A human-in-the-loop lead triage workflow that is actually dependable

Start with the routing policy

Before opening an automation platform, write down the decision rules in plain language. What makes an inquiry ready for discovery? Which services need a senior review? What is the fallback when budget, timing, or decision-maker information is missing?

If the team cannot explain those rules consistently, an AI model will only create a faster inconsistent answer. The first deliverable is a written policy with a named owner.

Separate three jobs

A useful triage workflow separates normalisation, recommendation, and action. This makes failures easier to diagnose and keeps consequences proportional to confidence.

• Normalise: validate input, remove obvious spam, and standardise the fields that downstream systems need.
• Recommend: calculate a transparent tier, identify missing information, and suggest an owner using explicit rules.
• Draft: use an LLM only where natural-language synthesis adds value, with the source fields included and the output marked as a draft.

Design normal and failure paths

The normal path is straightforward: an inquiry arrives, a record is created, the owner receives a review task, and the next action is suggested. The failure path deserves as much design attention.

If the CRM is unavailable, do not lose the inquiry. If enrichment times out, route with known information and flag the gap. If an automated draft contains unsupported commitments, block the send. If the same webhook arrives twice, use an idempotency key so it does not create duplicate records.

Measure the workflow, not novelty

Define a baseline before launch: median time to owner assignment, percentage of briefs missing required context, percentage of inquiries receiving a response within the target, duplicate CRM records, and manual edits per draft.

These metrics expose whether the system improves the operation. A high AI acceptance rate without lower response time or fewer missing fields is activity, not evidence of useful automation.

Keep authority with a person

The simplest governance design is usually sufficient: no automatic client-facing send, no automatic approval of commercial terms, and no deletion of source records. A person can approve, edit, or reject the recommendation, and the system records that decision.

This protects the agency relationship and creates the feedback needed to improve routing rules over time.