Thinking ahead

What makes an AI assistant proactive?

A proactive AI assistant notices a relevant change and initiates a useful next step without a new request every time. It still needs a goal, access to the right information and boundaries on what it may do. Proactive does not mean unrestricted.

A reminder is one part of the picture

A reminder repeats something you already decided to remember. Proactive assistance can go further by connecting a new event with an existing goal: a late flight affects your pickup, a moved appointment affects the school run, or an unanswered request needs a follow-up.

The important test is relevance. An assistant that sends more notifications has not necessarily reduced your mental load. A useful intervention explains what changed, why it matters to you and the next decision or action.

Look for five connected parts

Use this sequence to judge a proposed proactive feature. If a step is missing, ask how it will work rather than assuming the assistant can fill the gap.

  • A signal: a specific change the assistant is permitted to observe.
  • Context: the goal or preference that makes the change relevant.
  • Authority: an existing permission, or a request for your decision.
  • Action: a concrete next step through an available service.
  • Evidence: a clear account of what happened and what remains open.

Set permissions around decisions

“Handle my calendar” is difficult to interpret. A narrower rule is easier to use: “You may move my own flexible focus blocks, but do not change events with other people.” You can expand it later after seeing how it behaves.

A useful permission includes the action, the scope, exceptions and when it expires. Permission to read an itinerary is different from permission to pay for a change. A preference for morning appointments is not permission to cancel an existing booking.

More initiative should mean less noise

Ask how the assistant decides whether to interrupt you. Some events deserve an immediate question; others can wait for a summary. A system should be able to say that nothing useful needs your attention.

Also ask how it handles uncertainty. A tentative flight update should not trigger a chain of irreversible changes. When sources conflict, the next useful step may be to verify the information instead of acting on it.

Test the boundary as well as the happy path

When evaluating a proactive assistant, try an ordinary change and an exception. Does it notice the first? Does it stop at the second? A record of attempted and confirmed work makes both outcomes easier to understand.

Research on agent design distinguishes a predetermined workflow from a model choosing its own next steps. A proactive product may use either approach. What matters to you is whether the trigger, permissions and result match the task.

Reference: Anthropic: workflows and agents

Twin’s direction

Twin’s ambition is to use your chosen context and standing permissions to notice useful next moves before you have to ask. The delayed-flight example on the Twin site illustrates that direction; it is not a report of a live customer booking.

The proactive product remains in development. The early-access invitation is a way to introduce yourself to the team, not to switch on background monitoring of your accounts.

Meet the brand behind these guides.

Your own Twin. More room for life.

We’re building a personal AI helper around the context and permissions you choose. Introduce yourself to the team for early access.

Ask about early accessLearn about Twin