What changed
Two things moved this year, and they pull in opposite directions.
The first is that the voice-agent platforms got cheap to start and honest about it. Vapi's public pricing page now says the quiet part out loud: you pay Vapi a $0.05 per minute hosting fee, and every model and provider underneath is passed through at cost with no markup. Their own calculator, for 1,000 minutes a month, lands at $82 to $129 all in: $50 of hosting, about $10 of Deepgram transcription, $8 to $45 of OpenAI, and $15 to $24 of ElevenLabs voice. That is $0.082 to $0.129 per minute, not $0.05.
ElevenLabs publishes the same structure with different words. An additional call minute on the Agents platform is $0.080, burst minutes are $0.160, and the page states plainly that the LLM and the telephony provider are billed separately, at cost. CloudZero's September 18 breakdown of ElevenLabs pricing puts the agent rate at $0.08 per minute on every tier, from the free plan up to the $990 Business plan, and quotes PxlPeak founder John V. Akgul saying that "conversation minutes dwarf TTS consumption inside a month" on a $303 January invoice covering six clients.
So the headline number is real, and it is also about half of what you will actually pay. Three meters run at once: the orchestration platform, the model, and the phone line.
The second thing that moved is the helpdesk side. Zendesk began an expanded-access rollout of its AI agent capabilities in May 2026, pushing the features that used to be sold as "AI agents – Advanced" out across Suite and Support plans. That includes an integration builder for third-party API calls mid-conversation and authorised actions the agent can take in connected systems, including looking up an order and triggering a refund. The useful warning in eesel's teardown: access to the capability and the metered resolution usage are two separate bills, and including the former in your plan does not make the latter unlimited. Zendesk's own pricing page confirms the shape, seats from $19 per month, with Voice and Action Builder billed on consumption above plan allowances.
Intercom's pricing page shows the same split a different way: you buy seats on Essential, Advanced or Expert, and then you buy Fin outcomes and messaging channels as usage on top.
Why this matters to a brand at $50K+/month in paid social
Because at that spend level your call volume is mostly one question, and it is the cheapest question in the business to answer.
Helo's 2026 WISMO guide puts order-status enquiries at 30 to 40 percent of total support volume for many ecommerce and logistics operations, rising to 50 percent or higher during peak sales periods, launches and holidays, at a fully loaded cost they put at $5 to $22 or more per ticket. Plivo's roundup of WISMO automation tools cites Gorgias's finding that order-status questions are about 18 percent of support tickets. Those two bands do not agree, and neither is independently audited. Treat them as a range, not a number: somewhere between a fifth and two-fifths of your inbound is a database lookup wearing a phone call.
Here is the part that decides buy versus build. A voice agent that can talk but cannot read your Shopify order record does not deflect a WISMO call. It collects a name and a phone number and creates a ticket, which is the work you already had. The deflection only happens when the agent can authenticate the caller against the order, read the current fulfilment and tracking state, and say a date out loud.
Sthambh's D2C voice playbook lays the stack out in six layers: telephony, speech-to-text, orchestration and LLM, commerce integration, text-to-speech, and analytics. The commerce layer is the one people underestimate. Their note is specific and matches what we see in builds: real-time order status needs webhook subscriptions, not polling, and returns integration has to respect per-SKU policy rules. They also give a latency target of sub-300ms time-to-first-audio, and observe that for WISMO, latency matters more than transcription accuracy.
The three price shapes, side by side
| Approach | Published price | What it is good for | Source of the number |
|---|---|---|---|
| Vertical helpdesk AI (Gorgias, Fin, Siena) | $0.60–$1.27 per resolution (Gorgias), $0.99 per resolution (Fin), $750+/mo plus usage (Siena) | Brands under about 1,000 daily contacts, no AI engineers on staff | Sthambh's vendor table |
| Voice-first platforms (Vapi, Retell, Bland, Ringly) | $0.07–$0.15 per minute plus telephony and LLM | Deep voice control, routing, multi-language | Sthambh's vendor table; Vapi's own calculator lands at $0.082–$0.129 |
| Assembled stack (Twilio + Deepgram + LLM + ElevenLabs + orchestration) | $0.03–$0.08 per minute at scale, excluding engineering | 2,000+ daily contacts, complex catalog, multi-warehouse | Sthambh's vendor table |
Note what the third row excludes. "Excluding engineering" is doing all the work in that sentence, and nobody publishes that figure.
The triggers
Four things flip a brand from off-the-shelf to wired-in. Any one of them is enough.
- Concurrency at peak. Vapi's free tier gives you 4 concurrent calls and ElevenLabs' free tier also gives 4. Vapi's Core package at $29 a month raises that to 10, Pro to 30, and extra lines are $10 per line per month. If your Black Friday hour needs 40 simultaneous calls, you are on an enterprise package whether or not you wanted one.
- Data retention and access control. On Vapi, raw data retention is 14 days with no package, 30 days on Core at $29, and 180 days on Pro. Role-based access control and zero data retention appear at Core and Pro. Pro is priced at 10 percent of your Vapi hosting fee with a $999 per month minimum, which means the governance tier costs roughly twenty times the compute until you are very large.
- Regulated data. Vapi lists HIPAA compliance, including the DPA and BAA, as a $2,000 per month add-on. If you sell supplements, medical devices or anything that touches health claims, price that line before you price minutes.
- Burst. ElevenLabs charges $0.160 per minute for burst calls, exactly double the $0.080 base. Your worst support day is also your most expensive per minute.
If none of those four apply, buy the off-the-shelf thing and stop reading vendor comparison posts.
What to do this month
1. Count your calls by intent for two weeks, before you price anything
Pull two weeks of inbound calls and tag each one: order status, return or exchange, product question, billing, everything else. You are looking for the order-status share. If it is under 15 percent of calls, a phone agent is not your highest-value AI project this quarter, regardless of what the 30 to 40 percent industry figure says. If it is over 30 percent, you have a business case you can compute yourself from your own ticket cost, and you no longer need anyone else's per-ticket estimate.
2. Run the three-meter math on your own volume, not the headline rate
Take your monthly call minutes and multiply by three numbers, separately: the platform fee, the model cost, and the telephony cost. Vapi's calculator is the clearest public worked example, and ElevenLabs' page states outright that LLM and telephony are billed at cost on top. Then add the package tier that your concurrency and retention requirements force you into. A brand modelling $0.05 per minute and landing at $0.12 has not been overcharged. It has been reading one of three bills.
3. Test the order lookup before you test the voice
Every demo you sit through will be a voice demo. The voice is the easy part now. Ask the vendor to look up a real order from your store, mid-call, with a caller who gives a partial email and the wrong order number, and to state the tracking status out loud. Then ask how that lookup is wired: webhook subscription or polling, which Shopify scopes, what happens when the carrier API times out, and who gets paged. Plivo's evaluation criteria for WISMO tools are the right checklist here, ecommerce platform, order management, helpdesk, warehouse and carrier APIs, plus webhooks, and they weight how much work it takes to get a tool into production. Make the vendor answer that on your data, in the demo, or the integration work becomes your problem after signature.
What we could not verify
- Setup hours. No vendor publishes a number, and no source we opened gives an audited implementation time for a Shopify-connected voice agent. Anyone quoting you "live in a weekend" is describing the voice, not the order lookup.
- Intercom's Fin outcome price. Intercom's pricing page confirmed the seats-plus-outcomes structure but the pricing tables did not render on fetch, so the $0.99 per resolution figure here comes from Sthambh's vendor table, not from Intercom.
- The Gorgias 18 percent figure is cited by Plivo, attributed to Gorgias. We could not open the original Gorgias post that carries it. It conflicts with Helo's 30 to 40 percent band, and both are vendor-published.
- Per-resolution prices for Gorgias and Siena come from one third-party table and were not confirmed on the vendors' own pages.
- Zendesk's metered resolution price is not on Zendesk's public pricing page, which points you to sales.
What we would watch next
Whether the helpdesk vendors close the voice gap before the voice vendors close the commerce gap. Zendesk's authorised actions and integration builder are aimed squarely at order lookup and refunds; if per-resolution pricing on voice lands near the chat rate, the reason to assemble your own stack shrinks to concurrency and data control. Watch ElevenLabs' rate card too. CloudZero's September write-up notes the company cut prices during 2026 and that comparison guides still quote the old Scale tier at $330. Re-price on the day you commit, not from a roundup post.