What Kill Criteria Should We Set Before Signing an AI SDR Contract?

The AI SDR argument almost never gets settled on the merits. It gets settled by attrition: two quarters in, the vendor points at volume, the sales leader points at meetings, finance points at the invoice, and nobody can say what would have counted as failure because nobody wrote it down before the signature.

Kill criteria are the fix, and they belong on the order form, not in a slide. A kill criterion is a named metric, a floor, a measurement window, and a consequence. If you cannot write the consequence, you do not have a criterion. You have a hope.

What the data already says before you pilot

Two things are true at the same time, and holding both is what makes a contract survivable.

Hybrid wins on efficiency, not volume. Bridge Group data covered in AI capacity lift, not headcount cut shows outbound volume moving from 1,150 to 7,400 touches while reply rates fell from 4.7 percent to 2.9 percent. The efficiency result sits underneath that: hybrid teams produced 1.9x meetings per dollar, and 2.4x versus human-only. The volume number is not the win. The per-dollar number is.

Time saved is not revenue. The Gartner findings in the AI reinvestment gap are blunt: 4.8 hours per week returned, and 72 percent of it not reinvested into anything that touches revenue. Teams that named where the hours went saw 2.2x to 3.1x the return of teams that did not, and the ROI split ran 25 percent versus 20 percent depending on whether reinvestment was a written decision or an assumption.

And the headcount-cut version of the ROI story is the one to refuse outright. The capacity-lift issue documents Gartner's finding that teams cutting up to 80 percent of a function saw near-zero ROI gap versus teams that did not cut, Forrester's 55 percent regret rate, Gartner's projection that half of organizations that cut will rehire by 2027, and Robert Half's 29 percent who already reopened the roles. A vendor whose business case is fewer humans is selling you a rehire.

Six kill criteria to write into the contract

1. Meetings per dollar, not meetings

The metric is qualified meetings held per fully loaded dollar of AI SDR spend, measured against your human-only baseline from the two quarters prior. Floor: parity by day 60, and 1.5x by day 90. Hybrid teams in the Bridge Group data hit 1.9x, so a floor below 1.5x by the end of the pilot means your implementation is not reaching the observed range. Consequence for a miss: no renewal, no expansion, and the pilot ends at the term. Write the baseline number into the order form so it cannot be renegotiated later.

2. A reply-rate floor, with volume capped

Volume is the easiest thing for an AI SDR to produce and the least informative. Set a reply-rate floor at no worse than 80 percent of your human baseline, and cap total sends so the vendor cannot buy the meeting number with volume. If reply rate drops below the floor for two consecutive measurement periods, sends pause until messaging is reworked. The industry drift from 4.7 percent to 2.9 percent happened because nobody capped anything.

3. Named redeployment of recovered hours

Before the pilot starts, write down where the recovered hours go: named activity, named owner, named measurement. Multi-threading into open opportunities. Second-call preparation. Customer expansion motions. *Anything except the team will have more time.*** The reinvestment-gap data is the whole justification: 72 percent unreinvested, and a 2.2x to 3.1x return for teams that named the destination. If the redeployment plan does not exist on day zero, the criterion fails at day zero and you should not sign yet.

4. Data quality gate, checked before and during

The AI SDR writes into your CRM. Set the pre-condition: contact and account data must pass your hygiene threshold before sequences start. Then set the running condition: bounce rate, duplicate creation rate, and bad-fit contact rate each carry a ceiling. Breach the ceiling twice and the write-back permission is revoked, not just flagged. An AI SDR that degrades your database is a negative-return purchase even when the meeting number looks fine.

5. Downstream conversion, not top-of-funnel

Meetings held is a mid-funnel metric that can be gamed. Track AI-sourced meetings through to opportunity created and to closed-won, measured against human-sourced meetings from the same period. Floor: AI-sourced meeting-to-opportunity conversion within 25 percent of the human-sourced rate by day 90. If AI-sourced meetings convert at half the rate, the per-dollar number is an illusion and the seller time spent on those meetings is a hidden cost.

6. No headcount-reduction clause, in either direction

State it explicitly in the internal business case: this purchase is not justified by removing SDR headcount, and no headcount decision will be made on pilot data inside the term. The evidence for why is already on the record. Near-zero ROI gap for the teams that cut deeply, 55 percent regret, half projected to rehire by 2027, 29 percent already reopening roles. Refusing the headcount story is what keeps the pilot measurable, because the moment the team believes the tool is here to replace them, every input metric you are measuring becomes political.

The 30, 60, 90

Day 30: instrumentation, not results

Baseline locked and signed by RevOps. All six metrics reporting from a single source. Redeployment plan live with named owners. Data quality gate passing. Nothing about the meeting number is judged at day 30, because a 30-day meeting number is mostly a function of list quality and nothing else. What you are checking is whether you can measure at all. If you cannot, that is the first real finding, and it is usually about your CRM rather than the vendor.

Day 60: parity or explain

Meetings per dollar at parity with the human-only baseline. Reply rate at or above the floor. Bounce and duplicate rates inside the ceiling. Recovered hours visibly landing in the named activity, with the activity's own metric moving. A miss here is not automatically a kill: it is a written explanation from the vendor with a specific change and a date. One documented recovery attempt, not three.

Day 90: the decision

1.5x meetings per dollar or better. Downstream conversion within 25 percent of human-sourced. Data quality clean. Reinvestment measurable. If those hold, expand and keep every criterion in force through the expansion. If they do not, the contract ends at the term with no renegotiation window and the recovered budget goes back to the redeployment activity that was already working.

The paragraph to put on the order form

Ask for this, verbatim, before signature. The pilot term is 90 days. Success is defined as qualified meetings per fully loaded dollar at or above 1.5x the documented human-only baseline, reply rate at or above 80 percent of the human baseline, AI-sourced meeting-to-opportunity conversion within 25 percent of human-sourced, and bounce, duplicate, and bad-fit ceilings unbreached. Metrics are reported weekly from a single mutually agreed source. Failure of any criterion at day 90 terminates at term with no auto-renewal and no early-termination fee. Total sends are capped at the agreed number for the term.

A vendor who has hit 1.9x elsewhere will sign something close to that. A vendor who will not sign any version of it is telling you the ranges in their deck do not reproduce in accounts like yours. That refusal is the cheapest piece of diligence available, and it costs one email.

Set the baseline these criteria are measured against with the AI SDR unit-economics worksheet, then work through the AI SDR hub for the evaluation rubric and the pipeline definitions the contract language depends on.