A $20 ChatGPT agent needs the whole inbox, and C.H. Robinson's staff is down a third

OpenAI's ChatGPT Work turns the chatbot into an agent that works through office tasks on its own, and it costs $20 a month. It connects to a person's Gmail, Google Calendar, Slack and GitHub through one plugin per app, and it runs on GPT-5.6. VentureBeat reported that the rollout started with Pro, Enterprise and Edu accounts before reaching Plus and Business.

To work an inbox, the agent has to be able to read the inbox, not the one thread it was pointed at. In a consultancy, that is every client attachment anyone has sent over the past five years, plus the Slack channels where the awkward cases get discussed.

What it does with that access is the ordinary chain of a coordinator's morning: find the thread, read the attachment, draft the reply, put the meeting on the calendar. It has no way to weigh whether the document it just pasted into a summary was the one covered by a confidentiality clause.

The $20 is a flat fee, and an agent working through a multistep task uses far more computing than answering a question in chat, so a subscriber who runs it all day pays the same as one who opens it twice a month. For now, OpenAI covers the difference between the $20 and what that work costs to run.

C.H. Robinson had about 14,990 employees at the start of 2024. By the end of 2025 it had about 12,085, and since 2022 its employment is down 31%, the Star Tribune reported in March.

Most of what those employees did arrived as email. The broker receives more than 11,000 messages a day from customers and carriers asking what a truckload will cost, and its quoting agent now returns a customer-specific price in 32 seconds. A second agent reads emailed load tenders, attachments included, and builds the order in about 90 seconds, 5,500 truckload orders a day, which the company puts at 600 hours of labor saved daily. A person used to spend an average of seven minutes keying in the details of one load, and a tender covering 20 loads meant doing that 20 times.

These are not tidy forms. One tender says only that there is a load for Tuesday, because the shipper knows the broker knows what it ships on Tuesdays; the next is a PDF about 20 loads with handwritten notes on it. Mark Albrecht, the company's vice president for AI, has described the agent working out what is missing from an email and going to fill in the blanks itself.

Company executives have tied the reductions to productivity gains rather than to falling freight volumes. C.H. Robinson's own count puts productivity up more than 60% since the end of 2022, with average headcount down 10.8% year over year in the second quarter of 2026. In March it offered voluntary severance to a limited group of senior managers.

Travelers took in $49 billion of revenue last year and employs 30,000 people. The model it uses for routine insurance questions is one the company built itself.

Travelers unveiled TravelersLLM in June, trained on its own curated insurance data. Mojgan Lefebvre, the insurer's chief technology and operations officer, says it is "meaningfully faster and less expensive than frontier models" on certain insurance tasks.

A narrow model can manage that because of what the queue looks like: the same questions, thousands of times over, in wording that barely changes, about policies and forms the company wrote. A frontier model charges for reasoning that work never calls on, and companies increasingly route each task automatically to the cheapest model that can handle it.

CIO Dive reported this week that Travelers sends insurance-specific questions to its own model and keeps broad reasoning, research and coding on the costlier frontier models.

Alibaba's Wan3.0 takes a PDF, a slide deck, a spreadsheet or a web page and returns a video of up to 30 seconds. On Alibaba Cloud Model Studio the price is by the second: 5 cents at 480p, 10 cents at 720p, 20 cents at 1080p. A 30-second clip costs $3 at 720p and $6 at 1080p.

That puts a product explainer or a training video inside the budget of a firm that has never paid for one. The Next Web noted that no outside group has benchmarked the results, so the only published comparison is price: Google's Veo 3.1 standard tier is 40 cents a second, four times the 720p rate.