Voice AI vs Philippines BPO for Outbound Calling 2026: The Real Cost Per Minute Math

The invoice says $14,200 for the month. You run outbound collections for a US consumer lender, your BPO partner in Cebu staffs 11 seats on your program, and the line items look reasonable one at a time: $1,150 per seat, a telco recharge for the Telnyx SIP trunk, a QA fee, a small technology surcharge. Then you divide the total by the number of right-party contacts the team actually produced last month, 1,940 of them, and the number stops looking reasonable. You are paying $7.32 per conversation. Not per resolution. Per conversation.

Your BPO account manager will tell you the seat rate is competitive, and it is. Philippines seat rates have been competitive for twenty years. The problem is that the seat rate is the least important number in outbound economics, and almost nobody buying offshore outbound ever works out the numbers that matter: cost per connected minute and cost per outcome. This post works them out, line by line, for three options: a Philippines BPO dialing over SIP, US-based agents, and a managed voice AI platform.

What this post argues

The fully loaded cost of a Philippines BPO seat running US-hours outbound is $1,400 to $2,100 per month once you count night differential, attrition, QA and the management layer, not the $900 to $1,200 sticker rate. At realistic connect rates and dialer utilization, that translates to $0.55 to $0.95 per connected minute. Voice AI platforms land between $0.09 and $0.25 per connected minute, and per-outcome pricing models change the risk allocation entirely. After reading this you should be able to rebuild your own program's cost per outcome in a spreadsheet in under an hour, and know which calls to leave with humans.

Why this math matters more in 2026 than it did in 2023

Three things moved.

Philippine wages stopped being flat. Entry-level CSR pay in Metro Manila crossed PHP 25,000 to PHP 30,000 per month for voice accounts in 2025, and the mandated night shift differential (10 percent under the Philippine Labor Code, and most voice BPOs pay 15 to 20 percent to hold graveyard-shift staff) applies to every hour of a US-hours program. Attrition on outbound voice accounts runs 40 to 60 percent annually, which means you are perpetually paying for recruitment and a training bench you never see on the invoice as a separate line.

US carriers got hostile to bulk outbound. STIR/SHAKEN attestation, carrier-level analytics from the likes of TNS and First Orion, and aggressive spam labeling mean connect rates on cold US consumer lists dropped from the low 20s to 12 to 18 percent for most programs. Every point of connect rate you lose raises your cost per conversation, because the seat costs the same whether the phone gets answered or not.

Voice AI stopped being a demo. Sub-second turn latency on telephony, barge-in handling, and per-outcome commercial models are now table stakes for serious platforms. The question in 2026 is not whether a machine can hold a collections reminder conversation. It can. The question is what each option costs per outcome, and that is a spreadsheet question, not a technology question.

The cost model, end to end

Outbound cost has four layers: labor (or compute), telco, overhead, and yield. Most bad comparisons stop at layer one.

Layer 1: What a Philippines seat actually costs

The rate card says $900 to $1,200 per seat per month for a dedicated voice agent on a US program. Here is what sits behind and on top of that number.

Cost component	Monthly, per productive seat	Notes
Base compensation	$480 to $620	PHP 27,000 to 35,000, voice account, 1 to 2 years experience
Night differential + allowances	$90 to $160	US hours are 9pm to 6am in Manila
Statutory benefits, 13th month	$80 to $120	SSS, PhilHealth, Pag-IBIG, 13th month pay accrued
Facility, seat, IT	$150 to $250	Office, redundancy power, dual ISP, licenses
Team lead + QA allocation	$110 to $180	1 TL per 12 to 15 agents, 1 QA per 20 to 25
Recruitment + training bench	$90 to $170	40 to 60 percent annual attrition amortized
BPO margin	$180 to $350	18 to 25 percent typical on managed seats
Fully loaded	$1,400 to $2,100	Against a $900 to $1,200 sticker rate

A US-based agent doing the same work costs $4,800 to $7,500 per month fully loaded ($18 to $24 per hour base, plus benefits, facilities or remote-work stipends, and management). Nobody disputes that offshore wins on layer one. The gap narrows in the layers below.

Layer 2: Telco, the part everyone asks about and the part that matters least

If you found this post searching for Telnyx Philippines outbound rates, here is the short version. Termination cost is real money at volume, but it is a rounding error next to labor.

Route	Telnyx (published range)	Twilio (published range)
Outbound to US (from anywhere, via SIP)	$0.0050 to $0.0070/min	$0.0130 to $0.0140/min
Outbound to Philippines mobile	$0.13 to $0.18/min	$0.15 to $0.21/min
Outbound to India	$0.010 to $0.025/min	$0.013 to $0.035/min

Two things follow. First, a Philippines BPO calling US consumers over a Telnyx SIP trunk pays US termination, not Philippines termination. The agents sit in Cebu; the calls terminate in Ohio. At $0.006 per minute and 3 connected minutes per conversation, telco is under 2 cents per conversation. Second, this is why comparing platforms on carrier rates misses the point. We covered the deeper platform trade-offs in voice AI vs Twilio Voice for US contact centers; the summary is that CPaaS pricing is the cheapest line on the bill and the most expensive thing to build on top of.

The exception: if you are calling into the Philippines or other high-termination markets (collections for OFW remittance products, for example), termination at $0.15 per minute genuinely changes the math and pushes you toward shorter, more scripted first-pass calls, which is exactly where voice AI is strongest.

Layer 3: Connect rate and utilization, where the money actually leaks

An agent on an 8-hour shift does not talk for 8 hours. On a well-run predictive dialer, expect 30 to 35 minutes of connected talk time per hour. On preview or progressive dialing (which TCPA risk pushes many US lenders toward), 18 to 25 minutes. Take breaks, coaching, after-call work and system downtime out, and a Philippines seat delivers roughly 90 to 130 connected hours per month.

Now the formula that should be on your whiteboard:

Cost per connected minute = fully loaded monthly seat cost ÷ (connected hours per month × 60)

Run it: $1,700 fully loaded ÷ (110 connected hours × 60 minutes) = $0.26 per connected minute at the dialer-optimized best case. Run it at the preview-dialing, compliance-constrained case: $1,700 ÷ (75 × 60) = $0.38 per connected minute. Add dialer licensing ($100 to $180 per seat per month for the usual suspects), list scrubbing, and the telco trickle, and real-world programs land at $0.30 to $0.55 per connected minute offshore and $1.10 to $1.90 for US agents.

Voice AI platforms charge one of two ways. Per-minute pricing runs $0.06 to $0.20 per connected minute at volume depending on market and stack (we published the India-market version of this math at voice AI pricing in India; US pricing runs 1.5 to 2.5 times higher per minute but follows the same structure). Per-outcome pricing charges per completed verification, per promise-to-pay, per qualified transfer. The crucial structural difference: a voice AI platform has no idle seat cost. You pay for connected minutes only. The 82 to 88 percent of dials that go unanswered cost you fractions of a cent in telco, not agent time.

Layer 4: Yield, or why cost per minute is still not the real number

The number your CFO cares about is cost per outcome. A connected minute that produces nothing is waste at any price. So the last step is:

Cost per outcome = cost per connected minute × average handle minutes ÷ outcome rate

A human agent converts connected collections calls to promises-to-pay at, say, 32 percent with a 4-minute AHT. A well-built voice AI agent on the same first-pass reminder work converts at 22 to 28 percent with a 2.5-minute AHT, because it wastes no time on small talk and never deviates from the compliant script. Lower conversion, but far lower cost and shorter calls. The math decides, not the intuition.

What goes wrong when people run this comparison

Five recurring mistakes, from deals we have seen evaluated.

Comparing sticker seat rate to voice AI list price. The $1,000 seat against $0.15 per minute. Both numbers are wrong in opposite directions: the seat is understated by 40 to 70 percent (layer 1 above) and the per-minute price usually negotiates down 20 to 40 percent at committed volume.

Assuming human connect rates transfer to AI. They do not transfer; they improve. Voice AI dials with clean, consistent caller ID reputation management and can legally attempt tighter recall windows because attempts cost nearly nothing. Programs typically see 2 to 4 points more right-party contact simply from attempt density humans cannot afford.

Ignoring the supervision cost of AI. Voice AI is not zero-labor. Budget 0.5 to 1 FTE per 100,000 monthly calls for prompt and flow maintenance, QA sampling and escalation handling. It is a tenth of the equivalent human management layer, but it is not zero, and vendors who tell you it is zero are selling a demo.

Measuring AHT as if shorter is worse. BPO contracts historically priced by the minute, which quietly rewards longer calls. When you move to per-outcome pricing the incentive flips, and 2.5-minute calls that resolve beat 4-minute calls that meander. Watch this in vendor pilots: a vendor paid per minute has no reason to shorten your calls.

Forgetting the floor on human quality at 3am Manila time. Attrition and graveyard fatigue show up as QA variance. The 90th percentile Filipino agent is excellent; the program is priced on the average, and the average at hour seven of a night shift is not the demo agent you met during vendor selection.

The worked example: 10,000 outbound calls a day

A US lender runs 10,000 dials a day, 22 days a month, on early-stage delinquency reminders. Connect rate 15 percent, so 1,500 conversations a day, 33,000 a month. Average handle time 3.5 minutes human, 2.5 minutes AI. Target outcome: promise-to-pay.

	Philippines BPO	US in-house	Managed voice AI
Capacity needed	42 seats	42 seats	Elastic
Fully loaded monthly cost	$71,000 (42 × $1,700)	$260,000 (42 × $6,200)	$18,600 (82,500 min × $0.20 + platform fee)
Cost per connected minute	$0.61	$2.25	$0.23
Outcome rate (PTP per conversation)	32%	34%	25%
Outcomes per month	10,560	11,220	8,250
Cost per outcome	$6.72	$23.17	$2.25

Three honest observations about this table. The BPO cost per connected minute came out higher than the earlier range because 42 seats sized for peak hour sit partially idle off-peak; elasticity is a real cost humans carry and machines do not. The AI outcome rate is deliberately conservative; mature programs close the gap to within 4 to 5 points of humans on scripted first-pass work. And the AI column produces 2,310 fewer outcomes per month, which matters if list exhaustion is your constraint rather than budget.

Which is why almost nobody sane runs a pure swap. The hybrid version: voice AI takes every first and second attempt, humans take flagged accounts, disputes, and balances above a threshold. In this example, AI handles 78 percent of conversations at $2.25 per outcome, a reduced 12-seat human pod handles the rest at roughly $7.10 per outcome, blended cost per outcome lands near $3.30, and total monthly spend drops from $71,000 to about $41,000 while outcomes stay within 3 percent of baseline. That is the shape of the deals actually closing in 2026. The organizational side of that transition, retraining, redeployment, contract restructuring, is its own topic, and we wrote the BPO to voice AI migration playbook for it.

When the Philippines BPO still wins

Take the vendor-skeptic view seriously, including of us. There are programs where offshore humans remain the right answer.

• High-stakes, high-variance conversations. Hardship negotiations, save-desk retention, anything where the counterparty cries or negotiates in ways a flow designer did not anticipate. Route these to people, always.
• Low volume. Under roughly 500 conversations a day, platform minimums and the supervision half-FTE eat the AI advantage. A 4-seat pod is simpler.
• Deep product complexity with fast-changing offers. If your script changes weekly and your CRM data is dirty, humans paper over the gaps. AI exposes them. (Eventually you want them exposed; you may not want it this quarter.)
• Outbound where the list is tiny and precious. 200 warm enterprise leads deserve a skilled SDR, not an optimization on cost per minute. AI qualification belongs on wide funnels, the pattern we describe in lead qualification and follow-up.

If your program is high-volume, scripted, compliance-sensitive first-pass work (reminders, confirmations, verifications, qualification), the math above will not rescue the seat model.

Compliance: the trade-offs are not symmetric

TCPA exposure is the number that can delete every saving in this post: statutory damages run $500 to $1,500 per call. Three asymmetries between the options.

Consent and dialer classification. Human agents on preview dial sidestep some ATDS arguments; AI outbound at scale must get consent architecture right from day one, including revocation handling under the FCC's 2024-25 consent rules. A serious platform ships this as workflow, not as your problem. The full treatment is in our guide to TCPA-compliant AI calling for US enterprises.

Disclosure. Several states, and the FCC's direction of travel on AI-generated voice under the TSR and TCPA, require disclosing that the caller is artificial. Build the disclosure into the greeting. Programs that A/B tested honest disclosure found completion-rate impact of 2 to 5 points, far cheaper than the alternative.

Consistency as a compliance asset. A human agent under monthly quota pressure improvises. Recordings from hour seven of a Manila night shift are where regulators find mini-Miranda violations and unauthorized settlement offers. An AI agent says exactly what it is configured to say on call one and call ninety thousand. In examinations, that uniformity is worth real money, though it cuts both ways: a configured mistake also repeats ninety thousand times, which is why release discipline on flows matters as much as code review.

Recording consent (two-party states), DNC scrub cadence, and offshore data residency (your BPO holds US consumer PII in the Philippines; your AI platform should let you keep processing in-region) round out the checklist. If you operate across markets, region-specific stacks differ enough that we maintain a separate overview of global deployments.

The 60-day playbook to find your own number

Do not migrate. Measure first.

Weeks 1 to 2: rebuild the baseline. Pull three months of BPO invoices and dialer logs. Compute fully loaded cost per seat (add the hidden layers from the table above), connected minutes per seat, and cost per outcome by call type. Most ops directors find their true cost per outcome is 1.6 to 2.2 times what they believed.

Weeks 3 to 4: segment the call mix. Tag every call type as scripted-first-pass, judgment-heavy, or relationship. Typical outbound programs are 60 to 80 percent first-pass by volume.

Weeks 5 to 8: pilot AI on one first-pass segment. One use case, one list segment, A/B split against the BPO on the same list. Insist on measuring cost per outcome, not cost per minute, and hold both channels to the same compliance QA sample. Negotiate pilot pricing with a volume-committed production rate attached, so the pilot price is not a bait rate.

Weeks 9 onward: rebalance, do not rip. Move first-pass volume to AI in 20 percent increments, shrink seats by attrition rather than termination (at 40 percent annual attrition, the seat count falls fast without a single hard conversation), and renegotiate the BPO contract around the judgment-heavy work that remains. Your BPO partner would rather keep 15 high-value seats than lose 42.

What changes in the next 12 months

Per-outcome pricing spreads from collections into verification and qualification, shifting connect-rate risk from buyer to platform, and BPOs respond by reselling voice AI under their own brand with a services margin on top (several large Philippine providers already do; ask whose platform is under the hood and what the pass-through markup is). US termination rates stay flat, but carrier spam analytics keep tightening, which favors whoever manages number reputation programmatically. And Philippine wage inflation plus the 2025-26 push on AI upskilling means the offshore seat gets 6 to 10 percent more expensive annually while per-minute AI pricing continues drifting down. Every quarter you delay running this math, the spread widens in one direction.

Bottom line

The seat rate is a decoy. Price outbound in cost per connected minute, then in cost per outcome, and the 2026 numbers come out roughly: Philippines BPO at $0.30 to $0.61 per connected minute and $5 to $9 per outcome, US agents at 3 to 4 times that, managed voice AI at $0.09 to $0.25 per connected minute and $2 to $4 per outcome on scripted first-pass work. Humans still win judgment-heavy conversations and small precious lists. The winning architecture is not a swap, it is a split: machines take the wide, repetitive top of the funnel; a smaller, better-paid human team takes the conversations that deserve one.