Journal / Kirill Naumov
The Home Depot is replacing store phone menus with a voice agent. The interesting retail AI question is whether the caller reaches a trustworthy answer or the right person.
By Kirill Naumov ·
Imagine calling a store because you need a fitting for a repair today. You are standing beside a leaking pipe. The last thing you want is a recorded voice cheerfully announcing eight departments, followed by a second menu. You want to know whether the right fitting is available, whether you have chosen the right size and when you can collect it. Those are three different questions. A quicker greeting only solves the first few seconds of the call.
The Home Depot says it is replacing traditional menus for calls to its US stores with an AI voice agent built with Google Gemini. The company describes a 50-store pilot, says the system identifies a caller’s intent in under ten seconds and reports that callers reach a solution four times faster than with a traditional menu. Those are company-reported pilot figures, not independent measurements of every store call or every kind of customer problem. A wider US rollout is planned.
The useful design decision is deeper than speech recognition. The agent is connected to the product catalogue and company knowledge. According to Home Depot, it can answer order and availability questions, build a cart using current inventory, text links, begin service requests and route callers to a human associate. That turns “How can I help?” into a potential action. It also raises a tougher question: how does the system know when it should stop?
Take the fitting. A catalogue might know the product name while the store inventory says there is one left. Is it on the shelf, reserved for another order or waiting to be counted? The agent should distinguish “listed in stock” from “I can promise it is ready for you”. If the information is uncertain, the honest answer is useful: explain the limit and offer a person or a stock check. False confidence creates another call, a wasted trip and a customer who has learned not to trust the next answer.
A retailer testing this kind of voice AI should inspect the handoff between systems. What time was the stock record updated? Does the cart respect a local price and the exact variant? Can the assistant distinguish a simple order-status request from product-safety or compatibility advice? Those details are less glamorous than a fluid conversation. They determine whether the conversation helps.
Home Depot says the caller retains a direct route to a human associate. That is important, but the route needs context. A good transfer carries the question, the product or order identified, what was checked and where uncertainty remains. Otherwise the customer repeats the story and the colleague starts from zero. For a complex project, the system might be best at gathering the first facts and arranging the right conversation, not pretending to finish the job.
The associate should also be able to correct an answer. If the agent repeatedly says a product is available when the shelf is empty, that signal needs an owner and a repair path. Otherwise the voice interface scales the same bad promise to more callers.
Before rollout, compare comparable calls by purpose. Did a shopper get the correct answer? Was a trip to the store successful? How often did a person have to repair an automated answer? Did staff receive better-context calls, or simply more difficult ones? For simple requests, time to resolution matters. For compatibility, service and safety questions, correctness and escalation matter more than speed.
The lesson for other retailers is straightforward: conversational AI can remove an annoying menu, but the real product is a dependable path from a spoken problem to a grounded answer or an informed human. If the system knows that boundary, “How can I help?” becomes a good question rather than a shiny new way to waste someone’s afternoon.
Explore the English AI for Retail workshop → · Discuss a retail AI consulting brief
Primary source: The Home Depot’s announcement of its AI voice agent. Pilot figures and system capabilities are company-reported; this evaluation framework is independent commentary, not a client project.