How to measure an AI SDR funnel

A funnel stage should correspond to something the system can support from the conversation or tool record. Avoid labels that depend on optimistic interpretation. Positive sentiment is not automatically qualified interest, and a proposed time is not a booked meeting.

Write stage rules before comparing agents. If definitions change between runs, the apparent movement can come from the measurement rather than the system.

The agent may announce a meeting before the calendar tool succeeds. It may classify an account as qualified while missing a constraint recorded earlier. The funnel should require the combination of evidence appropriate to each stage.

This connection also makes failures diagnosable. Teams can distinguish a persuasive conversation with a scheduling failure from a calendar action created without genuine buyer agreement.

Open the cohort that stalled between two stages and compare it with buyers who progressed. Look for differences in fit, authority, urgency, objections, timing, and the agent actions they received. The purpose is to find a pattern that can change the product or narrow the intended audience.

Synthetic rates should be used for controlled comparison and diagnosis, not presented as guaranteed pipeline. Production outcomes remain the external validation of commercial performance.

A useful funnel report points from the stalled stage to the exact conversations, actions, tool results, and telemetry behind it. The team can then change the relevant prompt, context, policy, tool, cadence, or orchestration layer.

Rerun the candidate against the same reviewed population. If progression improves without creating a guardrail or workflow regression, preserve the comparison as evidence for the release review.