Your AI Assistant Doesn't Know When to Leave You Alone

The interruption problem nobody is measuring properly

Most AI assistant demos are built around a flattering assumption: if the system can offer help, you will want it. A suggestion appears, a summary is generated, a reminder arrives, and somewhere in a product meeting a graph moves reassuringly upwards.

But attention is not an empty inbox waiting to be filled by software. It is a limited, fragile resource. Sometimes you are concentrating. Sometimes you are tired. Sometimes you are trying to complete a form while your brain is already negotiating three other jobs. In those moments, an assistant that pops up with, “Would you like me to help?” has achieved the digital equivalent of tapping you on the shoulder every seven seconds.

That is not assistance. That is a Labrador with a clipboard.

The trouble is that AI systems are generally rewarded for visible activity. A useful answer can be counted. A generated email can be logged. A notification can be attributed to engagement. Silence, meanwhile, looks suspiciously like the product has gone on strike.

Yet a genuinely helpful assistant needs to know that doing nothing is sometimes its most intelligent action.

Help is not neutral

Every prompt, notification and suggestion imposes a cost. It may be small, but it is not zero. You have to notice it, decide whether it matters, dismiss it or act on it, then find your way back to whatever you were doing before. Psychologists call some of this switching cost; ordinary people call it losing the plot.

For disabled users, that cost can be particularly sharp. Cognitive load, fatigue, pain, sensory overload and limited dexterity can all make a supposedly minor interruption more demanding than its designers expect. An assistant that keeps asking for confirmation may feel reassuring to the team that built it, but it can turn a straightforward task into a small endurance event.

This is not an argument for removing assistance. It is an argument for treating attention as something the system borrows, rather than something it owns.

There is an important engineering distinction here. A system can be capable of making a suggestion without being confident that this is a good moment to deliver it. Most assistants collapse those two questions into one. They detect an opportunity and fire the notification cannon. Splendid. Another red badge has been born.

Context is more than a calendar entry

AI companies talk endlessly about context, usually meaning that the assistant remembers the previous few messages or can inspect a document you have open. That is useful, but it is thin context.

Real context includes the consequences of interruption. Is the user driving? Are they in a meeting? Have they ignored the same prompt three times? Are they rapidly editing a document, suggesting concentration? Is the phone in Do Not Disturb mode? Has the user explicitly said they only want alerts about urgent matters?

Some signals are technical and relatively straightforward. Operating systems already expose focus modes, notification permissions, screen state and application activity. Others require care. Inferring stress, health or emotional state from behaviour is both unreliable and deeply personal. An assistant should not decide you are anxious because you typed quickly, then offer breathing exercises with the confidence of an underqualified life coach.

The sensible design is graduated uncertainty. When the system knows little, it should interrupt less. When a user has clearly defined preferences, it should follow them. When an event is genuinely urgent and time-sensitive, it can earn the right to be more assertive.

That sounds obvious, which is usually a sign that it will be ignored until somebody turns it into a compliance requirement.

The missing control: a quietness setting

We are familiar with privacy settings, accessibility settings and notification settings. AI products need something more explicit: a control for conversational intrusiveness.

Not merely a switch labelled “notifications on/off”, but a set of understandable choices about how the assistant behaves:

• Answer only when asked.
• Suggest help quietly, without interrupting the current task.
• Interrupt only for deadlines, safety issues or chosen priorities.
• Ask before using personal context to make proactive suggestions.
• Let the user silence a topic permanently, rather than treating “no” as an invitation to try again next Tuesday.

The last point matters. Repeated refusal is data. If I dismiss a suggestion to summarise every document I open, the system should learn that I prefer reading my own documents, eccentric though that may appear in 2026.

A good control surface also needs to be accessible. It cannot be buried behind six menus, a cheerful illustrated mascot and a questionnaire that asks you to rank your productivity ambitions. If quiet mode is important, it should be easy to find, easy to understand and easy to reverse.

Engineers need to measure interruption debt

There is a practical way to improve this: measure the harm caused by unnecessary interventions.

For every proactive prompt, a team can record whether it was accepted, dismissed, ignored, or followed by the user disabling notifications. It can measure repeated prompts, timing, task abandonment and how often people use “never ask again”. These metrics are imperfect, but they are better than treating every displayed suggestion as a success.

Call it interruption debt. Every unwanted nudge spends a little of the user's patience and trust. Spend enough of it and they stop listening altogether, including when the assistant finally has something useful to say.

Temperature monitoring systems taught me a related lesson. An alarm that fires constantly for trivial deviations soon becomes background noise. Operators learn to acknowledge it without thinking, and the alarm that genuinely matters risks being missed. Alarm management is not about producing the most alarms. It is about making the right alarm credible when it arrives.

AI assistants face the same problem, except they are now following us into our writing, work, homes and conversations. That makes restraint a design requirement, not a nice personality trait.

The best assistant will not be the one that eagerly completes every sentence, schedules every thought and offers to optimise your lunch. It will be the one that understands you are busy being a person, waits patiently nearby, and speaks when it has earned something worth saying.