Two automation stories, one industry
Cyber Week 2025 closed with retailers reporting the two headline numbers they wanted: record spending and a visible role for AI agents in generating it. Global sales across the five-day period from Thanksgiving through Cyber Monday hit $336.6 billion, up 7% year over year, with AI agents and assistants influencing roughly 20% of purchases and directly driving about $67 billion in sales, according to Salesforce (Salesforce, December 5, 2025). Retailers running their own branded agents saw about 13% U.S. sales growth over the prior seven weeks, versus 2% for those without one, a sevenfold gap Salesforce attributes to the agent itself rather than general holiday demand.
Adobe's numbers point the same direction. Cyber Monday alone brought in $14.25 billion, a new single-day record, up 7.1% year over year, and traffic to U.S. retail sites from generative AI chat tools and browsers rose 670% compared with 2024 (Adobe, December 1, 2025). Buy now, pay later usage hit an all-time high the same day, at just over $1 billion.
Set those numbers next to a second, quieter 2025 story: retailers kept removing self-checkout machines. Dollar General pulled self-checkout from roughly 12,000 stores, the majority of its footprint, citing shrink (Payments Dive, 2024). Five Below moved toward a goal of 75% of transactions handled by an associate chainwide and 100% at its highest-shrink locations (Retail Dive, March 21, 2024). Target capped its Express lanes at 10 items in 2024 and was still promoting the limit as company policy through 2025 (Target Corporate, May 2025). None of this reversed course in 2025. A LendingTree survey of 2,050 U.S. adults conducted in October 2025 and published in December found 69% of self-checkout users believe the format makes theft easier, and 27% admitted deliberately walking out with an unscanned item, up from 15% in a 2023 reading of the same question (LendingTree, December 2, 2025).
The same industry, opposite bets
That is a strange split for one sector to run at once. Retailers spent the year pouring engineering effort into letting software act on a customer's behalf earlier in the funnel: recommending products, answering questions, initiating returns, updating a delivery address. Salesforce reported agentic customer service conversations up 55% week over week during Cyber Week, with agent-handled actions like returns and address changes up 70% versus the prior week. At the same time, those same retailers kept walking back the one automation layer that touches money directly at the point of sale.
The two decisions are not contradictory once the risk each one carries is separated out. An AI agent that recommends a product or drafts a return request is making a suggestion inside a system that still checks itself: a retailer can reverse a bad recommendation, decline a fraudulent return, or flag an odd request before money moves. Self-checkout removes a human check at the one moment the transaction actually completes. The theft data reflects that difference in exposure, not a difference in how advanced the technology is. Scanning a barcode is a solved problem. Verifying that the right barcode was scanned, without a person watching, is the part retailers have not solved, and the shrink numbers behind Dollar General's and Five Below's decisions say they know it.
What this means going into 2026
The industry is not choosing between automation and staffing. It is choosing where to place automation based on how easily a mistake can be caught downstream. That is a more useful way to read next year's retail tech announcements than treating every "AI-powered" claim as equivalent. An agent that recommends or assists carries a different risk profile than an agent, or a machine, that finalizes a transaction unsupervised.
The open question for 2026 is what happens as agents move further into that second category. Salesforce's own numbers show agents already initiating returns and changing delivery details, tasks that touch money and inventory more directly than a product recommendation does. If theft and fraud losses show up in that layer the way they did at the self-checkout register, expect the same retreat, just a year or two later and aimed at a different piece of software.