· 6 min read
When to call a tool mid-call
Calling out to a CRM mid-conversation buys accuracy and costs silence. When that trade is worth making, and how to cover the wait.
by Woise Team
The difference between an agent that answers questions and an agent that gets something done is whether it can touch your systems while the caller is still there. That capability is also the easiest one to overuse, and an agent that calls a tool on every turn is slower and less predictable than one that does not.
Call a tool when the answer must be true right now
The clear case is anything where a stale answer is a wrong answer. Appointment availability, order status, stock, account state. These change between the time you could have cached them and the time the caller asks, and the whole point of answering live is that the caller does not have to be rung back.
The clear non-case is anything static. Opening hours, your returns policy, what a service includes. Those belong in the agent's material, where they cost nothing per call and cannot fail halfway through a sentence.
Every tool call is a wait the caller can hear
A lookup that takes a moment is a moment of silence in a conversation, and silence in a conversation is loud. This is the real budget you are spending, not the API quota.
So it is worth asking what happens during that wait. Acknowledging it in one short line is usually better than nothing, and asking the next question while the lookup runs is better still where the flow allows it.
Design the failure before you need it
The system will be slow or unavailable at some point, and that path deserves as much thought as the happy one. The choices are roughly: retry, fall back to collecting the details manually, or hand over to a person.
Which is right depends on the call. For a booking, taking the details manually is a good outcome, because the customer's job is done from their point of view. For an account lookup that gates everything after it, a handover is more honest than continuing without it.
Write, do not just read
An agent that reads your systems but cannot write to them leaves somebody reconciling afterwards, which is most of the work you were trying to remove. The booking should exist before the call ends.
That also makes the agent's behaviour checkable. If the record is written during the call, the record and the transcript can be compared, and you are not relying on the agent's own account of what it did.