001
/ Journal /

Why our agents run slower than they could

Latency we could remove and choose not to, because a cheaper model that waits beats a fast one that guesses.

Almost every workflow we run could be two seconds faster. We leave the two seconds in, and clients stop noticing within a week. Here is what they buy.

We check before we act

Before an agent writes anything back, it re-reads the record it is about to change. Stale context is the most common cause of a wrong action, and the read costs a fraction of the cleanup.

We batch on purpose

Invoice chasing runs once each morning rather than continuously. Nobody wants three reminders in an hour because a payment landed between two runs, and the batch is cheaper besides.

We let the gate interrupt

When a case falls outside the handoff rule, the run stops and waits for a person. Measured end to end that looks like poor latency. Measured in hours returned it is the whole point.

Where speed does matter

Anything a customer is waiting on. Support drafts and lead routing run hot, because a person is watching. Back-office work does not need to, and pretending otherwise costs money for no benefit.

002

/ Keep reading

From the journal