The send classifier is the gate that decides whether a broker draft is safe to send on its own or should pause for you. In the automated modes, every outbound message the agent writes passes through it first. This is a closer look at what it does — the mechanism behind "the agent sends the routine stuff and brings you the commitments."
One thing to be clear on up front: the classifier is a gate, not a chatbot. It doesn't talk to brokers or write emails. Its only job is to judge each draft the agents produce and route it to send or to you.
What it's fed
To make that call, the classifier reads a context block for the draft, including:
- The email thread — the conversation so far.
- Your prior-turn decision — whether you answered a question, gave feedback on a draft, or left a free-text instruction on the last turn.
- Any pending draft waiting on approval.
- Driver long-term memory — the remembered facts about your driver and lanes (see below for when this is off).
- The booking pre-authorization flag — whether you've hit Book It on this load.
What it enforces
With that context, the classifier holds the drafts to the rules that keep you in control. Two it exists specifically to enforce:
- Any claim grounded in remembered facts requires your approval. If a draft's substance leans on driver memory, it routes to you before it can reach a broker.
- An explicit instruction from your prior turn is honored. If you told the agent what to do last turn, the classifier respects that rather than overriding it.
When it's in the path
The classifier isn't always running:
- In Manual, the agent's send skips the classifier entirely — you're the gate, and there's no extra round-trip per email. See Manual mode.
- In Auto-Supervised and Fully Autonomous, every outbound draft goes through it. See Auto-Supervised mode.
The Book It exception
There's one case where the classifier stays in the path regardless of mode. Once you've hit Book It on a load, that thread stays on the checked path even in Manual — a booking is in flight and needs the guardrail.
After Book It, the flag tells the classifier the load is pre-authorized: when the broker replies, the agent is cleared to use commitment language to close without pinging you for a second approval. It still may not re-pitch the offer — pre-authorization closes the load you agreed to, it doesn't reopen the price.
Driver memory is off in Fully Autonomous
In Fully Autonomous, driver long-term memory is off for the classifier's context — the agent won't lean on remembered numbers to make broker-facing claims. In the supervised modes those claims are allowed, but only after they route through your approval.
Related
- Approvals and the human in the loop — what happens when the classifier pauses a draft for you.
- AI modes — how the gate's role changes per mode.
- Book It and the booking flow — the pre-authorization the classifier respects.
- Driver memory, voice, and limits — the remembered facts the gate guards.
Need help? Email support@numeo.ai or use Report a problem in the app.