NORTHWIND watches public social activity and flags people who look like they might need our client’s service. The obvious way to build something like this is to maximise volume: flag more people, message more people, book more calls. That is also how most tools in this category are built, and it is why most of them eventually get their accounts restricted.
Early on, the system was flagging around two hundred prospects a day. The client’s team could realistically follow up with thirty. So the real question was never “how do we detect more?”, it was “how do we detect the right thirty?” That reframing changed most of the engineering that followed.
What restraint looks like in practice
We added three kinds of hesitation to the system. First, confidence thresholds that err toward silence: if the signal is ambiguous, the system says nothing rather than guessing. Second, pacing rules that cap outreach well below what the platform technically allows, because accounts that behave like humans survive and accounts that behave like scripts do not. Third, a mandatory human review step before any first contact.
Throughput dropped 41% in the first week after these changes. The client noticed, and asked the reasonable question. Then the numbers that actually matter started moving: reply rates roughly doubled, and the proportion of conversations that turned into booked calls went up by more than that. Fewer, better-chosen conversations beat more of them.
“The platform rewards volume. The operator rewards judgement. You have to pick one, and only one of them pays the invoice.
Why we’re writing this down
Because “better” is a slippery word in detection systems. If you measure the system by how much it does, you will build a louder system. If you measure it by what happens after it acts, you will build a quieter one. We now put that choice in writing at the start of every detection engagement, before anyone argues about models.