Shoppers sent by AI assistants buy 42% more often, and claims adjuster jobs fell 21%

More than half of the shoppers who arrive at a Shopify store from an AI assistant land straight on a product page. From Google search, about one in five do.

Adobe Analytics, which tracks visits across US retail sites, put the change inside one year. In March 2025, visitors sent by ChatGPT, Gemini and similar assistants were 38% less likely to buy than other traffic. In March 2026 they were 42% more likely.

Those shoppers did their comparing in the chat window, before any store saw them. The assistant picks which stores to name by reading product descriptions, prices and availability, so three vague lines of description, or a stock count that was last correct on Friday, is what it has to work from.

That is the work the big chains described on last week's earnings calls. On Ulta's call, Steelman said the beauty chain is creating content and enriching its product information for AI platforms including ChatGPT, and pointing shoppers to its own assistant, Ulta AI. Kohl's chief executive Michael Bender said on August 26 that its AI shopping assistant is producing more purchases per visitor and more revenue per visit.

Xin Qin and seven colleagues pooled 163 experiments covering 82,078 people to settle a disagreement in their field. Some studies find that customers prefer a person; others find they prefer the machine. Writing in Psychological Bulletin, the group reports that two conditions decide the answer: customers accept AI when they believe it does that particular task better than a person would, and when the task needs nothing personal from whoever performs it. Where either condition fails they resist, and the preference for a human in those cases is about twice as strong as the preference for AI when both conditions hold.

Checking a policy limit or tracking an order meets both conditions, because the customer wants the fact and does not care who supplies it. A disputed bill fails the second one however accurate the machine is, because the customer wants somebody to answer for the decision.

Bad news is where the pattern surprises people. Because customers credit a machine with no motive of its own, a disappointing offer from one feels less personal. Aaron Garvey, TaeWoo Kim and Adam Duhachek found in the Journal of Marketing that people were more likely to buy when an AI delivered an offer worse than they expected, and that a human agent drew the warmer response when the offer was better than expected.

No occupation Glassdoor tracks is more negative about AI than claims adjusting. Of the comments adjusters left about AI in the year to May 2026, 98% were critical, against 53% across all jobs.

Their employment fell about 21% over that same year as insurers automated claims work. Across insurance carriers and related businesses, employment fell 2.5%.

Most of that decline shows up as openings that were never posted. Advertised jobs for claims adjusters are down about 55% from their post-pandemic peak, against about 36% for the labor market as a whole. A role can lose a fifth of its people with no announced layoff at all: an adjuster retires, a contract is not renewed, and the vacancy never comes back.

Among the negative reviews that named a specific criticism, 20% raised AI replacing workers, 14% said their employer forced them to use the tools, and 13% said the company was pursuing AI at the expense of its core business and its customers. Reviewers who wrote about AI negatively were six times as likely to mention layoffs and 3.7 times as likely to mention job insecurity or burnout. Glassdoor also found workers more positive where the employer supplied the tools, the training and the support to use them.

The Bureau of Labor Statistics expects the wider group of claims adjusters, appraisers, examiners and investigators to shrink 6% between 2025 and 2035. It also expects about 21,600 openings a year over that decade, as people retire or leave the work.

OpenAI has begun letting some large customers pay only after its AI finishes a task, joining Salesforce, Sierra and Cognition in charging for completed work rather than for access. The attempts that fail along the way are the vendor's cost now, which turns the definition of finished into a line in the contract, written by two parties with opposite interests in the wording.

Anthropic is making the same adjustment to subscribers who negotiate nothing. It will raise Claude Code's weekly usage limits permanently by 25% on September 14, and remove a temporary 50% boost on the same day. Both companies are pulling back the subsidized pricing that made AI coding tools cheap. From September 14, the same subscription buys about 17% less coding capacity than it does today.

A benchmark report published this week found that only 13% of companies get full use of the HR software they have already paid for, and it finds the same pattern repeating with AI tools in every department, where the buying comes before the training.

IBM's newest survey of chief executives found 76% of companies with a Chief AI Officer, against 26% in last year's edition. Demand for the title is cooling in hiring data, and the CIOs in the same survey say the job changing most is their own.