Chewy's AI assistant settles 30% of chats. The savings are $50m a year

Chewy says its in-app assistant Cai resolves about 30% of chats without a person, and expects AI to save about $50m a year by fiscal 2027.

Chewy, the US online pet retailer, says its in-app AI assistant, Cai, resolves about 30% of the customer chats it handles without a person. Chief executive Sumit Singh gave the figure on the 9 September 2026 earnings call, and said Chewy expects its AI work to save low tens of millions of dollars this fiscal year and about $50 million a year by fiscal 2027.

It is one of the few retailers to put a self-service rate and a dollar target on the record together. Both numbers come with conditions that matter more to an operator than the headline.

What Chewy actually built

Cai, short for Chewy AI, is a chat assistant inside the Chewy mobile app. Singh said on the call that it handles the most common reasons customers get in touch: where an order is, shipment status, changes to Autoship, Chewy's repeat-delivery subscription, and account management. Automated returns and refunds sit inside it, backed by machine learning Chewy already had in place. Singh described several agents working under an orchestrator that routes each request.

If a customer wants or needs a person, the chat passes to a care agent, which Chewy says happens within seconds. Singh told the Goldman Sachs retail conference on 14 September that Cai and Callie, a voice agent that confirms and schedules appointments at some of its vet clinics, are built in-house rather than bought.

Alongside Cai, Chewy has put AI tools in front of its own care agents to pull information together and cut the manual work on each contact, and uses AI in its pharmacy to automate data extraction and checks.

Chewy did not start with a chatbot. It started with returns, the most repetitive request it gets, and added the assistant months later.

How long did it take?

The order is the useful part. On the 25 March call, Singh said Chewy had launched self-service returns and refunds about eight weeks earlier, which puts it around late January 2026, and that tools for agents were in development to cut handle time and contact rates.

By 10 June it was repeating the target without new detail. Cai was originally due in the third quarter. Singh said Chewy brought it forward into the second quarter after lowering its sales guidance, to protect profit. When he gave the 30% figure on 9 September, Cai had been live for less than a month and fewer than 10% to 15% of app traffic could see it.

So the sequence ran over roughly eight months: one self-service flow, then agent tools, then a customer-facing assistant in a limited beta, with voice tested separately on appointment booking.

What does the 30% actually measure?

It measures chats resolved through self-service among customers in a beta, over a few weeks. Chewy has not said how it defines resolved, whether those customers contacted it again about the same order, or what share of all its contacts arrive by chat rather than phone or email. The figure is not 30% of Chewy's customer-service volume.

The customers in an early beta may also have simpler questions than the average. Order status and Autoship changes are the easiest requests to automate, and Singh named them as the first ones Cai covers. The rate is likely to move, in either direction, as the assistant reaches every app user and harder requests.

What does $50 million mean against Chewy's size?

It is small. Chewy's guidance for fiscal 2026 is $13.46 billion to $13.57 billion in sales and about $912 million in adjusted EBITDA at the midpoint. By our arithmetic, $50 million a year is about 0.4% of sales and about 5% of adjusted EBITDA. The savings figure also covers all of Chewy's AI work, including pharmacy and fulfilment, not customer service alone.

Singh was careful about where the money goes. On the September call he said investors should not add the savings on top of Chewy's margin plan, because they will also offset normal cost pressures such as wage inflation, and some may be reinvested. He called AI a way to keep margins rising, not a separate pool of profit.

Chewy has not said what Cai cost to build, what it pays for the underlying models, or whether care headcount has changed. Its last annual report put total staff at about 18,000.

Does this apply to a 200-person retailer?

Partly. The sequence applies, the build does not. Chewy has its own engineering team and data platform and built Cai itself. A mid-sized retailer will buy this from its help-desk or contact-centre vendor. What carries over is the order Chewy followed: automate the single most common request first, measure it, then widen.

What it does not mean is that a chatbot will take 30% off a support bill. Chewy's figure is a share of chats from a selected group of customers, and its own chief executive frames the savings as an offset to rising costs.

What to do next

Pull your last three months of contacts and rank them by reason. If order status and returns are near the top, as order status is for Chewy, those are the ones to automate first, ideally from data your order system already holds.

Before you switch anything on, decide how you will count a resolution. A chat that ends without a person but brings the customer back by phone the next day is not one. Track repeat contacts on the same order for at least two weeks.

Keep the hand-off to a person fast and visible. Chewy made it a design rule, and it is the part customers will judge you on.