The weaving of machine intelligence into familiar workflows, transforming standalone AI tools into extensions of the actor’s existing ways of thinking and acting.
Understanding intent
Translating user’s fuzzy, high-level goal into a precise, structured command the system can execute. This process involves two key steps:
Intent Detection
The system identifies signals that indicate a user’s goal. These signals can be:
- Keywords and natural language: Direct text input.
- User actions: Selecting text, dragging an element, hovering over an item for a certain duration.
- Navigation paths: The sequence of pages or views a user visits can imply a larger goal.
- Contextual cues: The current state of the application, the type of data being viewed, or the user’s role.
- User history
Intent interpretation
…
To-do
- Continuous feedback loop — the actor’s reaction to the embedded intelligence (accepting or rejecting a suggestion) provides feedback for improving the underlying model.
Resources & references
- Heer, J. (2019). Agency plus automation: Designing artificial intelligence into interactive systems. PNAS.
- Salovaara, A. (2004). Six modes of proactive resource management.
- Noessel, C. (2017). Designing agentive technology.
- Tennenhouse, D. (2000). Proactive computing.
- Nardi, B., Miller, J. R., & Wright, D. J. (1998). Collaborative, programmable intelligent agents. Communications of the ACM, 41(3), 96–104.
Consequences
- intelligence acts inside the existing workflow on context the system already holds, so the actor never restates what they're doing in order to get help with it
- the system is now acting on inferred intent rather than stated intent, which means it will sometimes be confidently wrong about what the actor wanted
- because the help is woven in, which part of the result was the actor's and which the system's stops being obvious — and that boundary only exists if something shows it
Related patterns
Precedes
- Filtering — natural language queries translated into structured filters
- Focus and context — infers what contextual information is most relevant to the actor's current focus
- Suggestion — the output is often a suggestion the actor can accept, reject, or modify
Enacts
- Adaptability — the system mixes rule-driven and inferred responses, so behaviour shifts with context rather than staying fixed
- Agency — the system decides and acts on its own behalf rather than only responding to the actor's explicit moves
Complements
- Command menu — a natural home for embedded intelligence, turning a command line into an intent-driven interface
- Next-best action — next-best action as intelligence woven into the workflow
Related
- Help — The proactive sibling to Help. If Help explains how to do X, Embedded Intelligence offers to do X for you.