bAIbel TMS is in development. It is not on sale, and nothing here is an offer.
Machine translation is metered work bought from somebody else, and most systems tell you what it cost after it happened. This one settles what a run may cost before it starts — and when it cannot work out what a run will cost, it refuses to start it.
What follows is exactly what the budget can and cannot promise. The limits are stated because a bound you have to discover is not a bound.
If an enforcing budget applies and any AI model in the plan has no rate in force, the run does not start. Not a warning, not an estimate marked approximate — a refusal, naming what could not be priced.
A budget cannot bound spend it cannot compute. Admitting the run anyway would be claiming a limit with no way to calculate it.
An estimate that errs low is the expensive direction, so enforcement does not reason with the raw figure. It multiplies by how wrong this workflow's estimates have actually been — the 90th percentile of its own history, at documents of this size, from the record kept on every completed run.
The margin is not somebody's cautious guess. It is that workflow's measured error.
When you ask what a run will cost, you are told the estimate. When a budget decides whether to admit it, it uses the padded figure. An under-estimate is harmless in a quote and expensive in a bound, so the two are deliberately not the same number.
Which budget. What is left. What this run needs. What was added and why. A refusal nobody can act on is just an outage.
The whole deployment, one model provider's pool, one client, or one project. By calendar month, calendar quarter, a rolling window of N days, a fixed period for a campaign or a pilot — or no period at all, when the amount is a total and that is the point.
Where several apply, the most restrictive decides.
Refuse it. Park it, to be resumed when there is room. Or hold it for a decision, with the shortfall attached, so somebody can grant the increase rather than being told to go and find the right screen.
Three different answers to "there isn't enough", because they are three different situations.
An admitted run reserves what it is expected to need, and the next run compares itself against what is spent and what is held. Without that, every run admitted in the same moment measures itself against the same remaining balance, each one fits, and the budget is overshot as many times as there were runs.
This is not hypothetical here: the client portal and the agent door both exist so that work can be admitted by nobody in particular.
A reservation expires. A run that parks gives its headroom back — least of all may a run parked because a budget ran out sit on the very money it is waiting for.
Admission bounds expected spend. The estimate is an estimate, so a second control watches actual spend as the run proceeds, checked at each call boundary rather than mid-call.
That means actual spend is bounded by the ceiling plus whatever was already in flight when it tripped — one call per stage running at the time. A call the provider has started is a call the provider will bill, and aborting it would destroy a result you have already paid for rather than save the money.
When a call is cut off after the provider answered, its cost is not known. It is filed as unmetered and flagged, rather than quietly counted as nothing — because a zero in a spend ledger is a claim, and a gap is the truth.
No enforcing budget means no gate. A small team that wants to see what things cost sets a budget to advisory, gets the reporting, and is never refused anything. The machinery arrives when you ask for it.
