DEV Community

Ujjwal Dubey
Ujjwal Dubey

Posted on

Route Easy Requests to a Smaller Model: A Cost-Aware Classifier Step for Small-Business Workflows (Python)

Not every incoming message needs your most capable model. "What time do you open?" and "Please read this contract and flag the risky clauses" should not be priced the same way.

A small routing step keeps AI automation costs down for a small business without hurting answers that matter.

The pattern

  1. Cheap rules first (keywords, length, attachments).
  2. If rules are unsure, a small model classifies difficulty.
  3. Only "hard" items go to the larger model.
  4. Log which route each item took, so you can check the split.
EASY_HINTS = ("opening hours", "address", "price list", "where are you")

def rule_route(msg):
    text = msg["text"].lower()
    if msg.get("attachments"):
        return "hard"
    if len(text) < 120 and any(h in text for h in EASY_HINTS):
        return "easy"
    return None

def route(msg, small_classify, log):
    r = rule_route(msg)
    if r is None:
        r = small_classify(msg["text"])  # returns "easy" or "hard"
    log({"id": msg["id"], "route": r})
    return r

def handle(msg, small_model, large_model, small_classify, log):
    if route(msg, small_classify, log) == "easy":
        return small_model(msg["text"])
    return large_model(msg["text"])
Enter fullscreen mode Exit fullscreen mode

Guard rails

  • Anything the small model answers still passes the same approval step.
  • Sample "easy" answers weekly; if quality drops, tighten the rules.
  • Keep the router dumb and readable. Its job is triage, not cleverness.

Check the effect

Compare a week before and after: how many items each route handled, and how many easy answers needed a human edit.

We use variations of this in workflow builds for small teams.

Top comments (0)