Custom voice AI agent development cost is set by four things: how many systems the agent touches, how many distinct jobs it does, the compliance tier it must meet, and the engineering rate of the team that builds it. A focused inbound agent with one integration costs $25,000 to $60,000. A multi-intent agent that reads and writes to your CRM costs $80,000 to $200,000. A HIPAA or PCI deployment starts around $180,000.
Ask three firms for a quote and you'll often get three numbers that differ by 4x. That isn't evasion. They're pricing different scopes at different rates. This guide shows you how to build the number yourself, what it costs to run, when it pays back, and what each person on the leadership team should check before signing.
What does a custom voice AI build include?
A production voice agent is a real-time system plus integrations, testing and monitoring; the conversational demo is the smallest part of the work.
A custom voice AI agent is software that answers or places calls, understands natural speech, takes action in your systems and hands off to people when it should. "Custom" means it's built for your call flows, your data and your infrastructure, rather than configured inside a vendor's dashboard.
A production build has seven parts, and every serious quote should price each one:
Discovery and call design. Mapping real calls to intents, scripts, data needs and escalation rules.
Real-time voice pipeline. Telephony, WebRTC media, speech-to-text, turn detection, the LLM and text-to-speech, tuned for sub-second replies.
Integrations. CRM, scheduling, billing, practice management or ERP, each with read and write logic and error handling.
Evaluation and testing. Regression tests built from recorded calls, plus load and accent testing.
Observability. Per-call traces, latency, cost and containment dashboards with alerts.
Compliance controls. Consent, recording retention, PII redaction, audit trails and the right agreements with every vendor in the chain.
Launch and handover. Staged rollout, runbooks, documentation and, if you own it, the source code.
A demo covers parts 2 and some of 1. That's why a demo can be built in a week and a production system takes months.
How much does a custom voice AI agent cost in 2026?
By scope, a custom build costs $25,000 to $60,000 for one flow, $80,000 to $200,000 for a CRM-integrated multi-intent agent, and $180,000 or more for regulated deployments.
The ranges below combine engineering effort by scope with 2026 market data. They cover the build only. Running and maintenance costs come later in this guide.
Scope | What it looks like | Typical timeline | Build cost range |
Configured platform agent | One flow set up inside Vapi, Retell or similar; light integration | 1 to 2 weeks | $5,000 to $15,000 one-time |
Single-flow custom agent | One intent (for example booking), one integration, inbound only | 4 to 9 weeks | $25,000 to $60,000 |
Multi-intent production agent | Several intents, CRM read and write, escalation with context, evals and monitoring | 14 to 23 weeks | $80,000 to $200,000 |
Regulated or compliance-grade | Everything above plus HIPAA or PCI, audit logs, data residency, often self-hosted parts, multiple languages | 21 to 35 weeks | $180,000 and up |
Effort bands and scope tiers adapted from Softcery's 2026 bill-of-materials analysis (Jul 2026), priced at $6,000 to $10,000 per senior engineering week.
Other published 2026 guides land in the same zone. Biz4Group puts most custom voice agents at $20,000 to $200,000, RaftLabs at $15,000 to $150,000+, and Master of Code puts a cloud-plus-custom build at $25,000 to $150,000. No official body publishes this number, so treat every range, including ours, as a model built from engineering time.
Why do quotes differ by 4x? The engineering-week math
The same 12-week scope costs about $48,000, $120,000 or $216,000 depending on the team's weekly rate, so always normalize quotes to a rate before comparing scope.
Labor rate is the biggest swing in any quote before scope even enters. Softcery models external build work at three blended weekly rates:
Delivery model | Blended rate per engineering week | 12-week build |
Mixed or offshore team | $4,000 to $7,000 | $48,000 to $84,000 |
Senior US or EU consultancy | $7,000 to $12,000 | $84,000 to $144,000 |
Specialized regulated, telephony or security work | $10,000 to $18,000 | $120,000 to $216,000 |
Rates from Softcery (Jul 2026). For context, the BLS median salary for US software developers was $131,450 in 2024, before benefits, overhead and margin.
This is why two honest quotes can sit far apart. When you compare vendors, ask each one for the weeks and the weekly rate behind their number. Then compare weeks to weeks.
A senior team in a lower-cost market, such as RTC LEAGUE engineering team, can deliver the same scope at the lower end of these rates. What you should never trade away for a lower rate is real-time experience. A team that has never tuned turn detection or hardened a SIP trunk will spend your budget learning.
The 5 Cost Drivers Scorecard
Score your project on five drivers; each "high" moves you up a tier, and integration count matters most.
We use the 5 Cost Drivers Scorecard to place a project inside the ranges above before anyone writes a quote. Score each driver low, medium or high.
Driver | Low | Medium | High | Why it moves cost |
1. Integrations | 1 system, read-only | 2 systems, some writes | 3+ systems, writes, or a legacy system with no documented API | Integration and its testing often take 20% to 50% of the build budget |
2. Intents | 1 job | 2 to 4 jobs | 5+ jobs | Each intent needs its own logic, tests and failure handling |
3. Call direction | Inbound only | Outbound informational | Outbound sales at volume | Outbound adds consent tracking, voicemail detection and calling-window rules |
4. Compliance | None | GDPR, recording consent | HIPAA, PCI, data residency | Changes the infrastructure design as well as the paperwork |
5. Ownership | Platform config | Custom build, vendor-hosted | Source code, your cloud, full handover | Owning the system means paying for the engineering to make it maintainable by others |
How to read it: mostly low lands in the single-flow tier. Two or more mediums put you in the multi-intent tier. Any high on integrations or compliance usually means the top two tiers.
Driver 1: integrations, the most underestimated line
One CRM is a known problem. A CRM, a scheduling system, a billing platform and a legacy system with no API is four problems, and the fourth is where quotes go wrong. "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 work. Name every system and every action in your request for proposal.
Driver 2: intents, where scope creep hides
An agent that books appointments is one scope. One that books, answers billing questions, processes cancellations and qualifies leads is four. Scope creep in voice AI almost always arrives as "can it also handle..." Fix the intent list before the contract, and price additions as change orders.
Driver 3: outbound calling
Outbound calls need consent records, answering-machine detection, call-time windows by state or country and opt-out handling. In the US, telemarketing calls also fall under TCPA rules. If your project includes outbound sales calls, budget legal review alongside engineering.
Driver 4: compliance
HIPAA, PCI or a data residency rule changes where the system runs and which vendors you can use. Each vendor in the chain may need a business associate agreement. Platform compliance isn't free either: Vapi lists its HIPAA add-on at $2,000 a month per organization.
Driver 5: ownership
Some engagements hand you the source code, infrastructure and model configuration. Others give you access to a system the vendor runs. Both are valid, but they're different purchases. Settle this before signing, because it decides whether you can switch vendors later.
What does it cost to run a voice AI agent per minute?
A custom stack on LiveKit runs about $0.05 a minute and hosted platforms run about $0.08 to $0.15 a minute for typical builds, against about $0.48 a minute for a human agent's wages.
Running cost is billed per minute of conversation. Here's how the main options compare at November 2026 list prices.
Option | Typical all-in cost per minute | What's included | Source (date) |
Custom stack on LiveKit Cloud | About $0.048 | Agent session, telephony, observability, STT, LLM, TTS (default models) | LiveKit calculator (Nov 2026) |
Vapi | About $0.08 to $0.13 | $0.05 hosting plus models at provider cost | Dupple, from Vapi's calculator (Oct 2026) |
Retell AI | $0.07 to $0.31; calculator default $0.11 | $0.055 voice infrastructure plus model, voice, telephony | Dupple, from Retell's pricing page (Oct 2026) |
Human customer service rep | About $0.36 wages only; about $0.48 at 75% talk time | Median wage of $21.53 an hour | BLS (May 2025) |
The human figure is wages only, before benefits, management, software and facilities. The 75% talk-time assumption is ours.
Two points matter for budgeting. First, the model you pick moves the bill more than the platform does; a premium LLM or voice can double the per-minute rate. Second, measure cost per resolved call, not cost per minute. If half of calls escalate to a person, the AI's cost per resolved call roughly doubles before you add the human's time.
Build vs platform: the break-even math
On our model, custom pays back in about 21 months at 50,000 minutes a month, 9 months at 100,000 and under 4 months at 250,000; below 20,000 minutes a platform wins.
A custom build trades a one-time cost for a lower per-minute rate. The question for the CFO is how many minutes it takes to earn back the difference.
Our worked assumptions: platform at $0.11 a minute (Retell's calculator default), custom at $0.05 (LiveKit's default stack), an extra $50,000 to build custom versus building the same integrations on a platform, and $667 a month more in maintenance for the custom system.
Minutes per month | Platform usage per month | Custom usage per month | Net monthly saving | Payback on $50,000 |
10,000 | $1,100 | $500 | Negative | Never |
20,000 | $2,200 | $1,000 | $533 | About 94 months |
50,000 | $5,500 | $2,500 | $2,333 | About 21 months |
100,000 | $11,000 | $5,000 | $5,333 | About 9 months |
250,000 | $27,500 | $12,500 | $14,333 | About 3.5 months |
All figures are our arithmetic on Nov 2026 list prices. Swap in your own quotes; the shape of the curve holds.
The rule of thumb that falls out: below about 20,000 minutes a month, a platform almost always wins on cost. Above about 100,000, a custom build usually pays back inside a year. Between the two, the decision turns on things other than price.
Three conditions override the volume math entirely:
Compliance a platform can't meet, such as data that must stay in your own cloud or region.
Integration depth a platform's connectors won't reach, such as write-backs into a legacy system.
Voice as part of your product, where you need to own the agent, its data and its roadmap.
Total cost of ownership over 3 years
At 100,000 minutes a month, a custom build costs about $320,000 over 3 years against about $462,000 on a platform; at 20,000 minutes the platform is about $31,000 cheaper.
Executives approve budgets, not per-minute rates. Here's a 3-year view of the same model.
3-year cost | Platform at 20,000 min/month | Custom at 20,000 min/month | Platform at 100,000 min/month | Custom at 100,000 min/month |
Build or setup | $30,000 | $80,000 | $30,000 | $80,000 |
Usage (36 months) | $79,200 | $36,000 | $396,000 | $180,000 |
Maintenance (36 months) | $36,000 | $60,000 | $36,000 | $60,000 |
Total | $145,200 | $176,000 | $462,000 | $320,000 |
Assumptions: platform build includes the same integrations ($30,000); custom build $80,000; maintenance $12,000 a year on a platform and 25% of build cost a year for custom. Usage at $0.11 and $0.05 a minute.
For scale, the same 100,000 minutes handled by people would cost about $48,000 a month in wages alone at the BLS median, or roughly $1.7 million over 3 years. Gartner forecast that conversational AI would cut contact center agent labor costs by $80 billion in 2026, while expecting only about 1 in 10 agent interactions to be automated. The realistic case is AI handling the routine share and making people faster on the rest, not replacing a team outright.
What the quote usually leaves out
Four costs arrive after launch: model and API usage, hosting, maintenance at 15% to 30% of build per year, and the eval runs during tuning.
These are the most common gaps between a quote and the real budget:
Usage. At $0.05 a minute, 50,000 minutes is about $2,500 a month before anything else.
Hosting and infrastructure. A managed cloud costs less to run than a dedicated or self-hosted deployment. The compliance rule that forces self-hosting also raises running cost.
Maintenance. Providers change APIs, models get retired and carriers change behavior. Softcery plans 15% to 30% of build cost a year for stable agents and 30% to 60% for regulated or fast-changing ones.
Tuning and testing spend. Running test suites against recorded calls during tuning is a real build-phase cost. Hosted test tools help: LiveKit lists audio simulations at $0.05 a turn on paid plans.
Phone numbers and messaging registration. Small, but they add up across locations and countries.
A firm that won't estimate your running cost is quoting you half a number.
What each executive should check before signing
The CEO owns the build-or-buy call and ownership terms, the CTO owns architecture and lock-in, the CFO owns TCO and payback, and the COO owns containment and escalation.
Role | The question to answer | What to ask the vendor |
CEO | Is voice a feature we buy or a capability we own? | What do we own at the end: source code, infrastructure, prompts, data? |
CTO | Will this hit sub-second latency and survive provider changes? | What is the measured reply time on a real phone line? Can we swap STT, LLM and TTS providers? |
CFO | What is the 3-year TCO and payback? | Weeks and weekly rate behind the quote; running cost at our volume; maintenance terms |
COO or Head of CX | How many calls will it resolve, and what happens to the rest? | Target containment rate, escalation design and how results are reported |
CISO or compliance | Can the data path meet our obligations? | Every vendor in the chain, their BAAs or DPAs, retention and redaction settings |
On accounting, a custom build may qualify for capitalization as internal-use software under US GAAP (ASC 350-40), while platform fees are operating expense. That changes how the cost shows on the P&L. Confirm treatment with your auditor before you model it.
Custom build options compared
Platforms win on speed and low volume, in-house teams win on long-term control if you can hire, and a specialist build partner sits between the two.
1. Hosted platform (Retell, Vapi)
Configure an agent in a dashboard or through an API and pay per minute. Retell bundles more out of the box; Vapi gives developers more control over each component.
Best for: Low to mid volume, standard call flows and teams that need to launch in weeks.
2. Specialist build partner (RTC LEAGUE)
A team that designs, builds and hands over a custom voice stack on open infrastructure such as LiveKit, with your integrations and compliance rules built in.
Best for: High volume, deep integrations, compliance constraints or voice as part of your product, without hiring a full real-time team.
3. In-house team
Hire engineers with WebRTC, telephony and LLM experience and build it yourself.
Best for: Companies where voice is the core product and the budget allows a permanent team of 3 or more specialists.
4. General software agency
A broad development firm that adds voice AI to its services.
Best for: Simple, single-flow agents where real-time tuning isn't critical.
How to get a quote you can compare
Send every vendor the same written brief with named systems, intents, volumes and compliance rules, and ask for weeks, rate and running cost.
Name every system the agent touches and whether it reads, writes or both.
List every intent the agent must handle at launch, and what's out of scope.
Give your volumes: minutes a month, average call length and peak simultaneous calls.
State compliance rules up front: HIPAA, PCI, GDPR, recording consent, data location.
Ask what you own at the end: code, infrastructure, prompts, voice and data.
Ask for weeks and weekly rate behind the total.
Ask for running cost at your volume and a maintenance proposal for year one.
Where RTC LEAGUE fits
We build the custom side, so we'll be direct about when that's the wrong choice.
A custom build earns its cost when you need integration depth platforms don't reach, compliance that rules out shared infrastructure, or volume high enough that per-minute fees become the biggest line. That's the work our AI voice platform and WebRTC development practice covers, with CRM integration built in because integration is usually why a platform stopped being enough.







