Context Switching in AI

Context switching is what happens when a customer, halfway through booking, asks whether you take Amex, then asks what time you close, then says 'ok let's do Thursday'. An agent that can only run one flow at a time either ignores the interruption, breaks the booking or answers the question and forgets the booking existed. In a voice or chat agent, context switching is the ability to suspend a task, handle a different one and resume the first with its collected details intact.

How it works. The agent keeps a structured state alongside the transcript: the active task, its filled slots (service, date, name), and a stack of suspended tasks. When intent recognition on a new turn returns something unrelated to the active task, the agent decides whether it is a side question (answer inline and continue), a new task (suspend the current one, start the new one) or a cancellation of the current task. Side questions are answered from retrieval without touching the state. When the new task completes, the agent checks the stack and resumes: 'Back to your booking, I still need a time for Thursday.' Language models handle the conversational side well; the reliability comes from the explicit state, not from the model re-reading the transcript.

Why it matters for completion rates. Many abandoned bookings and half-captured leads in transcripts happen at a topic change. The customer asked a reasonable question, the agent answered it and then started over, and the customer gave up. On voice this is worse, because a caller who has already said their name and problem once will not say it twice. Good context switching shows up as fewer re-asks, higher task completion per conversation and shorter handling time, because the agent carries the thread.

Example. A caller to a heating and cooling contractor's line starts an emergency repair request: no cooling, address collected, unit type collected. Then: 'actually, how much is the diagnostic fee?' The agent answers from the price list, 89 dollars, waived if the repair is booked, and continues: 'Shall I keep going with the emergency visit? I have your address on Elm Street, I just need a callback number.' The caller then asks whether the contractor services Ridgewood; the agent confirms from the service-area document and again returns to the callback number.

Context switching vs contextual awareness. Contextual awareness is remembering what was said: the name, the address, what 'it' refers to. Context switching is managing multiple tasks over that memory: knowing which one is active, which are waiting, and when to return. An agent can have good memory and still fail at switching, by answering a side question and then treating the next message as the start of a new conversation.

What to measure.

Botcadence voice agents handle compound intents and transfer with full context, and chat agents hand off with the transcript; see chat agents and multi-turn dialogue.