Build1 distinct publisher3 min readPublished
Pickup starts at 48% on fresh numbers and decays toward 20%. Only retry scheduling gets past 70% cumulative connect, and the arithmetic says that costs about three times the dials.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The retry arithmetic is the part of this worth doing by hand. Treat each dial as an independent draw at the fresh-number rate of 48% [2], and two attempts put you at 73% cumulative contact [8]. Do the same at the aged rate of 20% [3], and five attempts reach only 67%; you need six to clear 70% [9]. Same contact target, three times the dial budget [10], and every one of those dials is telephony spend, concurrency capacity and scheduler logic rather than model quality.
Six is a floor, not an estimate. Independence is a generous assumption here: a number that does not answer at 11am on Tuesday is disproportionately a number that will not answer at 11am on Wednesday. The source makes the same point in softer language, saying a retry is not simply another dial and that its timing, frequency and relation to previous attempts affect the odds of eventually reaching the person [13]. Which means the six attempts only work if they land in different windows. Stack them in one slot and you buy the dials without buying the contacts.
Set that against what the dataset says is not broken. User-initiated drop-off came in under 3% [4]. On an aged number at 20% pickup, 100 dials produce 20 conversations and fewer than one hangup by a human who did not want to talk to a machine, against 80 calls that never connected at all: the unanswered outnumber the offended by more than 130 to 1 [11]. That ratio is the ceiling on what a more human-sounding voice can win you. It is not zero, and it is not where the losses are.
Latency is the one place the engineering advice holds without arithmetic. Median response was under a second, with p95 at roughly 2.1 seconds [5], so the tail runs at more than twice the typical case [12]. The listed causes of those spikes include uncommon phrasing, extra network hops and peak concurrency [6]. Note that the last one is downstream of the retry finding: a scheduler that pushes six attempts per lead into the same evening window is manufacturing its own concurrency peak, and paying for it in the tail where conversations break.
The provenance deserves stating plainly. These are one vendor's own production numbers from the DialNexa platform, across more than a million AI-assisted business calls in India [1]. The 20% figure describes "some call categories" [3] and the 70% describes "some campaigns" [7], with no denominators for either subset and no mechanism offered for why aged numbers decay. The 48-to-20 fall is 28 points, a 58% relative decline in first-attempt pickup [16], and a curve that steep almost certainly has a carrier-side cause the post does not name. Read the shape, not the constants: first-attempt pickup is a depleting asset, and the recovery lives in the schedule.
Ranked by verification strength, evidence, and original report placement.
At a 48% per-attempt pickup rate treated as independent draws, two attempts give 72.96% cumulative contact.
At a 20% per-attempt pickup rate treated as independent draws, five attempts give 67.2% cumulative contact and six give 73.8%, so six attempts are the minimum to clear 70%.
Reaching the same 70% cumulative contact target at the aged 20% rate rather than the fresh 48% rate requires about three times as many dials per lead.
On 100 dials to an aged number at 20% pickup, 20 calls are answered, fewer than 0.6 end in a user-initiated hangup and 80 never connect, a ratio of more than 130 unanswered calls per user-initiated drop.
The 95th percentile latency is more than 2.1 times the median, since the median is below one second and p95 is about 2.1 seconds.
The fall from 48% to 20% first-attempt pickup is 28 percentage points, a 58% relative decline.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single first-party disclosure, no methodology
Every quantitative input comes from one dev.to post authored by the operator of the platform being analysed. No date range, campaign count, denominators, segment definitions, raw data or third-party audit is provided, and favourable results are scoped to unnamed 'some campaigns'. What is verifiable is only internal: the arithmetic derived from the stated pickup, latency and drop-off figures is self-consistent, which lifts the score slightly above the floor.
Real production volume, one operator
The disclosure describes live commercial deployment rather than a demo: over a million AI-assisted calls, roughly 84% outbound, with inbound calls averaging about 13 minutes and ~89% objective completion. That is meaningful evidence that AI voice agents are running production Indian call traffic. It remains one vendor's unaudited account with no customer names, seat counts, revenue or retention, so it cannot stand in for category-wide adoption.
Measured tone, overreaching conclusion
The prose is unusually restrained for vendor content — it hedges, distinguishes inbound from outbound, and declines to claim AI voice is universally solved. The overstatement is structural rather than rhetorical: a claim that the dialer rather than the voice is the bottleneck is a causal conclusion drawn from one operator's unaudited aggregates, with best-case retry results and no cost, carrier-filtering or compliance counterweight. The favourable 70% connect figure is selection-scoped while the unfavourable decay is presented as general.
Vendor content marketing on its own data
The author writes from the platform whose production traffic supplies every figure, publishes on a developer community channel, and reaches a conclusion that shifts buyer attention from voice-model quality to dialer orchestration, retry scheduling and telephony operations — precisely the layer a platform of this kind sells. The dataset scale doubles as a credibility claim. There is no disclosed funding, sponsorship or independent review to offset this alignment.
Provenance clear, substance unverifiable
Confidence in this assessment is moderate: the source, its authorship, its incentive structure and its stated numbers are all unambiguous, and the derived arithmetic follows deterministically from those numbers. But with one publisher, one item and no corroboration or contradiction available, judgements about whether the reported pickup decay, latency profile and drop-off rate generalise beyond this platform cannot be firmed up from the supplied material.
build
Three manual interventions in a month, and every guard was working as designed1 distinct publisher
build
Six MariaDB versions, one real difference: the only reason to leave 10.6 is the July 2026 clock1 distinct publisher
build
Force the tool call, then hand Lightsail a long-lived key1 distinct publisher
build
AI-written code fails the same four ways, and every gate you own reports green1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 26, 2026