From notification through suggestion to automatic execution is a continuum
Aliases: proactivity spectrum · mixed-initiative interaction
What it is
A system's proactivity is not an on/off switch but a spectrum: answer only when asked → proactively signal "something is up" → offer a suggestion and wait → suggestion with one-tap execution → execute automatically and report afterwards → fully automatic and imperceptible. Each step to the right has the system doing more and the user intervening later. Research on mixed-initiative interaction treats this spectrum as a primary design degree of freedom: where a function sits on the spectrum shapes the experience more than how "intelligent" it is.
The common design failure is binary placement: either fully manual or fully automatic. The spectrum view says the middle ground is large and often optimal — "execute automatically but keep undo" and "suggest, but only at low-cost moments" are legitimate positions.
Why it happens
Why a continuum rather than two settings? Because two costs move monotonically in opposite directions along it:
- Interruption cost falls as the level rises — a low level calls the human back to judge, and every call is an attention shift.
- Error cost rises with the level — the higher the level, the more directly an inference error becomes an executed action, and the later the human gate stands.
Choosing a level is finding the minimum of the sum of these two curves for the task at hand. This also explains why the same function may shift position for the same user over time: early on, error cost dominates (stay low); once the system is trusted, interruption cost dominates (move up).
A second role of the spectrum is responsibility allocation. The level simultaneously defines who is at fault when things go wrong: at the notification level a miss is the user's; at the automatic level the system erred on the user's behalf. Full automation moves responsibility wholesale onto the system — which is why it demands a steep increase in trust.
Studying it
- Horvitz's principles for mixed-initiative interfaces formalise the decision: intervene only when the expected value of the action exceeds the value of the user's attention it interrupts — both the timing and the magnitude of initiative become variables in an expected-utility comparison.
- Norman, in The Design of Future Things, argues the negative case: what makes automatic machines fail is the missing feedback channel, leaving people to accept "inexplicably clever" behaviour. His prescription effectively pins functions to the low end (augment and inform) until a system can explain itself.
- A standard design is the online comparison: build the same function at two levels (suggest vs. execute automatically), compare task completion time, realised error consequences, and trust scales, and watch how users migrate between levels over time.
One methodological caution: the effect of a level is conditioned on the user's prior impression of the system. The same suggestion-level function behaves differently for a user who has been burned by it and one who has not; when comparing level preferences across users, usage history must be a stratifying factor, or the conclusions are confounded.
Where it stops holding
- The spectrum presupposes controllable consequences. Irreversible, spillover actions (send, unlock, spend) have no safe high position on the spectrum; any automatic execution needs separate gates — that is the subject of the whole group on boundaries of automatic execution, not something the level choice alone resolves.
- "Fully automatic and imperceptible" holds only for highly reliable functions with minor consequences; treating it as the universal endpoint of every feature is vision narrative.
- More adjustability is not automatically better — turning level choice into a daily chore transfers the system's uncertainty onto the user.
Applying it
- Explicitly level every proactive feature: during product definition, answer "which cell of the spectrum is this on", rather than letting implementation details decide by default.
- Explore the middle first: suggestion with one-tap execution, and automatic execution with after-the-fact undo — the two positions that best trade off interruption against error cost, and the two most often skipped.
- Allow one-directional, gradual promotion with usage history: new users start low; after a clean record, offer to move up. Moving up must be explicit ("shall I start doing this automatically?"); moving down is available at any time.
- How to check: for each feature track two rates — how often users voluntarily lower the level, and how often they disable the feature within 24 hours of a misfire. Both low means the placement is right.