What is a voice agent for B2B sales?

A voice agent is a fully automated phone caller that holds a real voice conversation with prospects, not a pre-recorded message. It opens in English, can detect a Ukrainian or Russian accent and switch languages mid-sentence, asks qualifying questions, handles pushback, leaves personalized voicemails on missed calls, and books meetings when the prospect is ready. It runs 24 hours a day without breaks.

Does a voice agent actually work for B2B sales?

Yes, with realistic expectations. After 90 days on OD Granite, the agent makes 160 to 200 dials per day, achieves a 41% connect rate, holds conversations over 30 seconds with 22% of connects, and books 6 to 10 qualified meetings per week. The agent books meetings. Humans still close. Voice agents close a much smaller percentage of demos than a trained human sales rep.

What objections does a voice agent handle?

The agent handles standard cold call objections: not interested, too busy, already have a supplier, call me later. It also qualifies budget early, which most human SDRs avoid because it feels uncomfortable. Roughly 40% of connects get disqualified in the first minute on budget, which saves time on both sides. Technical product questions get escalated to a human via SMS within minutes.

How much does building a voice agent cost?

Platform options include Retell, Vapi, and Bland, all priced by call minutes. At 160 to 200 calls per day, the monthly cost is around $250 all in (platform, voice synthesis, and reasoning). Setup takes about 60 hours of work, not one weekend as vendor marketing often claims. As a service, a Fractional AI Officer builds and deploys a voice agent for a B2B business in two to eight weeks as part of a sprint engagement.

When is a voice agent the wrong answer for B2B?

Three situations: your product requires detailed on-call customization that changes per prospect, your market is highly relationship-driven where buyers will hold AI use against you, or you have no way to measure conversion and would be flying blind. Also not recommended if you expect to set it up and ignore it. The system takes two to four hours per week of maintenance to stay calibrated.

Voice Agent for B2B Sales: 90-Day Production Log (2026)

May 7, 2026 · 11 min read · by Dmytro Negodiuk

I run a small Ukrainian-granite distribution business called OD Granite in the US. 90 days ago I replaced my planned SDR hires with a voice agent. This is the uncensored operator log. What worked. What broke. What I changed. What I still cannot automate.

Written for the person deciding whether to pay a platform $200-$500 a month or hire a $60K SDR. I made the same decision 90 days ago. Here is what the data says on the other side.

The setup

Platform: Retell for call orchestration. Vapi evaluated, lost on cost + voice quality at the time (Jan 2026).
Voice synthesis: ElevenLabs for Ukrainian and Russian. Retell native for English.
Reasoning: Claude Sonnet via API for conversation logic. Real-time inference, sub-second latency.
Target list: 2,400 US commercial contractors and property managers scraped from public permit data, enriched with firmographics.
Total monthly cost: ~$250/month all in (platform, voice, reasoning, and calling minutes).
What I was going to spend on 2 SDRs: $120K/yr base + ~$30K overhead = $150K fully loaded.

The actual 90-day numbers

Metric	Day 0-30	Day 31-60	Day 61-90
Dials per day	80-120	140-180	160-200
Connect rate	32%	38%	41%
Conversations over 30 seconds	14%	19%	22%
Meetings booked / week	3-5	5-8	6-10
Meetings that actually happened	2-4	4-6	5-8
Qualified pipeline created	$8K	$24K	$42K
Breaks requiring fix	7	5	3

The learning curve is real. First 30 days the agent was mediocre. Day 60 onwards it's consistent enough that I can plan around it.

What worked from day one

Language switching

The biggest unexpected win. Ukrainian granite buyers in the US are often Ukrainian or Russian-speaking contractors themselves. When the agent opens in English, detects a Ukrainian accent, and switches mid-sentence, the conversation depth triples. Close rate on Ukrainian-language calls is about 4x English-language calls.

Budget qualification

Second unexpected win. The agent asks "what's your typical material budget on a project like this" early. Most SDR scripts avoid this because it feels rude. The agent doesn't care. Rude is fine at scale. We disqualify roughly 40% of connects in the first minute on budget mismatch, which saves everybody time.

Voicemail strategy

Agent leaves voicemails on unanswered calls. Short, specific, name-dropped a likely recent permit. Voicemail-triggered callbacks account for about 12% of our booked meetings. Nobody tells you about this in voice agent marketing.

What broke

Breaks requiring fixes, by type (90 days total)

Model update shifted voice persona: 5 times. Sonnet minor version rolls, conversation style drifts, prospects notice. Fix: re-tune opening prompt, takes 20 min.
Platform API token revoked: 3 times. Retell rotates creds on schedule, my cron doesn't catch immediately. Fix: re-auth script, takes 5 min, but calls lose ~2 hours before I notice.
ElevenLabs voice model deprecated: 1 time. The Ukrainian voice I was using got replaced with a newer one. Fix: swap voice ID, re-record IVR intro, 30 min.
Phone number flagged as spam likely: 2 times. Carrier-level spam score changed. Fix: rotate number, takes 10 min via Twilio.
Cron missed because VPS rebooted: 4 times. My $8/month Hetzner box gets security patches, reboots, cron jobs don't auto-restart. Fix: systemd + monitoring, one-time 2-hour setup.

Total time fixing breaks across 90 days: about 14 hours. That's ~9 min per day average. For context, one US-based SDR would cost $60K/yr plus the time I'd spend managing them, coaching them, replacing them when they quit.

What I cannot automate (and do not want to)

Real closes. The agent books meetings. I show up to meetings. Voice agents close 5-10% of demos. Humans close 30-40%. Not a model problem, a trust problem. Humans buy from humans.
Technical product questions. Contractors ask detailed stone-spec questions. The agent doesn't know our inventory deep enough yet. When asked, it hands off to me via SMS. I respond in 2-3 min. Working great.
Referral pitches. When a contractor says "I know another contractor who'd be interested," the agent should ask for the intro. It doesn't. I rewrite this prompt monthly and it still isn't natural. I suspect this requires human flex.
Complaint handling. One contractor called back upset about a misrouted call. The agent tried to apologize and escalate. It was fine, but a human would've been better. Routing all returned / angry calls to me manually now.

The platform comparison (Retell vs Vapi vs Bland)

I tested all three in evaluation. Retell won for my use case in Jan 2026. Your mileage may vary.

Retell: Best conversation latency. Easy Claude integration. Cost at my volume: ~$120/mo.
Vapi: Better phone number handling. Slightly more complex for cross-language. Cost similar.
Bland: Fastest setup, best docs. Cost was higher at my call volume.

All three will work for most B2B use cases. Pick based on your existing stack integrations.

Setup time, honestly

The "one weekend" setup timeline is a lie. Real timeline:

Day 1-2: Platform signup, phone number provisioning, basic English conversation flow. ~8 hours.
Day 3-4: Ukrainian / Russian voice synthesis, accent detection, language switching. ~6 hours.
Day 5-7: Prompt engineering for the actual sales conversation. This is where 70% of the real work is. ~20 hours.
Day 8-14: Watching the first 100 calls, rewriting prompts, fixing edge cases. ~15 hours.
Day 15-30: Calibrating qualification questions, voicemail variants, handoff triggers. ~10 hours.

Total hours to working agent: ~60. That's a full working week + 20. Not a weekend.

On the upside: once it's running, it takes 2-4 hours a week to maintain.

Break-even math

One SDR in NYC fully loaded: $150K/yr = $2,885/week.

One voice agent: $250/month = $58/week.

Setup opportunity cost: ~60 hours of founder time at imputed $150/hr = $9,000 one-time.

Break-even point: Week 4. By Week 6 the voice agent has paid for itself vs the SDR alternative.

When a voice agent is the wrong answer

Your product needs detailed on-call customization that changes per prospect. Stick with humans.
Your market is extremely relationship-driven (enterprise SaaS, BigPharma). Buyers will sniff out AI and hold it against you.
You have no way to measure conversion. Flying blind. Don't deploy.
You expect "set and forget." You will spend 2-4 hours a week on this. Budget for it or it decays.

What I'd do differently

Start with ONE language. I tried three from day one. Wasted 2 weeks on Ukrainian + Russian voice tuning I could've done after proving English worked.
Build the handoff-to-human SMS workflow BEFORE the agent is live. I did it in week 2, should've been week 0.
Record a dozen of my own best cold calls and use them as prompt examples. I wrote prompts from scratch. Waste.
Set up cron monitoring from day zero. The 4 cron-missed incidents were preventable.

Considering a voice agent for your B2B business?

I deploy these for $1M-$10M SMBs. 2-8 weeks from audit to live. Every setup is field-tested in my own businesses first. Book a 30-min call below.

Book a 30-minute audit

More on the Fractional AI Officer model · Fractional AI Officer for distributors · The $600 operator stack