“We were already on our third iteration of the SDR team. Familiar helped us scale outbound without additional headcount, while delivering outcomes in line with our top rep.” · archit jain, co-founder, polyborg
Across 357 categorized live conversations between Familiar and a human sales rep, we observed three useful signals:
- similar rates of disqualification or “not right now” · Familiar 18.9%, Human 20.7%
- similar booking rates · Familiar 1.2%, Human 0.5%
- gaps in lead targeting that Polyborg could address · about half of calls reached a wrong prospect
Familiar delivered these outcomes at roughly 20x lower cost per live conversation.
We ran the comparison with Polyborg, which helps US manufacturers automate palletizing and case packing through robotics-as-a-service. Phone calls help Polyborg reach manufacturing decision-makers who are difficult to engage through email or LinkedIn.
For each Human–Familiar comparison, we looked at whether the observed gap fell within the approximate 95% range we would expect at this sample size if Human and Familiar were performing similarly.
Similar disqualification outcomes
For this comparison, we combined conversations ending in “disqualified” or “not right now.” These prospects broadly fit the target profile but weren't opportunities worth pursuing today. They weren't interested, were too small or weren't ready yet.
within expected range
gap 1.8 pp · expected ±8.3 pp
These conversations helped Polyborg decide which prospects to set aside and which to revisit later.
Similar booking rates
within expected range
gap 0.7 pp · expected ±1.9 pp
Familiar booked more meetings than the human rep in this sample. It asked qualifying questions, communicated Polyborg's value proposition and handled the back-and-forth needed to agree on a suitable time.
The calls also helped improve targeting
The trial also exposed a targeting problem. Some businesses looked relevant on paper but weren't a fit. Retailers and transportation companies, for example, operated in adjacent industries, making them difficult to exclude through list filtering alone. The calls established that they lacked the manufacturing workflows Polyborg automates. We tracked these separately as wrong prospects.
gap larger than chance
gap 11.7 pp · expected ±10.4 pp
Polyborg used that feedback to quickly refine its targeting and focus on businesses with relevant manufacturing needs.
Discovering these mismatches through human reps carries an opportunity cost. Time spent screening poor-fit businesses is time unavailable for promising prospects. Familiar offers a way to expand that screening while reducing the time required from the sales team.
Better screening → better targeting → more time for real opportunities
The job is the benchmark
Current voice benchmarks help measure capabilities such as tool-call reliability, turn-taking and interruption handling. Those measurements alone don't establish whether a model can qualify a prospect or book a meeting in a live conversation.
Live qualification calls are messy from the first few seconds. Prospects interrupt the intro, ask who is calling, speak over the agent, answer from noisy environments or change direction mid-sentence. Getting through that opening is often the difference between reaching a useful outcome and losing the call.
Familiar is built for these conditions. It is a full-duplex speech model that listens while it speaks, allowing it to naturally handle interruptions and overlapping speech. It is also trained on real sales calls, so it learns how to establish relevance, handle objections and ask the questions that determine whether a prospect is a fit.
That is why we benchmark it on the job itself.
Where Familiar failed
Familiar made identifiable model errors in 13.6% of live conversations. We tracked those errors separately to understand which sales outcomes they affected.
Correcting those errors would not have changed the booking rate in this sample, but it would have improved several other outcomes.
errors fixed vs human
within expected range
gap 4.1 pp · expected ±8.7 pp
errors fixed vs human
within expected range
gap 5.2 pp · expected ±10.4 pp
Adjusted Familiar results are counterfactual estimates, not measured production results.
We also looked at early hang-ups, combining calls that ended after the pitch or during identity verification. Familiar's rate fell outside the expected range at this sample size.
gap larger than chance
gap 8.6 pp · expected ±6.8 pp
These are the failure modes we are prioritizing in the next model iteration.
The economics of expanding qualification
Expanding a human qualification team means more hiring, onboarding and training. Turnover means repeating that work. There is also a practical limit to how many calls a human team can make during the hours when prospects are most likely to answer.
Familiar lets teams expand qualification capacity without increasing headcount at the same rate. More calls can be placed during those high-value windows without requiring a larger team.
Teams like Polyborg can review Familiar's calls and give feedback just as they would to a human rep. That feedback gets incorporated through post-training and fine-tuning, improving future calls for the business while also strengthening the underlying model for other customers.
human sdr
$3.50–4.25
familiar
$0.05–0.30
What comes next
The Polyborg trial produced similar qualification rates, booked meetings and actionable feedback on targeting.
Our goal is to deliver human-level sales outcomes while making qualification cheaper and easier to scale.
Try Familiar on a batch of prospects and compare the results.
All percentages are shares of categorized live calls, and they describe how calls ended, not detection accuracy. Each expected range is the largest gap chance alone would produce at this sample size if Human and Familiar performed the same.