Lowe's Achieves $1.2M in AI Cost Avoidance via OpenAI Model Optimisation

The Challenge: Scaling Mylow Without a Cost Framework

Lowe's decision to partner with OpenAI on its Mylow virtual adviser programme was one of the most ambitious AI deployments in US retail. When Mylow Companion rolled out to all 1,700+ Lowe's stores simultaneously in May 2025 — handling nearly one million customer questions every month — the scale of the API commitment became commercially significant almost overnight.

The technology was working: conversion rates more than doubled when customers engaged with Mylow during online visits, and in-store customer satisfaction scores rose 200 basis points when associates used Mylow Companion. But the commercial structure underpinning the deployment had not kept pace. Lowe's was running approximately 50 AI models across search, recommendation, demand planning, and pricing — alongside its Mylow customer-facing applications — all on OpenAI API pricing that had never been benchmarked or renegotiated since the original partnership was established.

When Lowe's technology procurement team conducted an annual AI spend review, the projection showed API costs tracking toward $2.4M over the following 24 months at then-current usage rates and list pricing. There was no volume discount in place, no committed spend tier, and no model-tiering strategy to route cheaper queries to lower-cost models. Lowe's engaged Redress Compliance to assess whether the commercial structure matched the deployment's actual profile.

The Approach: Model Tiering, Committed Spend, and Rate Optimisation

Query Classification and Model Routing

The first phase of the engagement was a token consumption audit across Lowe's three primary OpenAI use cases: Mylow customer enquiries, Mylow Companion associate support, and product metadata enrichment. The audit revealed that approximately 68% of Mylow queries were transactional in nature — order status, store hours, basic product availability — and did not require the full capability of the flagship model being used for all traffic.

Redress Compliance modelled a tiered routing architecture: transactional queries routed to a lightweight model at 1/12th the per-token cost, while complex product advice and project planning queries continued on the flagship model. Applied to actual monthly token consumption, this single change projected $740K in savings over 24 months with no degradation to customer outcomes.

Committed Spend Negotiation

With Lowe's usage trajectory clearly established and growing, Redress Compliance negotiated a committed annual spend tier with OpenAI that unlocked a 28% volume discount versus list pricing on the traffic that continued on the flagship model. The agreement included rate-lock protection for 18 months, eliminating exposure to mid-contract price adjustments on Lowe's highest-volume model.

Async Batch Processing

Product metadata enrichment — which accounted for approximately 14% of total token spend — was restructured from real-time API calls to asynchronous batch processing. OpenAI's batch mode offers approximately 50% off standard token pricing with up to 24-hour latency, which was entirely compatible with the metadata pipeline's overnight batch cadence. This change delivered a further $180K in cost avoidance without any engineering rework beyond a configuration change.

The Outcome: $1.2M in AI Cost Avoidance

The restructured commercial model reduced Lowe's projected AI API spend from $2.4M to $1.2M over 24 months — a 50% reduction — while maintaining the performance characteristics that had delivered the conversion rate and customer satisfaction improvements that justified the Mylow investment. The engagement was completed in four weeks, before Lowe's annual technology budget was finalised.

Lowe's also exited the engagement with a reusable cost governance framework: a token consumption model that can be updated monthly, a vendor review protocol for its 50+ AI models, and a template for renegotiating AI vendor agreements as usage scales.

Key Lessons for Enterprise AI Buyers

Three patterns from this engagement apply broadly to enterprises scaling AI deployments. First, most large-scale deployments are running a single model tier across heterogeneous query types. Classifying queries and routing by complexity is almost always the highest-value optimisation available — with no impact on user experience. Second, committed spend discounts are available but not offered proactively. OpenAI will negotiate volume tiers for buyers who demonstrate consistent usage, but only when asked by a buyer who knows what to ask for. Third, batch processing is underutilised for non-real-time workloads. Any AI pipeline that does not require sub-second latency is a candidate for 50% cost reduction through async batch mode.

For retailers specifically, AI cost management is becoming a board-level issue as deployments scale beyond pilot programmes. The gap between what AI costs at list pricing and what it costs under a well-structured enterprise agreement can be substantial — and unlike many software categories, AI pricing is still negotiable. Buyers with market data and commercial expertise consistently achieve materially better outcomes than those who accept the first proposal.