AI SDR metrics: the six numbers to track and how to read them

Track six AI SDR metrics: reply rate, positive reply rate, meetings held, cost per held meeting, approval edits and deliverability. Reply rate is the easiest to read and the easiest to misread, because it can stay flat while meetings fall. Decide by cost per held meeting, and use the other five to find out why it moved.

Most dashboards show activity: messages sent, invitations accepted, replies. Activity is cheap for a machine to produce, so it says little about whether the tool earns its place. The useful question is what each number lets you decide. This guide takes the six metrics in the order a conversation moves, shows how to calculate each, and says what to do when it looks wrong.

What is the funnel you are measuring?

Keep the stages as separate counts. Merging them is how a weak campaign looks healthy.

Count people, not messages. A person who gets five follow ups is one person messaged. Counting messages makes every rate look smaller and hides which audiences work.

1. How do you read reply rate?

Reply rate is replies divided by people messaged. It is a signal about the first message and the audience, nothing more. A reply of "not interested" counts, and so does an out of office, unless your tool filters those. Sellinger does: an absence notice is not counted as a reply from the person, and automatic replies from helpdesks and survey bots are marked as machine mail and never count as a human reply. Ask any tool you compare whether it does the same, or your reply rate includes robots.

The trap is treating it as the goal. On our own LinkedIn account, 40% of people messaged replied, and the rate stayed between 35% and 45% in each of twelve monthly cohorts. Meetings per person messaged peaked at 2.4% in November 2025 and drifted below 1% from June 2026 (case study). Replies stayed flat while meetings fell, so a reply rate alone would have told us nothing was wrong. That is one account and one audience, so do not read 40% as a target. Compare your own campaigns against each other.

What to do when it moves:

• Falls: look at the first message and the audience before touching anything else.
• Rises while meetings do not: check the positive reply rate next.

2. Why separate positive replies?

Positive reply rate is positive replies divided by people messaged. Define "positive" before the campaign starts, in one line, for example: asked a question about the offer, asked for a time, or agreed to a call. A person who must decide each case after the fact will drift toward counting what they hoped for.

Read the ratio of positive replies to all replies. If most replies are negative, the offer or the list is wrong, however many replies there are. If positive replies are common but meetings are not, the problem is later in the thread: slow answers, unclear booking, a calendar link nobody uses.

3. What counts as a meeting held?

Booked is not held. People cancel, forget and ghost, and an agent that books an easy yes can produce a calendar full of no shows. Track both counts and the gap between them.

Define a held meeting as one where the person attended and the topic was the one you pitched. Take it from your calendar or CRM, not from the agent's own report. Connect the calendar the agent books into (Sellinger works with Cal.com and Calendly) so the booking record exists outside the conversation.

If the booked to held gap is wide, shorten the time between agreement and meeting, send a reminder, and check whether the audience said yes to be polite.

4. How do you calculate cost per held meeting?

This is the decision metric. Add the full monthly cost of running the campaign and divide by meetings held in that month.

Worked example, with made up numbers:

If that month holds 5 meetings, the cost per held meeting is $153. If it holds 2, it is $382.50. The software did not change; the meetings did. The Start price is our list price and the other lines are invented for the sum, so replace them with your own. For what belongs in the cost lines, see the real cost of owning an AI SDR, and for the comparison with a person, the ROI of an AI SDR vs a human SDR.

Do not divide by meetings booked. It flatters the tool.

5. What are approval edits and why do they matter?

This is the metric nobody reports and the best early warning. When a campaign runs in approval mode, every message waits for a person to approve it. Count, per week, how many drafts you approved unchanged, edited or rejected.

• High unchanged share: the agent writes what you would sign. You can consider moving the campaign to send automatically.
• Edits cluster on one thing such as tone or a claim: fix it once in the instructions or knowledge, not draft by draft.
• High rejection: stop, and read why AI SDRs fail before sending more.

Edits also measure your time cost. A campaign that needs 30 minutes of fixing per day is not cheap, whatever it costs per seat.

6. Which deliverability numbers matter?

Applies to email. A message that lands in spam is not a reply you lost, it is a message that never happened. Watch:

• Bounce rate. A rising rate means a bad list or an aging data source. Verify addresses before sending.
• Spam rate in Google Postmaster Tools. Google's sender guidelines say to keep it below 0.3% and recommend staying under 0.10% (Google).
• Authentication. SPF, DKIM and DMARC set up on the sending domain. Google lists these for bulk senders (more than 5,000 messages a day to Gmail), and they are sensible for anyone.

For LinkedIn the equivalent is the acceptance rate and whether invitations pile up unanswered, which LinkedIn names among the reasons it limits accounts (LinkedIn Help). See LinkedIn invitation limits.

How should you review the numbers each week?

1. Open the funnel table for each campaign, split by channel.
2. Compare each count with the stage before it. Find the stage with the largest drop.
3. Check approval edits for that campaign. A drop in quality usually shows there first.
4. Check deliverability or acceptance for the channel.
5. Change one thing, write down what and when, and wait for a full cycle of follow ups before judging.

Changing three things at once leaves you with a number that moved and no idea why.

What should you not track?

• Messages sent. It measures effort, and the agent's effort is free.
• Open rate. Email privacy features inflate opens, so it no longer shows that a person read anything.
• Revenue in the first month. Sales cycles are longer than the test. Judge the test on held meetings, and judge the pipeline later.

For how long to run the test and what a fair trial looks like, read do AI SDRs actually work. The Bridge Group's SDR metrics report is a source for how human teams are measured, if you need to compare like with like.

What can't we promise?

We cannot give you a benchmark for your market. Rates differ by channel, audience and offer, and the only fair comparison is your own campaigns over time. Our figures here come from one account and from invented examples labelled as such.