Shared and stakes

Not everything AI touches needs the same care. A first draft of your own email needs very little. A tool that screens job applicants needs a great deal. Two questions sort one from the other, and they're the same two questions whether someone has built something or is simply using AI in their work.

Two questions, for things and for uses

Most AI governance starts from what gets built: a new tool, an app, an agent. That misses a lot of the risk, because a lot of the risk sits in how people use tools they already have. A recruiter ranking CVs builds nothing, but both questions still apply. The same goes for a lawyer drafting a court filing, or a finance lead pulling numbers together for the board. Nothing new exists, and the stakes can be as high as anything you'll build.

Is it shared?

This question is about tools, workflows and standing practices, not one-off outputs. The single email you drafted isn't shared. The template your whole team drafts from is. Something is shared if any of these is true:

• someone else depends on it
• it acts beyond your own work
• it writes into a shared system or a system of record
• outsiders use it directly, with no person in between

Anything shared needs an owner. Giving it one is module 7's job, but the question gets asked here, at the front door, so nothing shared slips through without one.

What's at stake?

There are two levels, low and high. Something is high stakes if any of these is true:

• it helps decide something about a person
• it touches regulated or confidential data
• it takes action outside your own work without a person checking
• it makes statements or commitments to outsiders on the organisation's behalf
• it could harm someone who never uses it

An organisation can add a middle band in its stakes rules, but it has to write down exactly what goes in it. An undefined middle is where hard cases go to avoid a decision.

What each combination needs

Put the two answers together and you get four combinations, each with its own minimum.

• Personal and low stakes: approved tools and the output standard. That covers most everyday use.
• Shared and low stakes: an owner, with a backup who can stop it, monitoring, an off switch, a tested way to undo it and a review date.
• Shared and high stakes: Build's full proof, with evidence from someone who didn't build it, then full ownership.
• Personal and high stakes: a tool that has passed Build's proof, plus a named person who checks the output.

For both high-stakes combinations, the people affected are told that AI is involved, and the records the stakes require are kept. Try it below with the examples, then with your own uses. Each one you add goes into this module's output.

The named check

Personal high-stakes use is where AI governance most often fails, because nothing was built that anyone could review. The framework's answer is a named person who checks the output and has the time, information and authority to reject it. Each word does work.

• Time: enough to check properly, at the volume the work actually arrives.
• Information: the sources and context needed to judge the output, not just the output.
• Authority: the standing to say no, without asking permission or worrying about what happens next.

Take any one of them away and the check becomes a rubber stamp. The name is still on it. The checking isn't.

When the check is a stamp

In 2023, a federal court in New York sanctioned two lawyers and their firm after they filed papers citing six court decisions that didn't exist. An AI chatbot had invented them, complete with quotes. The lawyer who signed the filing was the named check. Before signing, the judge found, he did not review any of the cases it cited. The case was Mata v. Avianca, and the judge was careful about where the fault lay.

The same opinion says there is “nothing inherently improper” about using a reliable AI tool for help. The tool wasn't the failure. The check was. That's why the practices layer spells out what a check means for each kind of use, so that “I looked at it” isn't mistaken for checking.

Outsiders hold you to what it says

Shared, high-stakes things fail in public. In February 2024, a Canadian tribunal ruled on an airline's website chatbot, which had told a grieving customer they could buy a full-price ticket and claim the bereavement fare afterwards. The airline's own policy said they couldn't. In its defence, the airline suggested the chatbot was a separate legal entity, responsible for its own actions.

The tribunal called that “a remarkable submission”. It held that the airline is responsible for all the information on its website, whether it comes from a static page or a chatbot, and ordered it to pay. The sum was small. The principle travels: whatever speaks for you, you answer for.

In the framework's terms, that chatbot was shared, because outsiders used it directly, and high stakes, because it made statements on the organisation's behalf. Nobody can check every answer a chatbot gives. That's why every output has to trace back to someone who answers for it, either directly or as the owner of the system that produced it.