← All articles

AI SDR Metrics That Matter (and Vanity Numbers to Ignore)

October 19, 2026 · 9 min read · by the 0effort team

TL;DR: Judge an AI SDR on outcomes: positive replies, meetings held and qualified pipeline per qualified account reached. Pair them with failure metrics: off-ICP sends, personalization errors, human-intervention rate and spam complaints below Google's recommended 0.1% (0.3% is the hard ceiling). Emails sent and sequences launched cost an AI nothing, and machines inflate opens, so none measures performance.

Open an AI SDR dashboard and the biggest number is climbing. Emails sent: 14,200. Contacts enriched: 6,800. Sequences running: 41. It looks like a department at full tilt, and it's the one part of the report a machine can inflate by changing a setting. We build an AI SDR, and our view is blunt: the activity half of that screen is decoration.

Why activity metrics break for an AI SDR

Because they were never about the work. A human SDR's call count and email count stood in for effort, and effort was the scarce thing. Forty personalized emails a day was a real day's labor, so counting them told you something.

An AI SDR's marginal email costs close to nothing. Activity stops being a proxy for effort and becomes a dial you set. Worse, the dial points the wrong way: sending ten times more to a loosely targeted list moves "emails sent" up and every outcome you care about down, because complaints and bounces rise while the replies stay the same.

So the rule we work by is simple. Any metric the AI can raise without a prospect doing anything is an input. Report it for capacity planning if you like, and never as performance.

The vanity numbers to ignore

Most of them. Here are the usual suspects and what to look at in their place.

Opens and clicks aren't just noisy; machines contaminate them, so we won't build a decision on either. "Meetings booked" is the sneakiest, because it sounds like pipeline. It isn't. A calendar slot is a promise, and a meeting held is the first moment anything real happened.

The north-star metric: qualified pipeline per account reached

If you track one outcome number, make it this one, and note that the denominator does most of the work.

Qualified pipeline per qualified account reached = pipeline value from qualified opportunities created from the AI SDR's meetings, divided by the number of in-ICP accounts that actually received a message from it. Three definitions matter. "Qualified account" means it passes the written ICP rule, not that someone liked the logo. "Reached" means delivered to a right-role contact, so bounces don't count. "Qualified opportunity" means whatever stage your sales lead already uses for real pipeline, fixed before the pilot starts.

Why divide by accounts? Because it punishes the behavior activity metrics reward. If the AI doubles volume by wandering outside your ICP, the numerator barely moves, the denominator balloons, and the number falls. You can't game it by sending more.

It lags by a full sales cycle, so it can't be your only metric. Everything below is a leading indicator of which way it will move.

Reply quality, not reply volume

Sort every reply into a bucket and report each separately: positive (wants to talk), referral ("try my colleague"), negative, unsubscribe, and auto-reply (out-of-office, ticket bots). The headline is the positive reply rate: positive replies divided by qualified accounts reached. Exclude auto-replies from the reply counts; they aren't people.

Track referrals on their own line; lumping a warm introduction in with "no" understates what the system found. We cover how a reply gets sorted in how AI SDRs handle replies.

One habit pays for itself: the classification is a metric too. Pull 20 labeled replies a week and read them yourself. If the AI filed a rejection as interest, every number built on that label is wrong, and no dashboard will tell you.

Metrics that catch the AI going wrong

These are the autonomy-specific failure costs, and in our experience few dashboards report them. A human SDR who emails the wrong person has had a bad afternoon; an AI doing it at scale causes a reputation event. Three numbers keep you honest:

• Off-ICP send rate. Sample 25 sends a week and check each recipient against your written ICP rule and exclusion list (customers, open deals, competitors). Off-ICP sends divided by sampled sends. The target is near zero, and any customer or open-deal contact is a failure on its own, regardless of the rate.
• Personalization error rate. In the same sample, check every factual claim in the first line against its source. Congratulating someone on a funding round that never happened is the classic. Emails with at least one wrong claim divided by emails checked. One bad claim costs more than ten bland openers.
• Human-intervention rate. Of the AI's reviewed actions (drafts, routing decisions, targeting picks), the share a human edited, rejected or reversed.

The third one needs care in both directions. A high rate early is normal; a rate that falls as you fix the brief means the system is learning your business. But a rate pinned at zero usually means nobody's looking, so pair it with the sampling above and trust neither alone. Separately, one rule has no rate: every opt-out request is honored, and every reply touching price or contract terms is routed to a human, with zero exceptions.

The hidden cost: sender reputation and spam complaints

Reputation is the balance sheet item the activity dashboard leaves off. Every send draws on it, and you only find out it's overdrawn when replies quietly stop.

The figure that matters is your spam-complaint rate. Gmail's and Yahoo's 2024 bulk-sender requirements set 0.3% as the ceiling (Google's email sender guidelines, accessed 19 October 2026; Yahoo's sender best practices, accessed 19 October 2026), and Google recommends staying below 0.1%. Treat 0.1% as your working limit; by the time you touch the ceiling you're already being filtered. Read it in Google Postmaster Tools per sending domain, never as one blended average, because a single bad domain hides inside a healthy mean. Postmaster Tools shows little for domains with small volume, so a low-volume pilot needs the other signals too: bounce rate, seed-test placement and a drop in positive replies while sends hold steady.

Plot these weekly. A complaint rate creeping from 0.02% toward 0.08% says slow down and tighten targeting, long before any line is crossed. Setup (authentication, warm-up, domain separation) lives in our deliverability checklist.

How long to wait before judging the metrics

Longer than a vendor demo suggests, and in layers. Early samples are tiny, so early weeks read failure metrics, not outcomes.

Weeks 1–2. Ignore reply rates. Read the failure numbers: off-ICP sends, personalization errors, bounces, intervention rate. At this volume they find problems faster than any rate can rank copy.

Weeks 3–4. Read the replies, not the rate. You'll have enough to see categories: a segment answering only "wrong person," an angle that irritates. Change one variable a week. First meetings may be held, which is a bonus, not a verdict.

Weeks 5–8. Now positive reply rate and meetings held are worth trending, against your own manual baseline. Ask whoever sat in each meeting if they'd take it again.

Month 3 onward. Qualified opportunities and pipeline per account. If your sales cycle runs 90 days, closed revenue from the first sends won't exist before the end of the quarter, so judge those cohorts on opportunity creation. Our implementation checklist covers the first 30 days in detail.

A scorecard your CFO will trust

Build it from things finance can audit: cost, meetings held, opportunities. Print your assumptions, because a scorecard with hidden ones is a marketing slide.

Here's an illustrative month. The inputs are made up to show the arithmetic, not benchmarks; replace every one.

• 1,000 qualified accounts reached
• 30 positive replies (3.0% positive reply rate)
• 12 meetings booked, 9 held
• 3 qualified opportunities, at an average $20,000 each
• Cost: $349 plan (0effort's Growth price) plus 10 hours of human review at $60 an hour = $949. Domain and data costs are left out to keep it simple; add yours.

Results: $949 / 9 = about $105 per meeting held, and $949 / 3 = about $316 per qualified opportunity. Pipeline is 3 × $20,000 = $60,000, or $60 per qualified account reached.

Notice what's in the cost line: the humans. In our experience, review time is the cost AI SDR business cases most often forget, and a CFO will find it in five minutes.

FAQ

What is a good positive reply rate for an AI SDR?

We won't give you a number, because any figure we printed would be a guess dressed as a benchmark. Vendors publish self-reported rates, and your market, list and offer will differ. Set your baseline from your own manual outbound, then watch whether your positive reply rate rises as targeting tightens. Direction against your own history beats a comparison with a stranger's.

Should I still track open rates for AI-driven cold email?

Not as a performance metric. Apple Mail Privacy Protection, introduced in 2021, pre-fetches images and inflates opens, and corporate security scanners auto-click links. Treat both as directional at best.

How do I attribute pipeline to an AI SDR when humans also touch the deal?

Pick a rule before the pilot and don't change it. We'd use "sourced": an opportunity counts for the AI SDR if its first held meeting came from the AI SDR's outreach, even if a rep closes it. Record the source on the opportunity when it's created. Report "influenced" pipeline separately, and never split one deal's credit into fractions.

What is a human-intervention rate and what level is acceptable?

It's the share of the AI's reviewed actions that a human edited, rejected or reversed. No universal acceptable level exists. An illustrative gate is fewer than one in five drafts changed by the end of the pilot, but your own trend matters more. Falling is good. Zero is a warning that nobody's reading. Missed opt-outs must be zero at any level.

How many meetings should an AI SDR book per month?

It depends on arithmetic you can do yourself: qualified accounts reached × positive reply rate × booking rate × show rate. With the made-up inputs above, 1,000 × 3% × 40% × 75% gives 9 meetings held. Change any input and the answer moves, which is why a quoted industry figure is worthless.

Which AI SDR metrics should be reviewed weekly versus monthly?

Weekly: the failure metrics (off-ICP sends, personalization errors, intervention rate) and reputation (complaint rate per domain, bounces, placement), which catch damage while it's cheap to fix. Monthly: positive reply rate, meetings held, qualified opportunities, cost per held meeting and pipeline per qualified account. Outcomes are too slow and noisy to read more often.