Measuring an AI SDR

What each tool calls a reply and a meeting

A reply rate is a share of something, and each tool picks a different something. Here is how four count a reply, and two a booked meeting, on their own pages.

The base moves the number. As an example, take 600 emails sent to 200 contacts, and 12 replies from 10 of them: per contact, the reply rate is 5%; per email sent, it’s 2%. Instantly’s benchmark counts per email sent, follow-up replies included, across all cold email on its own platform from 1 January to 18 December 2025: an average of 3.43%, and 5.5% for the top quartile. So work your rate out its way before you set the two side by side.

One funnel of your own

So keep one funnel of your own, counted the same way every month (support part 12 says why). Count people at every step, not emails. Follow each month’s new contacts as one group for a fixed time, 60 days for example, so that a meeting held in week six still counts for the month its contact came from. Here is a month at Example Software, an invented company that sells to finance teams, with two SDRs and an AI SDR.

An agent’s own report may end sooner. HubSpot’s Prospecting Agent lists its outcomes up to “Replied emails and booked meetings”. So have whoever ran each meeting mark in the CRM whether it took place, and take opportunities and pipeline from there. Then divide the month’s cost (the agent, the data, the sending) by the meetings held, and set that beside the same figure for your SDRs.

The warning counts beside it

A funnel can look healthy while the domain it sends from wears out. So read three counts beside it every week: the spam rate, unsubscribes and bounces. Google asks senders to keep the spam rate in Postmaster Tools below 0.10% and to avoid ever reaching 0.30%, and Yahoo asks for below 0.3% (part 11 has the rules, and which mail each count covers). Both work the rate out on mail delivered to the inbox, so complaints divided by all the mail you sent give a different number.

Two readings mislead. Postmaster Tools can leave out days when you send little, so a new agent may show nothing at all: an empty day isn’t a clean one. And when a high rate suddenly falls to 0, Google gives this possible cause: “many messages are automatically being sent to spam, and recipients aren’t reporting any inbox messages as spam.”

Count unsubscribes per contact delivered and bounces per email sent, and compare each week with your usual week. In HubSpot, both are in the sequence’s Email engagement report, apart from its reply and meeting rates, so copy them into your funnel’s sheet. If either jumps, pause the agent’s sending, then look: bounces point to your contact data (part 4), unsubscribes to who it writes to and what it says (part 5).

The agent and your SDRs, on one segment

Set the agent’s funnel beside your SDRs’ only for the same segment, in the same weeks, counted the same way. If the agent writes to the smaller firms and your SDRs to the larger ones, the totals compare firms more than senders (support part 13 shows how such a mix can turn a comparison round). So split one segment between them at random, such as finance directors at logistics firms of 200 to 2,000 people, and compare positive replies, meetings held and pipeline per contact, and the cost per meeting held.

If your SDRs call and the agent only emails, compare email replies alone: Outreach’s reply rate counts a prospect who answered a call as a reply. And give it time: with seven replies, one reply read differently moves the share of positive ones by 14 points.

Next in this series

Part 13: why did it email this person? The record behind each contact. See all 13 parts.

Why we write these guides