You own what it says

In February 2024 Air Canada told a tribunal that its chatbot was a separate legal entity, answerable for its own words. The tribunal member called that a remarkable submission and ordered the airline to pay. Whatever the software had said, the company had said.

That is the whole brief. An agent is not a contractor you can point at afterwards. It is a part of your company that talks to your customers, writes into your records and commits you to things. There is nobody behind it to take the blame.

What is being sold

The industry sells autonomy. Digital workers that find your tools by themselves, make their own plans and take the work off your hands. It is a good pitch, and the public record of it is short and bad.

New York City spent roughly $600,000 on a chatbot that told employers they could keep their staff’s tips and landlords they could turn away tenants holding housing vouchers. Both are illegal. The city decommissioned it in January 2026 as functionally unusable. McDonald’s ended a multi-year voice ordering pilot across more than a hundred drive-thrus in June 2024, beaten by engine noise and the way people actually speak. The National Eating Disorders Association replaced its helpline staff with a bot that started handing out calorie advice to people with eating disorders, and took it down within days. A courier’s support bot was talked into swearing at a customer and writing a poem about how bad its employer was.

These are not teething problems. In every one of them a probabilistic system was handed authority over something with consequences, and used it.

What the people shipping this actually do

They take capability away from the model. Anthropic’s own engineering guidance tells builders to find the simplest thing that works and add complexity only when something forces them to. GoDaddy found its breakthrough in breaking one enormous prompt into small ones with ordinary code in between. Charity Majors of Honeycomb puts the arithmetic plainly: if an answer can be computed, it probably should be. Hamel Husain, who has sat inside a lot of these projects, says the common root cause of the ones that fail is having no way to tell whether the thing is working.

None of that is a compromise anyone made reluctantly. It is the shape of the systems that survived.

Where we stand

The model reads. The machine decides. A person signs.

A language model is very good at one job: turning something unstructured into something typed. A photograph of a handwritten order, a voice note, a forty-page specification, a thread across four channels. Let it do that. Everything after belongs somewhere older and duller. Control flow in compiled code. State in a database with constraints. Prices and rules in a rules engine that gives the same answer twice. Anything consequential waits for a named person, with the arithmetic already done. And every run is on the record, reconstructable months later, line by line.

This is slower to build than a demo and it is the only version that is still running in a year.

It has already started

We are not forecasting. C.H. Robinson has agents running more than three million shipment tasks inside its own platform. ADI Global stopped keying seventeen hundred orders a day and got back a hundred and forty-two hours of its sales desk’s time daily. Leonardo Hotels closed two hundred and sixty-one thousand guest conversations in a year without a person touching them. Every one of those is named, dated and sourced in the playbooks on this site, with a label saying who did the measuring and whether they were selling something.

It does not arrive across a company. It arrives on one desk, in one process, and then the next one.

What would make us wrong

Two things, and we are watching both.

If models begin to prove their own output correct at the moment of generating it, the validation harness we build around them becomes dead weight, and we should stop building it. And if the law moves, if a legislature decides a company is not answerable for what its agent tells a customer, then the careful approval gates we insist on become a cost with no matching risk.

Neither has happened. Until one does, we will keep building the dull, checkable version, and we will keep publishing the evidence so you can argue with it.