C.H. Robinson runs 30 AI agents. Here's the part they skip.

What: C.H. Robinson reports 30+ AI agents that have executed 3 million-plus tasks — billing, document handling, pricing, scheduling, carrier vetting — claiming roughly 40% productivity gains (disclosed Q1 2026). Why it matters: when the largest 3PL normalizes agent-run back-office work, every TMS and brokerage vendor downstream resets what they pitch you. Do: ask your vendor which tasks they’ve actually automated — get the task list, not the percentage. [Source: C.H. Robinson.]

What: Bot Auto ran a fully humanless 231-mile commercial truckload across Texas on April 29 — no safety driver, no observer — and Aurora’s Q1 8-K reports 250,000-plus driverless miles with the Fort Worth–El Paso lane added. Why it matters: “no safety driver” is the line that changes the cost model; adoption is now milestones, not promises. Do: if you run southern lanes, price these corridors into 18-month capacity planning, not your five-year deck. [Source: FreightWaves, Aurora 8-K.]

What: Current API rates put the cheap tier at Claude Haiku 4.5 ($1/$5 per million tokens in/out), Gemini 2.5 Flash ($0.30/$2.50), and GPT-5.4-nano ($0.20/$1.25) — and a cache hit cuts input cost ~90%. Why it matters: the LLM brains of an AI dispatch tool now cost cents per conversation, which means the price on the invoice is markup and data feeds, not compute. Do: when a vendor quotes per-truck pricing, ask for the split — token cost vs. data feeds vs. margin. [Source: Anthropic pricing.]

EIA Weekly Diesel Survey, week of May 18, 2026.

The Call. Diesel printed $5.596/gal — down 4.3¢ week-over-week, but still well above year-ago levels. For carriers weighing whether to add an AI quoting tool: the rate environment, not the software, still sets your floor. This read tells you where diesel is — not where it’s going.

Indicative figure. Public source (EIA Weekly Diesel). Not a substitute for paid intelligence.

Strip the marketing off an “AI dispatcher” and you find a loop. A driver or dispatcher asks a question in plain English. An LLM reads it, decides which tools to call — route, fuel, rates, weather — fires them, reads what comes back, and writes the answer. Run it again if more is needed. That’s the agent. The model is the orchestrator and the writer. It is not the calculator.

That distinction shows up in the bill. The model layer is cheap now — with prompt caching, a full multi-tool conversation runs in cents. Routing is nearly free too. What you actually pay for is the gated freight-rate data (DAT, Truckstop — contract-priced, no public rate card) and the vendor’s markup. When a tool is sold as “AI,” the AI is the cheapest part.

And the failure mode that matters most: the second you let the model estimate instead of forcing a tool call, it hallucinates — confident, fluent, wrong miles and wrong rates. The most important safeguard in a well-built agent is a hard rule — never estimate miles, call the routing tool. A model narrates bad tool output as smoothly as good: garbage in, eloquent garbage out.

So: buy or build one to save your dispatchers five lookups a load. Don’t buy one to replace their judgment — that’s not what’s in the box.

Open at FMCSA this week — items worth tracking. Dates per the source notice.