PRAXIS glossary

The terms of the framework

PRAXIS names decisions that product design has been making for years without naming them. These are its core concepts: what they mean, when they apply, and how they show up in a real product.

← Back to the PRAXIS framework
01Delegation threshold

The boundary past which a system stops doing things for you and starts doing them instead of you. It is not a technical line that can be measured: it is a change of state, like the point where water turns to vapour. Crossing it changes the nature of the relationship between the person and the product, and it is almost always crossed without anyone deciding to.

Context

Most teams delegate by capability: if the machine can do it, it does. The threshold inverts the logic. The question is not whether the system can make the decision, but whether it should, and at what point delegating it leaves the person worse off.

Example

A clinical system can triage a waiting list more accurately than a human. The threshold is crossed when that prioritisation stops being a recommendation a professional reviews and becomes a decision no one questions because no one understands it.

02Dignified errors

A dignified error is one the person can understand, correct and forgive: visible, intelligible and reversible. It stands against the error that betrays, which is opaque, irreversible and silent. Designing for dignified errors means deciding in advance which failures are tolerable and which must be impossible by architecture.

Context

Every system fails. The design question is not how to prevent all error, but which kind of error the product is willing to accept, and whether those errors leave the person room to react or leave them with no way out.

Example

An assistant suggesting a wrong answer the user corrects is a dignified error. Executing an irreversible transfer on that same mistake is not. The difference is not in the model; it is in what it was allowed to decide on its own.

03Product character

A product's character is not what the team says it is: it is what the product does when no one specified it. It is the sum of its reactions in ambiguous moments, the way it insists or yields, how it treats whoever does not fit the expected pattern. A product that acts with initiative has character, whether it was designed or not.

Context

When a system acts on its own, it stops being a neutral tool. Every default, everything it does without asking, every limit it respects or ignores builds a personality the user perceives. Character is not a brand layer: it is the sum of behavioural decisions.

Example

Two assistants on the same underlying model can have opposite characters: one that interrupts and assumes, another that waits and asks. Technology does not define that difference; whoever designed where the system takes initiative does.

04Behavioural design for AI

The discipline that designs how an intelligent system acts, not just what it shows. It covers its initiative, its delegation boundaries, its conduct under uncertainty, and the preservation of the person’s agency. It is the discipline missing between user experience and AI engineering.

Context

Traditional design deals with interfaces: what the user sees and touches. When the system decides and acts on its own, the object of design stops being the screen and becomes the behaviour. Today that behaviour is designed by omission, in the code, without anyone treating it as a design decision.

Example

Deciding whether an agent confirms before sending an email, acts only within set limits, or never touches irreversible actions is not an engineering decision. It is behavioural design, and it determines whether the product can be trusted.

05Ethics as architecture

The principle that an ethical limit is not stated by saying 'this must not happen', but by designing the product so that, by its very nature, it cannot do what would cause harm. It works like the foundations of a building: red lines, irreversible harms and human-reserved decisions implemented as structure, not as a compliance layer bolted on at the end.

Context

Ethics treated as compliance arrives late and can be skipped. Treated as architecture, it defines what is impossible for the system to do, regardless of business pressure or what the user asks. The difference is the one between a rule and a load-bearing wall.

Example

A policy saying the system should not make a certain decision is compliance. The system being unable to make it because that decision is reserved by design to a person is architecture. Only the second holds when there are incentives to skip it.

06Human agency in the loop

The real, not nominal, capacity of a person to understand, question and reverse what a system decides. Having a human in the process is not enough: that human must have the information, the time and the effective power to intervene.

Context

'Human in the loop' has been emptied of meaning. In many systems the human is present but powerless: approving what they do not understand, under time pressure, with no real ability to reverse it. Preserving agency means designing so that human intervention is genuinely possible, not a symbolic signature.

Example

An operator who must approve a hundred system decisions an hour is not exercising agency: they are legitimising an automation. Agency is preserved when the design reduces volume, explains the why, and reserves time for judgement.