Home Depot let AI answer store phones. Callers got help 4x faster

In a 50-store pilot, Home Depot's AI phone agents understood callers in 10 seconds and got them to a fix 4x faster. Cost and error rates are not published.

The Home Depot, the home improvement retailer with 2,359 stores across the US, Canada and Mexico, has replaced the phone menu on calls to its stores with AI voice agents, and says that in a 50-store pilot the agents understood why a customer was calling in 10 seconds and got the customer to a solution four times faster than the old menu. The figures are in the company's press release of 22 April 2026, which also says the system will cover all US stores within a year. Home Depot has not published what it costs, how many calls it finishes without a person, or how often it gets things wrong.

This week brought another story about an AI agent loose on the internet doing things nobody asked it to. That is the fear most operators have about agents that act rather than talk. Home Depot's deployment is the useful counter-example, because it shows what a large retailer actually lets an agent do on a live customer line, and what it keeps for people.

What the phone agent actually does

A call to a Home Depot store used to start with a menu: press one for this, two for that. The new system starts with a question. The caller says why they are ringing, in their own words and, with real-time translation, in their own language. The agent works out the intent and either deals with it or hands over to an associate.

The release lists the jobs it finishes on its own: checking the status of an order, confirming whether a product is in stock, and giving store information such as hours. Those are read-only answers from systems the company already runs. It also lists a second tier that changes something: opening a service request, texting the caller a link to a cart it has already filled, and helping complete a purchase by phone. A third capability, describing a project and having the agent build the shopping list from live inventory, is written in the future tense. Treat it as planned, not running.

The agent takes the routine call. The associate keeps the hard one and the customer standing in the aisle.

The human exit is part of the design, not an afterthought. Both company documents say a caller always has a direct path to an associate. Gartner's customer service analyst, commenting to Customer Experience Dive at launch, made the same point from the other side: noise, accents, bad signal and high-value purchases such as appliances are where a voice agent should hand over.

What did four times faster actually measure?

It measured the agent against the phone menu, not against a person. Home Depot says callers reach a solution four times faster than by navigating the traditional menu, and that the agent identifies intent in 10 seconds. It does not say how long the old route took in minutes, what counts as a solution, or what share of calls ended without an associate. The comparison is real, but it is against a low bar.

The second claimed result has no number at all. Pilot stores reported higher job satisfaction among associates, who had more time for shoppers in the store. That is plausible: a ringing phone pulls someone off the floor at the worst moment. But it is a company statement, not a measured figure, and it should be read that way.

What did it cost?

Home Depot has not said. The release names the platform, Google Cloud's Gemini Enterprise for Customer Experience, and Google co-published it, but there is no price, no headcount change and no contract value. The company does not say whether any phone roles were removed. Its emphasis is on associates spending the freed time in the aisle.

What we can say is how cautious the rollout was. Fifty stores is about 2% of its 2,359 stores. The company tested on a small slice, published only speed figures, and gave itself a year to reach every US store. That is a pattern worth copying even if the price is not known.

Does this apply to a 200-person retailer?

Yes, more directly than most AI deployments, because the problem is identical at any size: the store phone rings while the only person who can answer is helping a customer. What changes at small scale is whether you need software at all.

Our own arithmetic, not Home Depot's: take a retailer with eight stores, each taking 40 calls a day, and suppose half are order status, stock checks or opening hours. That is 160 routine calls a day across the chain. At two minutes each, roughly five staff hours a day are spent answering questions a system could answer. Whether that justifies a voice agent depends on prices you will need quotes for. Whether it justifies fixing the problem does not.

What to do on Monday

First, count. For two weeks, have each store tally incoming calls by reason: order status, stock, hours, returns, complaints, other. Home Depot built its first tier around exactly the boring reasons, and most chains find those are the majority.

Second, check the data the agent would need. An agent can only confirm stock if your store-level stock figure is right. If staff routinely walk to the shelf to check, fix that before you buy anything that speaks.

Third, if you pilot, copy the order of trust. Start with answers that change nothing: order status, stock, hours. Add actions that create work for a person to review, such as a service request, next. Leave taking payment for last. Keep the press-zero route to a person, and measure transfers and repeat calls, the two figures Home Depot did not publish.