Understanding AI work through norms and intents

When we evaluate AI work, one of the harder problems is understanding what an action means in the organization where it happened. Knowing that an agent read a file, ran a command, or changed code tells us something about its behavior, but we still need to understand the expectations it operated under and the purpose the work served. The same action can be appropriate in one context and concerning in another.

At Origin, we use three points of reference to reason about this:

• Written policies describe what the organization explicitly requires.
• Norms describe how work is usually done.
• Intents describe recurring purposes we infer from work across the organization.

We keep these separate because learning how an organization works doesn't establish what it should accept. An agent can depart from an established practice for a good reason, while work that looks entirely routine can still violate a written rule. Those differences matter when deciding what deserves investigation and what needs to change.

We encounter a similar distinction when we learn what's expected in an unfamiliar setting. A visitor to our house can see us taking off our shoes and putting them on the rack just inside the door, and pick up the practice without anyone explaining it. But suppose they arrive with muddy boots and decide to leave them on the porch so they don't carry the mud inside.

They've understood the expectation and adjusted how they follow it to serve its purpose. If we compared their choice only with where everyone else left their shoes, we'd identify a difference while missing the reason it made sense.

When reviewing AI work, we ask whether an agent followed a relevant practice, departed from it, or did something the learned norms don't address. To call something a departure, we need to identify the practice and show what the agent did differently, then consider whether that difference introduced risk, wasted effort, or interfered with other work. A harmless difference doesn't independently give us a reason to escalate it.

For intents, we ask whether the work served an organizational objective, was tangential to it, or worked against it. A necessary prerequisite can serve an objective before the main task is complete, while unrelated work is tangential unless there is evidence that it actively works against that objective. There are also activities for which the learned intents don't provide a useful basis for judgment. Even when work is aligned, it can still introduce risk or violate a written rule.

Because we learn norms and intents from captured AI work, they describe the part of the organization we've observed. We look for recurring patterns across different people's work, with norms describing shared practices and learned intents describing purposes that persist across pieces of work. Those patterns still need local context: a practice that is common in finance may tell us little about what to expect in development.

A widespread practice can be harmful or outdated, and a recurring purpose inferred from work is not automatically an objective the organization has endorsed. Written policy carries authority that an observed pattern doesn't. A judgment against a policy, norm, or intent needs to identify the particular expectation it concerns and be supported by evidence in the captured work.

Keeping these sources of expectations separate helps us identify gaps between policy and practice. If work repeatedly follows an established norm while violating a written rule, that may suggest the rule has become a dead letter. It could also reveal a persistent practice that needs correcting. One instance wouldn't establish either conclusion, and repetition alone doesn't tell us which should change. Conversely, a reliable practice that policy doesn't address might deserve to be formalized. Those decisions remain with people.

This context is part of how we think about detection at Origin. We evaluate risky actions, sensitive material, and threats such as prompt injection alongside how work relates to the organization's policies, norms, and intents.

The goal is to help a reviewer understand what happened, identify which expectations apply, and decide whether the evidence calls for investigating an agent's action, correcting an established practice, or reconsidering a written rule.