AI in retail and customer service: e-commerce and CX benchmarks (with sources)

Retail is where AI shows up most to the end customer, and where a mistake becomes viral or a lawsuit on the same day. The sales and service gains are real; so are the disasters. This is the picture of both sides, sourced.

Adoption is broad. Per the Stanford AI Index, 78% of companies used AI in 2024, and generative-AI use in at least one function jumped from 33% to 71%. In retail specifically, McKinsey reports 82% have run generative-AI pilots in customer service and 36% are already scaling, in an estimated potential of $400 to $660 billion a year.

Where AI already delivers in retail

Application	Result (sourced)	Company
Integrated optimization (inventory, price, recommendation)	$42M saved, $110M in sales and $13M in profit per year	Alibaba (INFORMS study)
Customer service	2.3M chats in 30 days, two-thirds automated, resolution < 2 min	Klarna
Service deflection (forecast)	Conversational AI to cut $80B in call-center costs by 2026	Gartner
Personalization	Typical revenue lift of 10 to 15%	McKinsey
Virtual try-on	Online sales growth correlated with rollout (Sephora case)	Harvard d3

The most solid case is Alibaba, published in a peer-reviewed journal: a single system for forecasting, inventory and pricing generated tens of millions in annual savings and sales. Klarna, meanwhile, is the perfect example of both sides, because the same success case became, months later, a reversal case (see below).

Where AI broke in retail

Air Canada: a tribunal ordered the company to compensate for a fare policy its chatbot invented (2024) and rejected the argument that the bot was a "separate entity".
DPD (UK): after an update, the chatbot swore and wrote a poem calling the company "the worst"; the video passed 1.3 million views and the AI feature was disabled.
Chevrolet of Watsonville: a prompt injection made the bot "agree" to sell a ~$76,000 Tahoe for $1.
McDonald's: ended its drive-thru AI pilot at 100+ stores (2024) after viral videos of wrong orders.
Klarna: walked back its "all-AI" stance in 2025 and rehired humans after a drop in satisfaction and "generic" replies.

Regulation: you answer for what the bot says

The Air Canada case became the global reference: a company answers for its AI's statements just as it does for any content on its site. In Brazil, bill PL 2338/2023, approved by the Senate in December 2024, aligns with the LGPD and the Consumer Defense Code, bans manipulative techniques (dark patterns), and sets fines up to R$50 million or 2% of revenue per infraction.

In Brazil

Lu do Magalu, with more than 32 million followers, got Google generative AI in 2023 to recommend products and resolve customer issues. Mercado Livre weaves AI across its commerce, payments and logistics ecosystem, though without a clean public revenue-lift figure, which reinforces the point: integrated data advantage is worth more than the model alone.

The lesson for anyone implementing

Automate the repetitive volume, not the accountability. The cases that work keep a human in the loop and clear limits; the viral disasters come from letting the bot loose with no guardrails. Gartner projects agentic AI will resolve 80% of common service issues by 2029, but those who get there are the ones who built with governance. That is Reche's work: putting AI into customer service and e-commerce with process, testing and oversight.