On 2 November 2026, Salesforce signed a definitive agreement to acquire Cresta, a voice-first customer service platform whose agents resolve calls end to end. The deal is expected to close in the first quarter of Salesforce's fiscal 2028. Read the coverage and you will find enterprise names: Marriott, Comcast, Alaska Airlines. Read it again as the owner of a twelve-person agency or a plumbing company with four trucks, and it says something more useful — the layer of voice AI you were about to rent is being absorbed upward, and the price you were quoted was never yours to keep.
That matters right now because tens of thousands of small businesses and agencies are in the middle of a build-versus-buy decision on an AI voice agent, and almost every guide answering it was written for somebody else.
Salesforce described Cresta as a voice-first operating system for customer experience, and its Agentforce business is already past $1.5 billion in annual recurring revenue, up 240% year over year. Cresta had raised more than $270 million, and its own published case study with Brinks Home claims quality-management costs down 50% and an NPS gain of 30 points.
None of that is small-business news on its face. Here is the part that is:
Put those two facts together and the real decision is not "build or buy." It is which layer you own, which layer you rent, and how exposed you are when the layer you rent changes hands.
Search the question today and you get a dozen 2026 guides. Most are competent, and most share three flaws that matter at your scale.
A voice agent is not a project you finish. It is a system that has to keep answering the phone in month nineteen, when the models underneath it have changed and the person who built it has left.
A production agent is six layers, not one prompt. Every layer has a failure mode you own:
| Layer | What it does | Typical choices | What breaks |
|---|---|---|---|
| Telephony | Places and receives calls | Twilio, Telnyx, SIP carriers | Number reputation, carrier registration, failed transfers |
| Speech recognition | Turns caller audio into text | Deepgram, Google, Azure | Accents, names, background noise |
| Language model | Decides what to say and do | OpenAI, Anthropic, Gemini | Hallucinated prices, wrong appointment slots |
| Text-to-speech | The voice the caller hears | ElevenLabs, Cartesia, Deepgram | Prosody that sounds robotic under stress |
| Orchestration | Turn-taking, interruptions, tool calls | Vapi, LiveKit, custom code | Dead air, talking over callers, loops |
| Data and tools | Calendar, CRM, job records | Your CRM, Supabase, internal APIs | Bookings that never reach the schedule |
The component prices look almost free. Published rates as of late 2026:
| Component | Published rate |
|---|---|
| Twilio US local outbound / inbound | $0.0140 / $0.0085 per minute |
| Streaming speech recognition (Deepgram) | ~$0.0065–$0.0077 per minute |
| Text-to-speech (ElevenLabs class) | ~$0.0146–$0.0238 per minute |
| Orchestration hosting (Vapi) | $0.05 per minute, plus $29/mo Core or $999/mo Pro |
| Managed all-in (Retell) | $0.07–$0.31 per minute, plus $8 per concurrent call past 20 |
So a "build" on paper is 5 to 10 cents a minute. The bill that kills it is not the per-minute rate — it is the person. A 2026 analysis using the US median software developer wage of $135,980 puts one developer at roughly $2,615 a week, salary only, before tax, benefits or equity. That analysis prices a ten-week build on your own stack at about $26,150 to launch, then $2,266 a month in upkeep; four weeks on a developer platform at about $10,460 to launch, then $1,133 a month.
And that is the optimistic version, because it assumes the build is a normal engineering project. It is not. Conversational latency is a hard failure mode: research cited in build-versus-buy analyses finds voice interaction degrading past roughly 0.4 seconds of delay, with callers starting to talk over the agent after 0.5. Tuning that is a skill, not a ticket.
Packaged voice AI bills three ways, often simultaneously: per minute of conversation, per seat for the phone system wrapped around it, and per concurrent call. The sticker is rarely the bill. Knowledge-base fees, PII redaction, call recording retention and separate telephony charges all appear on the invoice later.
For a business that just wants its phone answered, that is manageable. For an agency that wants to resell it, buying creates a structural problem no pricing page mentions:
This is why "buy" and "own" are not the same word. Buying a seat is the fastest way to get a working agent. It is also the fastest way to build a book of business on ground you do not control.
Build economics only work at volume, and the gap is wider than most guides admit. Using midpoint rates and engineering at the median wage:
| Minutes per month | Build it yourself | Developer platform | Buy a platform |
|---|---|---|---|
| 1,000 | ~$4,517 | ~$2,127 | ~$150 |
| 10,000 | ~$5,163 | ~$3,222 | ~$1,500 |
| 100,000 | ~$11,620 | ~$14,180 | ~$15,000 |
At a thousand minutes a month, building costs roughly thirty times what buying costs, and 98% of that is the engineer. The crossover sits somewhere around 50,000 to 57,000 minutes a month. Below it, building is a hobby with a payroll. Above it, the per-minute margin starts paying a salary — and keeps paying it every month.
Take an HVAC company answering 1,200 inbound minutes a month at an average job value of $480, booking on roughly 8% of answered calls. This is illustrative arithmetic, not a benchmark:
The number that decides this is not the demo. It is minutes per month, and almost nobody measures it before choosing.
Between "write it yourself" and "rent a seat" sits a model that fits agencies and multi-location service businesses: license a finished voice platform and operate it as your own product.
You keep the parts that are genuinely yours — the script, the qualifying questions, the tone, the escalation rules, the client relationship, the invoice, the logo. You stop owning the parts that are commodities — speech recognition, model orchestration, telephony plumbing, latency tuning, on-call at 2am.
| Build in-house | Subscribe to a vendor | License a white-label platform | |
|---|---|---|---|
| Time to a trustworthy agent | 4–9 months | 5–14 days | Days to weeks |
| Who maintains the stack | Your engineers, permanently | The vendor | The vendor ships, you run |
| Whose brand the client sees | Yours | The vendor's | Yours |
| Control over behaviour | Total | Configuration only | Total within the platform |
| Exposure to repricing | None | High | Contractual |
| Who owns the customer | You | Shared, at best | You |
That is the model Nida is built for: a white-label AI phone agent, a client portal and a website generator in one platform, so an agency can put an answering front desk in front of a client under its own brand, and a service business can have one without hiring an engineer. See plans and pricing, or read the breakdown of white-label AI receptionist platforms for agencies.
Being honest about the other side matters, because the build case is real in four narrow situations:
If none of those four is true, building is not control. It is an unfunded maintenance obligation with a launch date.
For the operating numbers behind that pilot, the real running costs of an AI voice agent platform and the ROI statistics for small business are the two most useful reads before you sign anything. If you are comparing developer platforms specifically, Vapi vs Retell vs Bland breaks down where each one's all-in rate actually lands.
Buying or licensing is cheaper for almost every business under about 50,000 minutes a month. A ten-week own-stack build runs roughly $26,000 to launch plus $2,266 a month in upkeep, against $0.07–$0.31 per minute for a managed platform. The build only wins at sustained high volume, or when voice is the product itself.
A managed or licensed platform typically goes live in five to fourteen days, mostly spent on call-flow design and testing. A custom build that handles interruptions, transfer and CRM write-back reliably takes four to nine months, and a further period before you would trust it with paying customers.
Raw components land around $0.05–$0.10 a minute: telephony near $0.01, speech recognition under a cent, text-to-speech in the low two cents, plus orchestration. Managed platforms charge $0.07–$0.31 all in, and the difference is the platform fee, concurrency charges and the extras added after the headline rate.
Yes, but only on a platform that supports private labelling — otherwise you are reselling a vendor's product under your own invoice. Confirm three things before signing: whether the client-facing interface carries your brand, whether pricing is contractual rather than a public rate card, and who owns the client record and the phone number if you leave.
Usually nothing on day one, and then a slow narrowing. Acquisitions redirect roadmaps toward the enterprise buyer, freeze the small-account plan and raise the renewal. If your agent sits on a licensed platform you operate, the risk is a contract discussion; if it sits on a rented seat, the risk is your price.
Build if voice is the product and you have the volume and the engineer. Buy a seat if you need an agent live this week and do not mind that the rails are someone else's. If you are an agency, or a service business that intends to be on the phone for the next five years, license a platform you can put your own name on — own the conversation and the customer, rent only what has become a commodity.
The Salesforce–Cresta deal is not a warning about one vendor. It is a reminder that the top of this stack is consolidating fast, and the businesses that come out of it well are the ones that kept the part that was theirs. Start with the Nida plans, or work out what an AI receptionist would return for a contractor business before you spend a dollar on infrastructure.
← Back to the blog