The AI budget guard, end to end

Why a separate guard exists at all#

Telephony minutes and model spend are two different accounts, and the carrier's cap says nothing about whether the model can be paid for. The outbound path booked minutes against the telephony ceiling and nothing asked the other question, which is how the AI ledger ran past a configured limit with blocking switched on and calls still connecting.

The guard closes that gap by asking, before a word is spoken, what one call *could* cost if it ran to the full configured ceiling — both sides of the conversation at the audio rate for the whole cap. Not what it will cost: what it could, which is the only figure available before the call exists.

The two refusals#

Stage by stage#

1. Trigger — an outbound call is about to be placed, or an inbound call has arrived at a line and is about to be answered.
2. User or external event — for inbound, a real person is on the line right now. That fact governs every choice below.
3. Authentication and workspace resolution — the workspace comes from the line that was dialled, not from the caller's number, which is only ever data about the call.
4. Ingest or request — the engine asks the guard, passing the voice configuration so the ceiling and the call cap are the ones actually in force.
5. Canonical record — a refused inbound call still creates a call row. Before that, a refusal was a call nobody could see afterwards.
6. Reasoning — none yet. The guard runs before the model session opens, which is the entire point: an unaffordable call should not first cost a model call to discover it.
7. Knowledge, memory and rules — the closed-line message is composed from the business's own persona, so a refusal sounds like the business rather than like a system.
8. Autonomy and approval — untouched. The guard is about affordability, and a workspace's rules about placing calls are checked separately, on their own terms.
9. Action through a provider — on a refusal, nothing is asked of the carrier for an outbound call; for inbound, the worker speaks one sentence and hangs up.
10. Result — outbound: a refusal carrying the reason. Inbound: an ended call, zero duration, hung up by the agent, with the refusal code as its outcome.
11. Relationship, timeline and memory — the call appears in the caller's history as a call that was answered and closed with a reason, not as a mystery.
12. Audit, usage and Needs You — the refusal is recorded, the summary says the caller heard the closed-line message, and the line's health surfaces for a person.

What the caller hears#

A sentence, in the business's voice, and then the line ends. It is not a good outcome — a caller who wanted an answer did not get one — but it is an outcome, and the alternative is worse than it sounds. Answering into a model that cannot speak produces a connected call with nothing on it, which the caller experiences as being ignored by a business that picked up.

The record matters as much as the sentence. A refused inbound call is a call row with a zero duration, an outcome naming the refusal, and a summary saying the caller heard the closed-line message. Somebody reviewing the day can see that three people rang and were turned away, which is exactly the signal that was missing when the room was simply deleted without a word.

• The line being switched off produces the same shape of refusal, with its own code.
• A call outside the line's configured hours produces one too, unless the line routes after-hours calls to the agent anyway.
• All three are visible on the call record and none of them is silence.

Getting calls flowing again#

1. Find out which of the two refusals it is. The reason is on the call record and on the line's health.

   Result You know whether to change a budget or fix an account, which are unrelated actions.

2. If the model account is refusing, resolve it at the provider — the key or the account itself.

   Result The breaker closes on the next successful call rather than on a timer you have to wait out.

3. If it is the budget, raise the ceiling or let the period turn over.

   Result The guard's next answer is different immediately; nothing is cached and no restart is needed.

4. Place one test call before assuming the line is well.

   Result A single successful call is the only evidence that both accounts are answering.

Questions#

Related#

• The budget that stops a call
• What a call costs
• The AI budget is spent
• A provider quota breaker opening and closing
• The call connected and nobody spoke
• Switching the line off