How Much It Costs to Build an AI Voice Agent (2026)

A production voice agent costs what its scope, integrations, and compliance tier cost in engineering weeks. The models are the cheapest line on the bill. That is why a quote can read $5,000 or $500,000 for what sounds like the same product: the two numbers price different objects, and the label “voice agent” hides the difference.

For a founder or CTO comparing those quotes, the only way to tell overbuild from under-scope is to assemble the number from its parts and price each one.

The 100x gap is real because the quotes price different things

Three genuinely different systems get sold under one name.

What the quote actually describes	Estimate band
Configured platform agent (one flow, light integration, set up by one person)	$1,000–$15,000 one-time, or $200–$500/mo
Custom integrated build (tool-calling, escalation, evals, monitoring)	$80,000–$200,000
Compliance-grade bespoke system (HIPAA, self-hosted components, multilingual, audit trails)	$180,000+

These are our estimate bands, built from engineering weeks, not published market rates. No official body reports what a custom voice agent costs.

Six drivers open the gap, ranked by how much they move the number:

1. Scope collapse in the label. A $5,000 “voice agent” is a no-code platform configured in a week for one flow. A $500,000 “voice agent” is a bespoke multi-intent system with deep integrations, a compliance surface, an evaluation harness, and an operations function. Same two words, different systems.
2. Labor rate, up to 5–8x on its own. A US senior engineer-week and an offshore agency week are not the same unit cost. Rate alone can swing a 12-week build by more than $150,000 before scope enters.
3. Integration depth. In our production work, integration and its testing routinely consume 20–50% of a build budget. Each proprietary system (CRM, EHR, booking engine, auth, payments) is discrete engineering plus its own test surface.
4. Compliance tier. HIPAA and a business associate agreement (BAA), PII redaction, recording consent, audit trails, and data residency are a different security surface entirely. A regulated agent is a different object from a booking bot.
5. What gets counted as “the build.” Cheap quotes exclude evals, monitoring, telephony hardening, and post-launch tuning. The demo works, then the re-quote arrives.
6. Buy-vs-build baseline. A platform subscription competes with a six-figure build for the same visible capability at low volume. Comparing a subscription to a custom quote compares a run cost to a capital cost.

“How much does it cost” has no answer. The question that produces a real number: what am I buying, integrated to what, at which compliance tier, and who runs it after launch.

What an engineering week actually costs

Before scope, the unit price of labor is the single largest swing in a quote. Two numbers get conflated here and should not be: salary and billed rate.

Salary is the floor. The US Bureau of Labor Statistics reports 2024 median pay for software developers, QA analysts, and testers at “$131,450 per year / $63.20 per hour.” That is base plus cash, before benefits, taxes, overhead, and margin. It is not what a vendor bills.

Billed rate sits well above salary because it carries margin, utilization risk, and no benefits. We model external build work at three tiers:

Delivery model	Blended rate per engineering week
Mixed or offshore team	$4,000–$7,000
Senior US/EU consultancy	$7,000–$12,000
Specialized regulated, telephony, or security work	$10,000–$18,000

The consequence is arithmetic. The same 12-week scope quotes at roughly $48,000, $120,000, or $216,000 depending on team mix and delivery model. That spread explains more of the 100x than model prices ever will. Normalize any two quotes to a rate before comparing their scope.

Where the money goes, and where it does not

Rank the cost stack for a real production build and the models come last:

1. Integration and tool-calling
2. Evaluation and observability harness
3. Conversation design and turn-taking tuning
4. Telephony hardening
5. The models themselves

A quote heavy on model choice (“we’ll use the latest frontier model”) and thin on CRM contracts and evaluation is under-scoped. Weight the detail in a quote the way the stack is weighted, not the way the sales deck is.

The per-minute component rates prove the point. These are runtime costs, not build costs, and they are small. Provider rates below checked July 2026; they churn quarterly, so re-verify before signing anything.

Component	Representative published rate	Source
Telephony (inbound)	Local calls receive $0.0085/min; SIP interface $0.0040/min	Twilio
Voice-suitable LLM	Claude Haiku 4.5 at $1.00 in / $5.00 out per 1M tokens	Anthropic
Speech-to-text (bundled agent)	Deepgram Voice Agent API “Standard $0.075/min”, “Advanced $0.163/min”	Deepgram
Text-to-speech	ElevenLabs “$0.05 (Flash/Turbo)” per 1,000 characters	ElevenLabs

At a voice-suitable model like Haiku 4.5, the LLM runs on the order of a cent to a few cents per minute of conversation. It is the smallest number in the build.

One caution on models: reasoning modes are unusable in a live call. Time-to-first-token runs into tens of seconds and drops the conversation, so reasoning models stay out of the live path, and a quote should never frame the model as the cost driver.

The per-minute economics, and the platform tiers that wrap them, sit in the deployment and scaling guide and the platform comparison.

Before anyone can price it: the inputs you must supply

No honest number exists without these. A vendor who quotes a firm price for a full production agent before asking them is either padding for unknowns or under-scoping to win the deal.

• Monthly minutes and average call length
• Peak concurrency (simultaneous calls at the busy hour, not the monthly average)
• Inbound, outbound, or both
• Languages, and whether callers switch languages mid-call
• Number of tools the agent calls, and whether each is read-only or writes back
• Compliance class (PHI, PCI, recording consent, data residency)
• Recording retention and deletion requirements
• Human escalation model
• The success metric the agent is measured against

Peak concurrency deserves its own emphasis. Monthly minutes underprice launch risk. 10,000 minutes spread evenly is a different capacity problem than the same minutes arriving in a two-hour window, and platform tiers are priced on concurrent lines, not monthly totals.

The quote built in front of you: three reference agents

Take the engineering hours each part requires, price them at a blended $6,000–$10,000 per senior engineering week, and the bands fall out. State the rate so a reader can rescale at their own team’s cost. These are illustrative models, not quotes from any Softcery engagement.

Agent 1: inbound booking (single flow, calendar)

One inbound number, one conversational flow, calendar availability lookup, appointment creation, caller confirmation, fallback to a human, transcripts, a basic test set, one deployment path.

Line item	Effort
Discovery and call flow	0.5–1 week
Platform or orchestration setup	0.5–1.5 weeks
Calendar integration	1–2 weeks
Prompt, voice, fallback	0.5–1.5 weeks
Testing	1–2 weeks
Launch and monitoring	0.5–1 week
Total	4–9 weeks, roughly $25,000–$60,000 custom

Configured on a platform by one person, the same capability is 1–2 weeks and $5,000–$15,000 one-time, or a $200–$500/mo subscription. Verdict: this tier is usually a buy, not a build. Custom is justified only when the flow touches proprietary systems or a compliance driver exists.

Agent 2: CRM-integrated support (tool-calling, escalation, evals)

Identity lookup, CRM read and write, ticket creation, knowledge retrieval, escalation with context handed to the human, authentication for tools, event logging, regression evals, PII handling.

Line item	Effort
Discovery and KPI design	1–3 weeks
Conversation design and tool policy	2–5 weeks
CRM and tool integration	4–10 weeks
Evaluation harness and replay suite	2–6 weeks
Observability and analytics	1–4 weeks
Telephony hardening	1–2 weeks
Pilot, tuning, handoff	2–5 weeks
Total	14–23 weeks, roughly $80,000–$200,000

Integration and evals dominate; the models are a rounding error. Most of the range hides inside one ambiguity: “CRM integration” can mean a read-only lookup, a ticket creation, or an account write-back with permissions and rollback. Those are three different amounts of engineering. A legacy CRM or a write action that touches money pushes toward the top of the band.

This is the modal “real” build. Our published CaseGen work shows the shape of it: runtime switching across LLM, STT, and TTS providers with fallback for rate limits, warm transfer with an AI-generated briefing for the human, and LLM-based quality evaluation on every call. None of that appears in a demo.

Agent 3: compliance-heavy (HIPAA, self-hosted components, multilingual)

Everything in Agent 2, plus a BAA chain, recording consent, retention and deletion workflows, audit logs, role-based access, redaction, possibly self-hosting or a VPC, and per-language QA.

Roughly 21–35 weeks, $180,000 and up, often well beyond. The high end buys the security, compliance, infrastructure, and operations built around the agent, not the agent itself. Two facts set the cost:

• HIPAA and zero-retention pull against each other. Anthropic states that covered models “require 30-day data retention and aren’t available with zero data retention (ZDR).” An agent that needs both has to architect around the constraint, for example by self-hosting. Design constraint equals cost.
• Compliance plumbing on a platform is a line item, not a checkbox. Vapi publishes its HIPAA add-on at “$2000/mo” and zero data retention at “$1000/mo”, and in HIPAA mode “Vapi will not store call recordings or transcripts”, so “you must configure a custom storage bucket.” The retention nobody wanted to build becomes the customer’s build.

Self-hosting is the clearest build-up, run-down tradeoff. It raises the build cost (model serving, GPU ops, monitoring) and lowers the run cost at volume. Our Skipify/Ella work put a co-located pipeline in place that cut total voice infrastructure cost by more than 50% versus a vendor stack, with speech an order of magnitude cheaper at saturation. That only pays off above a real volume floor, which is why it is an Agent-3 decision, not a starter move.

What you pay for that a demo does not include

A demo is the happy path on a developer’s laptop. It proves the stack can speak. It does not prove the system can be trusted with customers. Production adds line items the demo never touches, and their absence from a quote predicts a re-quote after launch.

• Evaluation harness and datasets. This is engineering weeks, not a SaaS seat: building regression suites from real call recordings, defining scoring criteria, and gating releases. On CaseGen, every change runs against the full real-call dataset before it ships.
• Multi-provider fallback. Building and testing failover across LLM, STT, and TTS providers so one outage does not drop the call.
• Telephony hardening. PSTN audio at 8kHz degrades recognition and naturalness. DTMF handling, answering-machine detection, warm transfer, SIP trunk config, jitter and codec handling. None of it shows on a laptop. The telephony stack piece covers the surface.
• Turn-taking and barge-in tuning. Endpointing and interruption recovery are iterative work against real callers, not a config toggle. See the latency budget.
• Observability. Traces, per-call cost and latency tracking, anomaly alerts, per-customer spend limits. Build and wire-up effort.
• Voice conversation design. Prompts that work in text break in real-time audio and need rewriting for speech.
• Compliance plumbing. PII redaction, consent capture, retention controls, and the BAA work described above.

These items are the 20–50% integration share plus the hidden total-cost gap. The demo is the cheap 10%. The failure mode is covered in why voice agents sound great in demos but fail in production.

How to read your quote, and what to cut to hit a budget

Validate a quote by its exclusions and line items, not its headline number. The cheapest quote is often cheapest because it shifts post-launch risk onto your team. Treat these as red flags:

• A single lump sum with no line items. Impossible to tell what it buys, or to cut anything.
• No line for evaluation or observability. The demo-grade 10% is being sold as the whole.
• No mention of telephony hardening, provider fallback, or turn-taking tuning. The production surface is missing.
• The model framed as the cost driver. Models run cents per minute; if the model is the headline, the integration and eval work is under-scoped.
• A six-figure quote for Agent-1 scope (single flow, no proprietary data). Overbuild. A platform config fits.
• A $5,000–$15,000 quote for Agent-2 or Agent-3 scope. Under-scope. Expect a rebuild inside a year.

On contract shape: a fixed-bid pilot into a time-and-materials or milestone-based production build fits how voice-agent scope actually evolves. A firm fixed price for a full production agent quoted before any pilot is guesswork, priced either with heavy padding or with corners hidden.

To hit a fixed budget, these levers move the number without wrecking the agent:

• Down: use a platform instead of self-assembling the stack; cut to the one integration that matters; single language; managed models instead of self-hosting; a narrower eval scope at launch, expanded later; inbound only, since outbound adds dialing, answering-machine detection, and consent compliance.
• Up: each proprietary integration; the compliance tier; self-hosting; multilingual and code-switching; outbound at volume; strict latency targets that demand more turn-taking tuning.

When custom build is the wrong answer

Below a volume and complexity threshold, custom is waste. The cost logic is simple; the full decision procedure sits in the build-vs-buy piece.

A custom build has to beat a platform subscription plus usage over its amortization window. The math rarely works at low volume. At 10,000 minutes a month, a $0.10/min platform is about $1,000/month in usage. A $100,000 custom rebuild takes years to pay back on usage savings alone. At 500,000 minutes a month, that platform is roughly $50,000/month, and custom infrastructure or bring-your-own-carrier starts to make economic sense, provided quality and operations are staffed.

Buy when the flow is a single generic pattern, touches no proprietary systems, and has no compliance driver. Build when tool logic is proprietary or write-heavy, when compliance forces ownership of the data path, when voice is the product rather than a feature, or when volume clears the economic threshold.

Right-sizing works in the other direction too. Our STRAI work shipped a messaging agent at $0.01–$0.10 per message in three iterations over eight weeks, handling the majority of messages without a human. That is a build scoped to the job, not an over-engineered one. Sometimes the honest answer is that the reader does not need to build anything yet.

What it costs to keep running

A voice agent is not done at launch. Product changes, model revisions, CRM API changes, and caller behavior all demand ongoing work. Post-deployment monitoring is a recognized requirement, not an optional extra: the NIST AI Risk Management Framework calls for “Post-deployment AI system monitoring plans” covering “incident response, recovery, and change management.”

We plan annual maintenance at 15–30% of the initial build for a stable production agent, and 30–60% for regulated, high-volume, or frequently changing workflows. Treat that as our planning estimate tied to a specific build number, not an industry constant, and do not average the band; the driver is scope. The recurring lines behind it:

• Model drift and provider churn: re-tuning prompts and re-running evals on every model swap
• Prompt and conversation-policy upkeep
• Eval runs as a real recurring cost in both compute and engineering time
• On-call and incident response for fallback and latency regressions
• Integration maintenance when CRM, calendar, or helpdesk APIs change

Below roughly 10,000 minutes a month, a platform plus a light retainer covers this. Above it, or for an Agent-2 or Agent-3 system, plan for a standing engineering function or a partner retainer.

A few one-time costs beyond engineering surprise people:

• Phone numbers. Twilio lists local numbers at “$1.15 / mo” and toll-free at “$2.15 / mo.”
• Carrier registration, only if the agent sends SMS (appointment texts, document collection). US A2P 10DLC brand registration runs from “$4.50” for low-volume through “$46” standard including secondary vetting, plus a “$15 vetting fee” per campaign and “$1.50-$10/month” per campaign. Voice-only inbound agents skip most of it.
• First-month API and eval-run spend. Running the eval suite repeatedly during tuning is a real build-phase cost, separate from steady-state runtime.

The business metric that matters is cost per resolved call, not cost per minute: total AI cost divided by calls contained without a human. If half of calls escalate, the AI cost per resolved call roughly doubles before human cost enters. The per-minute and per-call runtime math sits in the deployment and scaling guide.

Get the bill of materials first

No source publishes what a custom voice agent should cost, because the number is a function of scope, integrations, compliance tier, and who owns operations after launch. The defensible first deliverable is a line-itemed bill of materials pinned to those four things. With it in hand, a reader can price their own agent, read any quote by its exclusions, and see exactly where the money goes.

That mapping, cost and feasibility before budget commits, is what our advisory engagement does. Test your own minutes, platform, model, concurrency, and compliance assumptions in the voice agent cost calculator, and read the full build guide for the architecture behind these numbers.