Skip to content

Automated Classification and Routing

Software tool — instantiates Net-Additive Contribution Intake

Reads each standardized offer and mechanically sends it to the right destination — review queue, self-service path, alternate recipient, or decline — applying triage rules at volume without staff touching every one.

Automated Classification and Routing is the tool that turns a written triage policy into an executable one. Given a standardized offer record, it classifies the offer by its declared attributes and dispatches it to the correct next destination — a human review queue, a self-service path, an alternate recipient, or a decline notice. Its defining property is that it is a mechanical executor, not a judge: it applies the disposition rule at whatever volume arrives, consistently and without staff attention, but it decides only where each offer goes next, never whether the offer is genuinely worth taking. That deeper judgment stays with the human review it feeds.

Example

A large environmental nonprofit's "offer to help" inbox receives everything from a donated pallet of laptops to a volunteer's spreadsheet of grant leads to a developer's offer to rebuild the website. A coordinator used to read each one and figure out who should handle it — hours a week, inconsistently. They wire up Automated Classification and Routing on top of their intake form. Now an offer tagged physical goods, out-of-region is routed straight to the alternate-recipient path with a suggested local partner; an offer tagged small, self-installable digital asset is sent to a self-service upload page; anything tagged skilled labor, ongoing commitment is dropped into the human review queue where a person decides if it's actually additive.

The coordinator's inbox stops being a sorting job and becomes a review job: they now see only the ~15% of offers that genuinely need a judgment call, each already in the right queue, while the rest have been dispatched consistently the moment they arrived.

How it works

What distinguishes the router from a person doing the same sorting is that it operationalizes the rules:

  • Classify from the record. It reads the standardized attributes an offer already carries and assigns it to a category — no re-interviewing the contributor.
  • Apply the disposition rule. Each category maps to one of the five dispositions (accept-to-review, reshape, redirect, defer, decline), so the routing decision is the disposition decision made explicit.
  • Dispatch, don't decide worth. It moves the offer to that destination; whether an accepted offer is net-additive is left to the review it routes to.

Tuning parameters

  • Confidence threshold — how sure the classifier must be before auto-routing versus escalating to a human. A high bar routes fewer offers automatically but makes fewer wrong calls; a low bar clears volume at the risk of confident misfiling.
  • Rules vs. learned model — hand-written rules (transparent, brittle on novelty) versus a trained classifier (handles variety, harder to audit). The trade-off is explainability against coverage.
  • Destination granularity — how many distinct routes exist. More routes place offers more precisely but multiply the rules to maintain and the ways to misroute.
  • Override friction — how easily a human can reclassify a routed offer. Easy overrides catch machine errors and generate training signal; frictionless ones let staff quietly bypass the routing entirely.

When it helps, and when it misleads

Its strength is throughput with consistency: it applies the same triage rule to the hundredth offer as to the first, at a speed no human sorter matches, and it frees scarce staff to spend their judgment only where judgment is needed.

It misleads through automation bias — once routing is automatic, people tend to trust the destination even when the classification is plainly wrong, so a miscategorized offer sails to the wrong place unchallenged.[1] It is also brittle on genuinely novel offers that fit no existing category, and it invites a quiet misuse: tightening the rules to auto-decline more offers as a way to shrink the queue, which hides demand rather than handling it. The discipline that keeps it honest is routine audit-sampling of routed offers and a low-friction human override, so the machine sorts but people stay accountable for where things land.

How it implements the components

Automated Classification and Routing realizes the sorting-and-dispatch components — the executable core of triage:

  • triage_policy — it is the running instantiation of the triage rules, applying them to every offer at volume.
  • accept_reshape_redirect_defer_decline_rule — each classification maps to one of the disposition outcomes, so routing is the disposition rule enacted mechanically.

It consumes the standardized record produced by Pre-Screening Form rather than creating it, it does not judge whether an accepted offer is net-additive (that is the human contribution_net_value_review it routes to), and it does not own the alternate destinations it points at — those belong to Overflow Redirection.

Notes

The router decides where, never whether it's worth it. It can send an offer confidently to the accept-review queue and still be routing something that should ultimately be declined — the net-value judgment is a separate human review the routing feeds, not a step the routing performs.

References

[1] Automation bias — the tendency to over-trust the output of an automated system and under-scrutinize it, so an obviously miscategorized offer passes unchallenged simply because the machine placed it. Audit-sampling and easy overrides are the standard correctives. ↩