AI cold calling in 2026: the benchmark report no one else will publish

Every vendor pitch in 2026 includes an AI cold-calling demo that sounds indistinguishable from a human. The reality in production is messier. Connect rates have been declining for three years as call recipients adapt. Spam-label risk is real, and it scales with dial volume. And the booking-rate headline numbers — the ones on the pitch decks — almost always cherry-pick a single vertical and a single best week. We're publishing the numbers we actually see across 140 active Signal deployments running AI-assisted outbound in the US.

The actual benchmark numbers

These figures come from production Signal deployments running outbound AI voice campaigns in Q1–Q2 2026. All campaigns used STIR/SHAKEN-attested numbers, verified US origination, and compliant opt-in lists. We've removed the top and bottom decile to strip outliers.

The gap between the AI-only and hybrid columns is the single most important finding in this data. Hybrid deployments — where an AI voice agent qualifies the lead and a human agent closes — book 2.1 to 2.4 times more appointments than fully-automated pipelines running end-to-end on AI. More on why below.

The four failure modes we see most often

Most underperforming deployments fail in one of four places. They're fixable, but only once you can name them.

Why hybrid outperforms AI-only by 2.3×

The intuitive answer — "humans are better at closing" — is incomplete. The more precise explanation has two parts.

First, trust calibration. Prospects who speak to an AI and then get transferred to a human report significantly higher trust in the human rep than prospects who receive a cold call directly from a human. The AI call acts as a pre-qualification signal — by the time the human joins, the prospect has already confirmed interest once, which changes the entire dynamic of the conversation.

Second, cognitive load reduction on the agent. When a human rep takes a warm transfer, they receive a real-time summary: the prospect's confirmed intent, the specific need flagged, and the objections raised during the AI call. Reps working with these summaries close at 2.1× the rate of reps taking cold calls, even on the same list. The information advantage compounds over a shift — reps in hybrid programs report 40% less mental fatigue by end-of-day.

STIR/SHAKEN and spam-label risk at scale

AI dialers that operate at volume create a specific risk that manual outbound teams rarely encounter: algorithmic spam scoring. Here's what's happening at the carrier layer.

Tier-1 US carriers (AT&T, T-Mobile, Verizon) run behavioral reputation models that score numbers continuously — not just at registration time. The signals they weight most heavily in 2026 are call-back rate (do recipients call the number back?), call duration distribution (are calls suspiciously short or uniform in length?), and origination velocity (how many unique ANIs are originating from the same trunk?).

An AI dialer running at 500 dials/hour from a tight number pool will trigger velocity alerts within 3–4 hours of a campaign launch. The mitigation stack that works in production:

1. STIR/SHAKEN full attestation (A-level). Never operate on B or C attestation on outbound campaign numbers. A-attestation costs nothing extra with carriers who verify origination — and unattested or partially-attested numbers are scored 40–60% more aggressively by carrier reputation models.
2. CNAM registration per DID. Numbers with a registered caller name see 12–18% higher answer rates and are downgraded by reputation models more slowly because call-back rates are naturally higher when recipients recognise the brand.
3. Pacing variance.Uniform inter-dial intervals are a strong spam signal. Introduce ±20% jitter into your dialer's pacing algorithm. It costs negligible throughput and meaningfully reduces velocity scoring.
4. Pool-to-dial ratio.Maintain at minimum 1 number per 80 dials per day. At 500 dials/hour, that's a pool of at least 625 numbers across an 8-hour campaign window.

What the top 10% of deployments do differently

Across 140 deployments, the top decile by booking rate shares four operational patterns that the median deployment doesn't have.

• They instrument call quality, not just outcomes. Every call gets a quality score on five dimensions: audio clarity, AI latency (time-to-first-token), interruption rate, sentiment trajectory, and escalation timing. These scores are reviewed in a weekly cadence and used to update the qualification script.
• They treat the AI persona as a product, not a config.High performers A/B test voice persona, opening line, objection-handling branches, and pacing every two weeks. Low performers set the script in week one and don't touch it.
• They route on intent score, not completion. The qualification script is a floor, not a ceiling. If intent confidence exceeds the threshold mid-script, the best deployments interrupt and route — even if four questions remain unanswered.
• They own their number pool health. Top performers run daily reputation checks on every active DID using third-party reputation APIs, retire numbers proactively at the first yellow flag, and track pool health as a team KPI alongside booking rate.

The ROI math that actually holds up

AI cold calling is not a cost-elimination play — at least not in any deployment that performs well. It's a capacity-expansion play. The math looks like this: a human rep can work 40–50 live conversations per day before quality degrades. An AI agent running in parallel can qualify 300–400 leads in the same window and hand the human rep 20–30 warm transfers instead of cold calls. The rep's close rate on warm transfers is roughly double their cold-call rate. Total bookings per human rep roughly double. You're not replacing the rep — you're giving them a qualification infrastructure they couldn't afford to hire for.

The cost structure that typically pencils: Signal AI minutes at the Professional tier (1,000 minutes/seat/month) cover approximately 300–400 qualification calls, depending on average call duration. At a $49/seat price point, that's roughly $0.12–0.16 per qualified conversation before the lead reaches a human. In insurance and solar — where a booked appointment has a downstream value of $300–$2,000 — the math is unambiguous. In lower-value verticals with thinner margins, the ROI is tighter and the operational discipline described above matters more.