I Replaced an SDR Seat With an AI SDR for 60 Days. Here's the Real Cost Per Booked Meeting, and the Attribution Problem Nobody Publishes.

The AI SDR booked 11 meetings in its first month for $415 in software. The human SDR seat it was tested against booked 19, for roughly $8,000 all-in that month. The number that actually mattered was neither of those. It was what happened to attribution the week both were emailing the same accounts.

The result, before the setup

Cost per booked meeting over the 60-day window: $46 on the AI SDR stack, $312 on the loaded human SDR seat. That is the headline. It is also misleading on its own, which is why I am writing this instead of a screenshot.

When I corrected for show rate and SQL rate, the gap closed sharply. Cost per SQL was $184 on the AI stack and $397 on the human seat. The AI SDR won, but by 2.2x, not the 6.8x that the raw dashboard implied. The meetings the agent booked were not the meetings the human booked, and pretending otherwise is the trap most "I replaced a person with AI" posts walk into.

The second thing that distorted the numbers, harder to see and harder to fix, was attribution contamination. For roughly two weeks of the test, both the agent and the human were emailing into the same target accounts. Replies came back. Some went to the agent's inbox, some to the human's, some to a generic info@ that I had to route by hand. Both motions claimed credit. The vendor dashboard reported one set of numbers, our CRM reported another, and neither was right until I rebuilt the account split and reran attribution from a holdout group.

This is the part nobody publishes. The cost-per-meeting math is the easy half. The attribution problem is the half that decides whether your AI SDR is real margin or a deferred cost that quietly moved off the headcount line. Industry framing already pegs AI SDR software at a fraction of a human seat's loaded cost, but the framing assumes clean attribution, and clean attribution is exactly what you do not get when an agent and a human work overlapping accounts.

The setup and the loaded-cost baseline

One outbound motion. One ICP: mid-market B2B SaaS, 200 to 1,500 employees, North America. One target account list of 4,200 companies, split 50/50 between the agent and the human SDR with a hard suppression rule between the two lists. 60 days, starting in late summer. Same offer, same calendar link target, same qualification rubric on the inbound side once a reply landed.

The fully loaded human SDR cost I tested against was not a guess. I built it line by line so the comparison was defensible against a finance team:

Annualized, that is just over $140K all-in for one seat, which is consistent with the published industry framing that a human SDR runs north of $60K base with a fully loaded cost that exceeds it materially once you stack the load on top.

The AI SDR side was built the same way, like for like. Software subscription, usage, and the data layer underneath. No hidden headcount, no "we already had this seat" math, no pro-rating that flatters the AI number. I will itemize that next.

The AI SDR stack, every tool and every price

The agent itself was the AI SDR layer. The deliverability and data infrastructure underneath it was a separate stack, because no AI SDR product I have tested ships with enough of either to run alone, despite vendor framing that suggests otherwise. The published definition of an AI SDR covers prospecting, sequencing, and reply handling. In practice, the prospecting layer needs an external data source and the sequencing layer needs an external inbox infrastructure, or you get throttled inside a month.

The agent-layer number lands inside the $3,000 to $6,000 per year published band for AI SDR software. The full stack lands at around $8,400 annualized once you include the data and inbox layer the agent rides on top of. The all-in stack is still under one month of the loaded human seat.

One important scope correction. The cold-outbound motion is one model. The inbound model, an agent that answers and routes inbound site traffic and form fills, is a different product solving a different problem. I tested the outbound variant. Inbound numbers from this piece do not transfer.

The data layer is the part most "I tried an AI SDR" posts skip. The agent is only as good as the rows it sends to. For more detail on that piece, I covered the Apollo, Clay, and Smartlead combination underneath any AI SDR separately. The short version: you cannot evaluate an AI SDR without first deciding whether your data and inbox layer is the bottleneck.

The 30-day deployment playbook

Day 1 to 5: ICP, list, and suppression

Define the ICP tighter than you think you need to. The agent's cost per meeting is wrecked by a loose ICP because the volume amplifies the mistake. Build the account list. Enrich. Build the suppression list: the accounts the human SDR is already working, the accounts in active opportunity, and anything from the last 90 days of inbound. Without this list, the attribution problem starts on day one.

Day 6 to 12: messaging, review gate, deliverability

Write the sequences. Run every variant through a human review gate before any send. I rejected the first three sets of agent-generated openers as too obviously templated – the kind of "I came across your profile" opener that gets a domain blacklisted inside a week. Warm the inboxes for 14 days minimum on new domains, longer if you can wait. Cap the initial daily send at 30 per inbox and ramp.

Day 13 to 21: ramp and hard escalation rule

Push send volume to target. The hard rule, which mattered more than any other operational choice, was that any positive reply triggered a human handoff within minutes, not hours. The agent classified intent. A human owned the next message. This is where most fully autonomous setups go wrong, because the agent's reply to a real buyer reads like an agent's reply to a real buyer, and you spend the trust you spent six emails earning.

Day 22 to 30: kill, tune, lock the holdout

By day 22, you have enough reply data to kill the worst-performing sequence branch. I cut one of three branches that was driving 60% of the negative-sentiment replies. Lock a holdout group of 400 accounts that neither the agent nor the human touches, so you have a clean attribution baseline for the back half of the test.

The one rule that prevented a brand-risk incident: no fully autonomous send to tier-one accounts. The top 5% of the target list went through a human approval queue on every message. The agent drafted, the human pressed send. This is the rule I would not relax for any cost-per-meeting number.

The numbers at day 7, 30, and 60

| Metric | Day 7 | Day 30 | Day 60 | |---|---|---| | Meetings booked, AI | 3 | 11 | 26 | | Meetings booked, human | 5 | 19 | 38 | | Show rate, AI | 67% | 64% | 65% | | Show rate, human | 80% | 79% | 81% | | SQL rate of shows, AI | 50% | 43% | 41% | | SQL rate of shows, human | 70% | 68% | 66% | | Cost per booked meeting, AI | $164 | $64 | $46 | | Cost per booked meeting, human | $1,567 | $618 | $312 | | Cost per SQL, AI | $490 | $231 | $184 | | Cost per SQL, human | $2,791 | $1,148 | $397 |

The AI SDR won on cost per booked meeting at every checkpoint. The AI SDR also won by progressively less on cost per SQL as show rate and qualification were applied, which is the real curve. At day 7 the human seat was 5.7x more expensive per SQL. At day 60, 2.2x.

The raw dashboard the agent vendor surfaced reported 34 meetings booked for the AI side, not 26. The eight-meeting delta came from double-counts where the agent claimed credit for a meeting that the human had already opened, and from "meetings" that were actually replies asking to be removed which the agent's classifier had flagged as positive intent. I rebuilt the count from the CRM, not the vendor dashboard. The corrected number is the one in the table.

The attribution problem nobody publishes

Two weeks into the test, an account from the human SDR's list replied to an agent email. The human had emailed the same person three weeks earlier and had a meeting on the calendar that had no-showed. The reply went to the agent inbox. The agent classified it as positive, the handoff fired, and a new meeting was booked. The CRM logged it as agent-sourced. The vendor dashboard counted it. The human SDR, reasonably, flagged it as theirs.

Three things were happening at once. The agent and the human were emailing overlapping accounts despite the suppression rule, because the suppression rule covered active opportunities but not stale ones. Last-touch attribution inflated the agent's number because the agent emailed second. The vendor dashboard inflated it further because every positive-intent classification was logged as a "sourced meeting" before the meeting actually happened.

The fix, which took a full week and which I would build before the test next time:

Split the account list with a hard fence, not a soft suppression. The agent's 2,100 accounts are off-limits to the human for the full test window, and vice versa. No exceptions for "warm" accounts, no "the human already had a relationship" carve-outs. The contamination starts with the carve-outs.

Lock a holdout group of accounts neither motion touches. Without a holdout, you cannot tell whether the meetings you booked were caused by the outreach or were going to happen anyway from inbound and existing pipeline. The holdout is small and expensive in foregone pipeline, but it is the only thing that lets you say a number out loud with confidence.

Use first-meaningful-reply as the unit, not "meeting booked." A meeting booked is downstream of three or four reportable events and gets contaminated. A first-meaningful-reply is an event with a timestamp on a specific email from a specific sender, and you can attribute it cleanly to one motion.

The general principle: never run an AI SDR and a human SDR against the same list without a hard split and a holdout, or the numbers are unrecoverable, and you will spend two weeks of the test explaining the discrepancy to a sales lead who, correctly, does not trust the agent dashboard.

What the AI SDR could not do

Multi-threading a buying committee after the first reply. When a director replied with interest, a human SDR pulled in the VP and the IT lead within the same thread by day three. The agent kept emailing the director. Five of the eight deals that progressed past first meeting did so because a human added the second and third stakeholder, and the agent has no model of buying committee that survives contact with a real org chart.

Reading timing on a one-line reply. "Reach out next quarter" went two ways in the test data. From a director with budget, it meant October. From a manager without budget, it meant never. The human SDR caught the difference inside ten seconds on a Slack flag. The agent treated both the same and re-engaged both in 90 days, which burned the second relationship.

Objection handling beyond the scripted branch. The agent had three branches for three common objections. The fourth objection, which came up roughly 15% of the time, got a generic response that read as evasive and ended the thread.

Brand risk. One agent message went out with a competitor's product name in the offer line because the personalization layer pulled the wrong field. It went to 14 accounts. Two flagged it, one was a former customer. The cost of that incident does not show up in cost per meeting and it should.

The repeatable decision rule

Replace the seat with an AI SDR stack when three conditions hold. Deal size is under roughly $25K ACV, where the unit economics will not support a human SDR's cost per SQL even at decent conversion. List volume is high enough – north of 3,000 net-new target accounts a quarter – that the agent's throughput advantage compounds. Motion is top-of-funnel only, with a separate human team owning multi-threading and qualification after the first reply.

Augment with an AI SDR, do not replace, when deal size is above $50K ACV, when the buying committee is three people or more by default, or when a single misfire to a named account is a brand-risk event you would have to explain to a CEO. In that range, the AI SDR adds capacity, which is real but is not "we cut a seat."

The middle, deal size $25K to $50K, is where most teams sit, and it is where the decision is hard. The right move there is not the vendor demo. It is a 60-day split test with a hard fence and a holdout, built the way I have described, run before you cut anyone's seat. Most of the cost-per-meeting wins I have seen in vendor case studies disappear under a clean attribution rebuild. Some survive. You will not know which without running it.

FAQ

What is an AI SDR?

Software that runs the prospecting, sequencing, and reply-handling tasks a human SDR does. Scope varies sharply by vendor and by motion; cold outbound, inbound answering, and voice are different products with different unit economics.

Are SDRs being replaced by AI?

At most deal sizes and volumes, AI is changing the SDR job more than removing it. The repetitive layer – list research, first-draft sequences, reply classification – is being automated. Multi-threading a buying committee and reading intent nuance on a one-line reply are not, at least not yet.

How much does an AI SDR cost?

The agent software typically lands in the $3,000 to $6,000 per year band, with usage pricing pushing the real number higher. A fully loaded human SDR seat exceeds $140K annualized once base, variable, payroll load, tooling, ramp, and management overhead are included.

What's the best AI SDR?

Depends on the motion. Cold outbound, inbound answering, and voice agents are different products. The right pick is keyed to your deal size, list volume, and whether your bottleneck is the agent layer or the data and inbox layer underneath it.