productivity from changes that improve an entire operation—such as higher throughput, fewer late shipments, lower overtime or postponing the need for another distribution center.

Unified commerce and the cost of silos

Discussing Manhattan’s Unified Commerce Benchmark, Aaron explains why separate systems for inventory, orders, stores and fulfillment make it harder to meet customer expectations. A shared operational view helps retailers coordinate decisions across channels, personalize experiences and support faster fulfillment.

How Manhattan combines AI reasoning with reliable execution

Manhattan uses third-party language models, with Google Gemini as its default approach, while supplying the business context, tools, security and workflow controls around them. Aaron describes this as a “probabilistic brain and a deterministic spine”: AI helps interpret a problem, while defined software processes execute inventory and order changes.

A practical warehouse example

Aaron describes a customer whose wave coordinator investigates why orders cannot be released to warehouse workers. AI helps trace inventory and allocation problems, reducing an investigation from roughly an hour to half an hour in his example. The benefits extend downstream: work starts earlier, overtime falls, shipments leave more predictably and expedited freight can be reduced.

Experimenting quickly with human oversight

Aaron outlines a roughly 90-day engagement in which Manhattan establishes a baseline, deploys standard and custom agents, and measures the resulting changes. Human review allows teams to test recommendations before authorizing actions that affect inventory or customer commitments.

Faster delivery changes the economics of retail

The discussion connects rising expectations for rapid delivery with the operational systems needed to support it. Faster fulfillment can become a paid service and a source of revenue. Stores also become fulfillment locations, requiring coordination between the shopping experience, store inventory and the wider supply chain.

Manhattan Editions and growing on one platform

Aaron explains Editions as different entry levels into Manhattan Active, rather than a separate product. Essentials, Enterprise and Enterprise Premier let businesses select capabilities appropriate to their operations. A logistics provider can use different editions across small and large facilities while maintaining a common platform.

Product development, APIs and changing interfaces

The conversation covers how customer feedback, product management and partners shape Manhattan’s roadmap. Aaron also discusses AI in software development, the value of Manhattan’s API-based architecture, and interfaces that could dynamically present the information and actions relevant to a particular task.

Choosing models and controlling AI costs

Both speakers question the pressure to always use the newest model. The practical goal is to select a model that performs the task reliably at an appropriate cost. Aaron also acknowledges that model capability still matters: removing access to stronger models created problems for some internal teams.

Personal agents, robotics and future interfaces

The discussion expands to personal assistants, glasses, dedicated AI devices and warehouse robotics. Aaron expects consumer experiences to influence enterprise expectations, while noting that specialized business applications often have clearer tasks and success measures.

The people and operating model behind AI

Agent deployment introduces ongoing questions: when should an agent run, when should it pause, who investigates failures, and what happens after a model changes? Aaron emphasizes adoption, governance, testing and accountability, alongside the technology.

His closing advice is to experiment with an outcome in mind, measure improvements across complete workflows, and build toward a longer-term operating vision.

 

Aaron’s closing advice: experiment with an outcome in mind and focus on improving complete workflows.

Learn more:
Manhattan Associates: https://www.manh.com
Retail Tech Podcast: https://retailtechpodcast.com

Subscribe for more conversations with the people building and operating retail technology.

#RetailTech #AI #AgenticAI #SupplyChain #UnifiedCommerce #ManhattanAssociates