Amazon’s Catalog AI: Lessons for Utilities and Retail

Amazon didn’t improve shopping by building a flashier homepage. It improved shopping by fixing the thing most customers never think about: the product catalog.

Over the past few months, many Amazon listings have gotten noticeably clearer—more images, more consistent titles, more complete specifications. That cleaner catalog is also what powers faster predictive search suggestions while you type. Under the hood, this is driven by Amazon’s Catalog AI, launched in July and led by engineering leader Abhishek Agrawal.

This post is part of our AI in Retail & E-Commerce series, where we track how personalization, search, demand forecasting, and automation are reshaping the customer experience. But I’m going to take a stance: the most transferable lesson from Catalog AI isn’t “use generative AI.” It’s “treat messy information as an operational risk.” That’s as true for e-commerce as it is for energy and utilities—where “catalog data” shows up as asset registries, work orders, tariffs, interconnection documents, and outage notes.

Catalog AI works because it fixes the data layer first

Amazon’s Catalog AI is essentially an automated system for product information management (PIM) at massive scale. The immediate outcomes shoppers see—better titles, fuller descriptions, more accurate attributes—are just the surface.

What’s actually happening is more strategic: Catalog AI collects product information across the web, then uses large language models to fill gaps, correct errors, standardize terminology, and rewrite specifications for clarity. That upgraded structure is what makes search feel smarter.

There’s a parallel here that energy and utility leaders will recognize instantly: you can’t optimize operations or customer experience on top of inconsistent data. If your outage cause codes are unreliable, your reliability analytics will be unreliable. If your asset attributes are missing (manufacturer, install year, configuration), your predictive maintenance models will drift.

The “catalog” is a system of record, not marketing copy

Most companies treat catalogs as presentation. Amazon treats the catalog like operations.

That mindset shift matters because:

• Search and personalization depend on structured attributes, not just free text.
• Customer trust depends on accuracy, especially for specs.
• Operational automation depends on standard language, otherwise every workflow becomes exception handling.

In utilities, the same logic applies to:

• Asset hierarchies (substation → feeder → transformer → meter)
• Equipment specs (ratings, firmware, manufacturer, install date)
• Work orders and inspection notes
• Tariffs and program eligibility rules

When these are incomplete, every downstream system pays the “data tax.”

Predictive search is really demand prediction in disguise

Amazon’s predictive search suggestions look like a UX feature. They’re also a real-time forecasting problem.

Catalog AI supports an experience where, as a shopper types (say, “red mixer”), the system can:

1. Interpret intent (color attribute + product type)
2. Map the query to standardized catalog terminology
3. Suggest likely matches immediately

That’s not far from what utilities do when they forecast customer load or manage flexible demand:

• Interpreting signals (weather, time of day, customer history)
• Mapping them to a structured model (rate class, device ownership, program enrollment)
• Recommending the next best action (price signal, alert, dispatch)

The shared principle: prediction only works when your underlying entities are well-defined.

Retail calls them SKUs and attributes. Utilities call them meters, service points, devices, and constraints.

What retail gets right that utilities often miss

Many utilities try to start AI with grid optimization, DER orchestration, or enterprise copilots—big, expensive initiatives.

A more reliable path is what Amazon did:

• Standardize the glossary
• Automate enrichment
• Improve the “search” layer
• Then scale personalization and optimization

In a utility context, that might look like:

• Standardize asset and outage taxonomies
• Automate extraction of attributes from PDFs, inspections, and notes
• Improve internal search for operations and engineering
• Then deploy advanced forecasting, dispatch, and reliability modeling

From manual glossary building to automated enrichment: the real shift

Agrawal’s team initially built a glossary by manually collecting catalog information (dimensions, color, manufacturer) to standardize descriptions. This mattered most for third-party sellers, who might enter too much, too little, or inconsistent information.

Then the catalog got too large to maintain by hand. That’s the moment many organizations hit: the scale ceiling.

So they moved from manual collection to an AI tool that:

• Pulls product info from across the web
• Identifies missing attributes
• Corrects errors
• Produces clearer titles/specs using large language models

This pattern—manual standardization followed by automated enrichment—is one of the most practical playbooks for enterprise AI.

Utility equivalent: stop treating free text as “unusable”

Utilities sit on mountains of unstructured and semi-structured data:

• Crew notes in work orders
• Inspection narratives
• Interconnection applications
• Equipment manuals and commissioning reports
• Customer emails and call center transcripts

A modern NLP and LLM stack can extract structured fields from this mess. Not perfectly, but well enough to create step-change improvements in:

• Asset registries (fewer missing fields)
• Failure mode tagging (better root-cause analysis)
• Job scoping and dispatch (less rework)
• Customer support routing (faster resolution)

Snippet-worthy truth: If your best knowledge lives in PDFs and notes, your AI program is stuck at “pilot” until you convert that knowledge into structured, searchable fields.

A/B experimentation: the most underrated part of the story

One detail from Agrawal’s Microsoft years deserves more attention: he helped build an online A/B experimentation platform to test features before launch, track metrics, and produce scorecards.

That capability—fast, disciplined experimentation—is why teams can safely improve search ranking, relevance, and UX without blowing up the customer experience.

Retailers already live and die by controlled experiments: conversion rate, revenue per search, cart size, and return rates.

Utilities, by contrast, often ship changes slowly and evaluate them with blunt instruments. But experimentation is becoming non-negotiable as utilities digitize customer engagement and grid operations.

What “A/B testing” looks like in energy & utilities

You can’t A/B test a feeder the same way you A/B test a checkout page. But you can run controlled tests in places that matter:

• Customer communications: message timing, channel, and content for outage alerts or demand response events
• Call center workflows: agent assist prompts, knowledge suggestions, and wrap-up automation
• Field operations: routing algorithms, parts pre-staging logic, or inspection prioritization policies
• Digital self-service: payment plan recommendations, high bill investigations, and personalized energy tips

A disciplined experimentation loop reduces risk and makes AI measurable.

What energy and utility leaders should copy from Catalog AI

Catalog AI is a retail example, but the operating model maps neatly to regulated infrastructure businesses.

Here’s what I’d replicate—without pretending utilities are e-commerce.

1. Build a “utility catalog” mindset

Define your core entities and attributes the same way retail defines SKUs.

Examples:

• Transformers: kVA, impedance, manufacturer, install date, location class
• Meters: type, firmware, comms module, last read quality
• Service points: rate class, premise type, DER present, program eligibility
• Work orders: job type taxonomy, safety tags, true cause codes

Then treat missing attributes as a defect, not a nuisance.

2. Use LLMs for enrichment, not decision-making first

A safe, high-ROI early use case is enriching and standardizing records:

• Extract structured fields from PDFs and free text
• Normalize terminology across systems
• Identify contradictions (two different install years)
• Flag low-confidence records for review

This is how you earn trust and build momentum.

3. Pair predictive models with better “search”

Amazon’s predictive search feels magical because the catalog is clean.

Utilities can create similar wins by improving internal and customer-facing search:

• Customers: “Why is my bill high?” “How do I apply for EV rates?”
• Agents: “What’s the eligibility rule for this program?”
• Engineers: “Show past failures for this asset family.”

Search is where data quality becomes visible fast.

4. Measure with scorecards that executives actually read

Amazon expects Catalog AI to drive material sales impact (a July report cited a projected $7.5B lift). Whether or not you focus on that specific number, the message is clear: the initiative is being measured as a business system.

For utilities, scorecards should tie to:

• Reduction in truck rolls
• Shorter outage restoration times
• Higher first-contact resolution
• Lower call handle time
• Improved forecast accuracy
• Lower rework in field jobs

If your AI program can’t explain value in these terms, it won’t survive budgeting season.

Where this fits in the AI in Retail & E-Commerce narrative

In this series, we often talk about AI for personalization, dynamic pricing, and customer behavior analytics. Catalog AI is a reminder that those wins depend on something less glamorous: data standardization at scale.

Retailers that get this right build experiences that feel effortless. Utilities that get this right build operations that feel controlled—fewer surprises, fewer exceptions, faster service restoration, and better customer communication.

If you’re leading AI in an energy or utility organization, here’s the forward-looking question worth sitting with: what’s the equivalent of your “product catalog,” and how much of your operational friction is just missing or inconsistent information?

The fastest AI wins usually come from cleaning the language your business runs on—names, attributes, codes, and the messy text people type when they’re busy.

If you want leads from AI initiatives in 2026, don’t start by promising big transformations. Start by promising fewer defects in the data layer, a better search experience, and measurable operational outcomes.