Contents
Fast answer
For almost every small business, buy first and revisit later. The published rates as of August 2026 make the buy case easy to state: Bland bundles the model, transcription and premium voices into $0.14 a minute with no monthly fee on its Start plan, Retell publishes an itemised all-in range of $0.07 to $0.31, and Vapi charges $0.05 a minute for orchestration with the providers billed separately. On the reference workload of 500 calls a month at three minutes each - 1,500 minutes - that is $210, $105 to $465, and roughly $120 to $200 all-in respectively. Building means assembling transcription, a language model, speech synthesis and telephony yourself, plus turn-taking, interruption handling, latency tuning and carrier work. We are not going to put a dollar figure on that build because it depends entirely on your team, but the honest framing is that the per-minute saving from building is a few tens of dollars a month at this volume and the engineering is weeks. Build when voice is your product, not when voice is your phone line.
Facts last verified 2026-08
Quick facts
- Bland Start
- $0.14/min bundled, no monthly fee - $210 at 1,500 min (as of Aug 2026)
- Retell all-in
- $0.07-$0.31/min itemised - $105-$465 at 1,500 min
- Vapi
- $0.05/min orchestration plus providers - roughly $120-$200 all-in at 1,500 min
- Reference workload
- 500 calls a month at 3 minutes each = 1,500 billable minutes
- The hard part of a build
- Turn-taking, endpointing, interruption handling and streaming latency
- Most underestimated cost
- Maintenance, not the initial build
- Compliance
- Vapi HIPAA $2,000/mo, ZDR $1,000/mo; Retell PII removal $0.01/min
- Build cost figure
- Not published here - it depends on your team and scope, not the technology
On this page
What this page covers
A comparison visitor should understand the tradeoff, the best-fit scenario, and the next diagnostic tool to confirm the choice.
- 01Quick verdict
- 02What buying costs
- 03What building costs
- 04Hidden work
- 05Decision table
- 06Where both are wrong
Why does this matter now?
Build versus buy is where automation budgets get destroyed, usually by optimism about maintenance rather than about the initial build. A voice agent is a live, public system: it fails in front of customers, at the busiest moment, and the failure is audible. The platforms in this category have already solved the parts that are hard to get right and easy to underestimate - endpointing, interruptions, latency across four providers, carrier quirks. Paying for that is not laziness, it is buying the part of the problem that has no business value to solve twice.
Internal path
Where to go next from this page
These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.
What you should be able to do after this
- Price the buy option honestly
- Price your own engineering time
- Spot the workloads that justify a build
- Choose a reversible starting point
How do you do it, step by step?
1. Write down what the agent must do and must never do
Before pricing anything, define the job: which calls it answers, what it books or captures, what it must never say, and when it hands off to a person. This is the part that determines whether the project succeeds, and it is identical whether you build or buy. Most failed voice projects failed here rather than on the platform choice.
2. Price the buy option at your real minutes
Take your monthly minutes and run them against the published rates as of August 2026: Bland Start at $0.14 with no monthly fee, Retell in its $0.07 to $0.31 band depending on model choice, Vapi at $0.05 plus separately billed providers. At 1,500 minutes those come to $210, $105 to $465, and roughly $120 to $200. Write the annual figure down, because that is the number a build has to beat.
3. List the components a build actually needs
Speech-to-text, a language model, text-to-speech, telephony and carrier configuration, plus the orchestration that handles turn-taking, endpointing, interruptions and streaming latency. The first four you can rent. The fifth is the hard part, and it is exactly what platforms in this category sell. Be specific about which of these you intend to own before deciding you are building.
4. Price maintenance, not just the build
The build is a one-off; the maintenance is forever. Model deprecations, carrier changes, voice provider updates, latency regressions and the ongoing prompt tuning that every voice agent needs all land on whoever built it. Estimate hours per month rather than hours to launch, and multiply by twelve. That figure, not the build estimate, is what you compare against a platform's annual cost.
5. Check whether compliance changes the answer
If you handle health data, look at the published add-on prices before anything else. Vapi lists HIPAA at $2,000 per month and Zero Data Retention at $1,000 per month on either plan as of August 2026; Retell prices PII removal at $0.01 a minute instead. For a small clinic those figures can exceed the value of the automation, and in that case self-hosting parts of the stack becomes a genuine argument rather than a preference.
6. Start bought, keep the assets portable
Whichever you choose, keep your prompts, call scripts, business rules and transcripts in your own repository rather than only inside a vendor console. Prompts are plain text and port anywhere, numbers can generally be ported between carriers, and transcripts become the test set for any future migration. Doing this from day one makes the build-versus-buy decision reversible rather than final.
7. Revisit the decision at a volume threshold, not on a hunch
Set a trigger in advance: a monthly minute count or a monthly bill above which you re-run this comparison. That keeps the decision evidence-driven rather than a reaction to one frustrating month. For most small businesses the trigger never fires, which is itself the answer. For businesses where voice becomes the product, it fires clearly and early.
Buy when, build when
Buy when voice is how customers reach you rather than what you sell. On the reference workload of 1,500 minutes a month, the published August 2026 rates put a bought agent at $210 on Bland Start, $105 to $465 on Retell depending on model choice, or roughly $120 to $200 all-in on Vapi. Those are small numbers against the cost of a missed call, and they include the part of the problem that is genuinely hard: turn-taking, endpointing, interruption handling and streaming latency across several providers.
Build when voice is central to the product you sell, when your call volume makes per-minute fees a serious line item rather than a rounding error, or when a data requirement means inference cannot happen on someone else's infrastructure. Those are the three honest reasons. Wanting more control, disliking vendors and enjoying the engineering are not reasons - they are preferences, and they get expensive quietly.
There is a middle path that most technically capable teams end up on, and it deserves naming. Vapi's unbundled model - $0.05 a minute for orchestration with your own provider keys, so its share of model and voice cost is zero - gives you component ownership and cost visibility without owning the real-time voice loop. You get most of what building offers and skip the part that takes weeks to get right.
Whichever you pick, make it reversible. Keep prompts, call scripts, business rules and transcripts in your own repository rather than only inside a vendor console. Prompts are plain text and port anywhere, numbers can generally be ported between carriers, and transcripts are the test set for any future migration. Doing this from day one costs nothing and turns a permanent decision into a revisitable one.
What we will not put a number on
We are not going to publish a dollar figure for building a voice agent. Figures in circulation range from a few thousand to hundreds of thousands, and every one of them is a function of team rate, scope, integration depth and compliance rather than of the technology. Quoting a number we have not verified would be exactly the kind of fabricated benchmark this site exists to avoid.
What we can state is the shape of the work, because it is the same on every build: renting speech-to-text, a language model and speech synthesis, arranging telephony and carrier configuration, and then writing the orchestration that handles turn-taking, endpointing, interruptions and streaming latency. The first four are ordinary API integration. The fifth is the hard part, and it is precisely what the platforms in this category sell.
We can also state which cost people underestimate: maintenance rather than build. Model deprecations, carrier changes, voice provider updates, latency regressions and the ongoing prompt tuning every voice agent needs all land on whoever built it, forever. Estimate hours per month rather than hours to launch and multiply by twelve. That annual figure, not the build estimate, is what a platform's annual cost has to beat.
And we can state the one number that reliably decides it: your monthly minutes. Below a few thousand minutes a month, the per-minute saving from building cannot cover any plausible engineering estimate. Above a genuinely high volume, it can. Set that threshold in advance so the decision is arithmetic rather than a reaction to a frustrating month.
Where both approaches are the wrong answer
Both are wrong if the calls need human judgement. A voice agent handles booking, qualifying, confirming and answering documented questions well. It handles a distressed customer, a complex complaint or a negotiation badly, and it does so with complete confidence, which is worse than doing it hesitantly. Split your call types before you price anything - the calls an agent should never take are not a build-versus-buy question.
Both are wrong if the underlying process is broken. If callers abandon because your opening hours are wrong, your booking system is slow or nobody has the information to answer, an agent will handle those calls more consistently and leave the actual problem untouched. Voice AI is an excellent answer to unanswered calls and a poor answer to a broken process behind the phone.
Both are wrong if you handle regulated data and have not priced the compliance line first. Vapi lists HIPAA at $2,000 a month and Zero Data Retention at $1,000 a month; Retell prices PII removal at $0.01 a minute instead. For a small clinic those figures can exceed the value of the automation entirely, and finding that out after a build is the most expensive version of this mistake. Price compliance before writing anything.
At a glance
Build vs buy an AI voice agent, published platform pricing as of August 2026
| Decision factor | Buy a platform | Build it yourself |
|---|---|---|
| Cost at 1,500 min/mo | $105-$465 depending on platform and model choice | Provider costs only, plus your engineering time |
| Time to first real call | Days | Weeks, and longer to make it reliable |
| Who owns turn-taking and latency | The platform | You, on every provider change |
| Ongoing maintenance | Prompt tuning and monitoring | Prompt tuning, monitoring, plus model, voice and carrier changes |
| Compliance | Priced add-ons or per-minute features you can budget | Yours to design, implement and evidence |
| Data location control | Depends on platform; strongest on unbundled providers | Complete, which is the strongest reason to build |
| Concurrency and scaling | Published limits you buy more of | Your infrastructure problem |
| Reversibility | High - prompts and numbers port, transcripts are yours | High on assets, low on the engineering investment |
| When it wins | Voice is a channel, volume is normal, no developer to spare | Voice is the product, volume is high, or data cannot leave |
What mistakes should you avoid?
- Comparing a build estimate against a platform's per-minute rate without adding a year of maintenance to the build
- Assuming the hard part is the model when it is turn-taking, interruptions and latency across four providers
- Building to save a few tens of dollars a month at small-business call volume
- Buying without checking the compliance add-on price when you handle health or financial data
- Keeping prompts, scripts and transcripts only inside a vendor console, which makes the decision irreversible
- Testing with your own voice in a quiet office rather than real callers on real phones
- Treating build versus buy as permanent instead of setting a volume threshold to revisit it
FAQ
- What does buying an AI voice agent cost?
- On 1,500 minutes a month as of August 2026: Bland Start is $210 at $0.14 a minute with no monthly fee, Retell falls in its published $0.07 to $0.31 range so $105 to $465 depending mostly on model choice, and Vapi is $75 of orchestration plus separately billed providers, typically $120 to $200 all-in. Telephony and transfer minutes are extra on several of these.
- What does building actually involve?
- Renting speech-to-text, a language model and speech synthesis, arranging telephony and carrier configuration, and then writing the orchestration that handles turn-taking, endpointing, interruption handling and streaming latency. The first three are straightforward API work. The orchestration is the hard part, and it is precisely what these platforms exist to sell you.
- When is building genuinely the right call?
- When voice is central to what you sell rather than how customers reach you, when your call volume makes per-minute platform fees a serious line item, or when a data requirement means inference cannot happen on someone else's infrastructure. Outside those, the per-minute saving at small-business volume does not cover the engineering, let alone the ongoing maintenance.
- Is there a middle option?
- Yes, and it is what most technically capable teams end up doing: use an unbundled platform like Vapi, which charges $0.05 a minute for orchestration and lets you bring your own provider keys so its share of model and voice costs is zero. You own the components and the cost visibility without owning the real-time voice loop, which is the genuinely hard engineering.
- How locked in am I if I buy?
- Less than people expect, and it varies. Prompts and call scripts are portable text on every platform here, and phone numbers can generally be ported. Vapi is the loosest because your providers are your own accounts. Bland is the tightest because the model, transcription and voice are all its own, so behaviour will not reproduce identically elsewhere. Export transcripts regularly either way.
- What breaks first in production?
- Latency, before anything else - a voice agent that pauses for a second feels broken to a caller in a way a slow chatbot never does, and the pause is usually the sum of several providers' response times. Then transcription degrading on accents and background noise, which propagates into wrong answers. Then the agent failing to notice a caller who wants a human.
Sources & further reading
Done for you
Want this built for your business?
Tell us the workflow. We scope, build, and ship the agent with guardrails and the numbers to prove it worked. The scoping call is free.
Free scoping call. You own the code.
Was this page helpful?

