Reading the meter, and the shape of a month

A refusal does not spend a reply. The material was checked, nothing cleared the threshold, the fallback wording was returned and no model was involved, so the allowance is untouched. An assistant refusing all day costs nothing and consumes nothing.

An answer served from the cache does not spend one either. Somebody asked a question close enough in meaning to one already answered for that assistant, the previous answer came back, and nothing was produced.

Everything else does. That includes the ordinary case of a question answered from your material, and it also includes the short conversational replies: a greeting, a thank you, an answer to somebody asking what the assistant can do, and the reply to somebody asking for a person. Those are answered without touching your material and without producing anything, they add nothing to the spend figure, and each of them still takes one off the allowance. It is the single most common surprise in this whole area, and it is worth knowing before you compare the meter to anything.

The messages figure on the insights page counts every assistant message over the last thirty rolling days, for one assistant, including the refusals and the cached answers. The meter counts only the replies that spend one, across the whole workspace, since the current period began.

Three differences at once, then: a different rule about what counts, a different scope, and a different window. Two figures that both look like a count of replies will not match, and neither is wrong. If you are reconciling them, start by noting that the meter is per workspace, so several assistants share one allowance and no screen splits it between them.

The period is the workspace's own rather than a calendar month. It runs from the date the current period began, and that date is set again when a plan change takes effect, which is why an upgrade does not inherit the usage that prompted it.

A total tells you whether you fitted. A shape tells you whether the meter is the right size. Even consumption across the period means the allowance matches the traffic, and a plan chosen on the total will keep matching.

Consumption concentrated into a few days means the total is the wrong thing to size against. A launch, a newsletter, an outage on another channel or a seasonal peak can spend a fortnight's worth in an afternoon, and the plan that comfortably fits the month runs out on a Tuesday. The daily series of messages and refusals for an assistant is what shows this, and it is worth looking at before deciding a tier is sufficient.

A rising floor is the third pattern and the one that creeps. Each week is a little heavier than the last, nothing spikes, and the ceiling arrives one month without anything having happened. That is normally growth, which is the good version, and it still needs noticing in advance rather than at the moment the widget stops.

When the count reaches the allowance the assistant stops answering. What appears in the panel is the offline message set on that assistant, whose wording by default says it is having trouble and to try again shortly. It does not mention a plan and it does not mention a limit.

That is the right behaviour for a whitelabel product and it puts an obligation on you. A visitor on a customer's site is going to read that sentence, so it should be a sentence you would be happy to have read. There is a second ceiling behind it, a spend limit, which stops answers in exactly the same way and shows exactly the same message, so the wording covers both.

The failure is total rather than degraded: the assistant does not answer more briefly or refuse more often, it stops. That is why the counter is in the header on every screen rather than tucked into a billing page.

Raise the threshold. Refusals do not spend a reply, so an assistant that refuses more lasts longer on the same plan. The meter reads as thrift and the product is doing less, and there is nothing on the usage screen that would show you which of the two happened.

A repetitive week does the same thing without a decision. The cache absorbs the duplicates, none of them spends a reply, and the allowance stretches for reasons entirely to do with the traffic mix. The following week of varied questions burns it faster at identical volume.

The accidental version is sizing a plan on a quiet month. The total fitted, the shape was never looked at, and the first busy week takes the assistant offline on somebody's live site. Read the daily series before believing a comfortable total.