AI agent development: voice and chat agents built on your data and run in production
We build voice and chat agents on your own data and then operate them. Pricing starts at $497 for a one-day build session, $4,950 plus $1,997 per month for a done-for-you production agent, or $9,950 plus $4,997 per month at multi-location scale. Live in 21 days.
What you get
A phone number that answers — A provisioned line, or a port of your existing number, wired to the agent with a warm-transfer path to a human and a voicemail fallback that still writes a record.
Function calling against your systems — The agent does not just talk. It dials, books, looks up a record and writes an outcome — the same four function classes running on our own production phone agent today.
A knowledge base built from your material — Your pricing, service area, hours, policies, objection handling and escalation rules, indexed and retrieved at conversation time rather than stuffed into a prompt.
Multi-provider model fallback — Anthropic first, then Groq Llama 3.3 70B, then Gemini 2.5 Flash. A provider outage degrades quality for a few turns instead of dropping the call. This chain runs in our own production stack.
An hourly health watch — A scheduled job calls the agent's health surface every hour and alerts on failure. Our own phone agent has run behind this watch since it went live.
Full transcripts and call outcomes — Every conversation stored with its transcript, the functions it called, the outcome it wrote and the recording where consent law permits it. Exportable at any time.
An evaluation suite before launch — Scripted conversations across the easy path, the hard path, the hostile caller, the out-of-scope request and the caller who switches language. Failures go back into the build, not into production.
A weekly tuning loop — We read the transcripts, find the turns where the agent hedged or misrouted, and change the prompt, the knowledge base or the function contract. This is the part nobody sells and everybody needs.
Escalation and handoff rules you approve — Written in plain English before the build starts: what the agent must never say, what it must transfer, and what it must log and end.
Published pricing
Agent in a Day — $497 one-time. A four-hour working session. You leave with a deployed agent on a real number and the artifacts to run it yourself. Timeline: One day.
Operators (done-for-you) — $4,950 build then $1,997 / month. Voice plus chat, function calling into your systems, evaluation suite, monitoring, weekly tuning. 5,000 minutes included. Timeline: 21 days.
Scale (multi-location) — $9,950 build then $4,997 / month. Multi-location or multi-brand routing, a dedicated engineer, deeper integration work, 25,000 minutes included. Timeline: 21 days.
Agent embedded in a larger build — Quoted inside the project then Folded into the retainer. When the agent is one component of a custom software or automation engagement rather than the product itself. Timeline: Set by the project.
How the work actually runs
01. Scope and ground truth (Days 1–3) — A working session with whoever actually answers the phone today. We record how they answer, what they get asked, where they escalate and what they are not allowed to say. We pull the last stretch of call history and the systems the agent will need to read and write. Output: A written spec: intents, functions, escalation rules, the things the agent must never do, and the definition of a successful call.
02. Build (Days 4–12) — Knowledge base indexed from your material. Functions implemented against your booking tool, CRM or scheduler with real credentials in a sandbox. Prompt and dialogue policy written against the spec. Model fallback chain wired: Anthropic, then Groq Llama 3.3 70B, then Gemini 2.5 Flash. Output: A working agent on a test number, callable by you, with transcripts and function traces you can read.
03. Evaluate (Days 13–17) — Scripted conversations across the easy path, the hard path, the hostile caller, the out-of-scope request and the language switch. We score booking accuracy, escalation appropriateness, factual accuracy against the knowledge base, and tone. Anything under threshold goes back to stage two. Output: An evaluation report you read before launch, including the calls the agent handled badly and what changed.
04. Launch (Days 18–21) — The agent goes onto the real number, usually behind a rollout you control — after hours first, then overflow, then front line. Monitoring and the hourly health watch go on at the same time, not later. Output: A live agent, an alerting path, and a documented rollback to your existing phone tree.
05. Operate (Ongoing, monthly) — We read transcripts every week, fix the turns that went wrong, add to the knowledge base as your business changes, and report on volume, outcomes and escalation rate. Model and provider changes are our problem, not yours. Output: A monthly written report and a changelog of every modification made to your agent.
Who this is for
Businesses losing revenue to unanswered calls, slow callbacks or after-hours silence.
Teams with a real system of record — a CRM, a booking tool, a scheduler — that an agent must read and write.
Operators who want the agent watched and tuned by someone else, not handed over as a prompt in a text file.
Who this is not for
Anyone wanting a chatbot bolted onto a website with no connection to a backend. That is a $99/month product, not a build.
Regulated call flows requiring a human in the loop on every interaction — the economics do not work.
Teams who want to own and self-operate the runtime from day one. We will happily scope that as custom software instead.
The receipts behind those claims
4 function classes in production — Our own production phone agent dials, books, looks up records and writes call outcomes. Those are the four function classes we build for clients, running against our own business first.
3 model providers in the fallback chain — Anthropic, then Groq Llama 3.3 70B, then Gemini 2.5 Flash. A single-provider agent is a single point of failure attached to your phone number.
Hourly automated health watch — A scheduled job exercises the agent's health surface every hour and alerts on failure, so an outage is discovered by a job rather than by a customer.
58 repositories in the operator's portfolio — The team that builds your agent has 58 repositories behind it, including one client engine carrying 54 edge functions and 40 scheduled jobs in production.
What is an AI agent, as opposed to a chatbot?
An AI agent takes actions in systems outside the conversation. A chatbot produces text; an agent produces a booked appointment, an updated CRM record, a transferred call or a logged outcome. The difference is function calling: a defined contract between the model and your software, with arguments the model must fill correctly and a result it has to reason about. That distinction is also where the cost sits. Writing a good prompt takes an afternoon. Making an agent write correctly into a booking system that rejects overlapping slots, handles a customer who gives a date in three different formats, and fails safely when the API is down takes the rest of the engagement. Vendors who quote a few hundred dollars are quoting the afternoon.
Should you start with a voice agent or a chat agent?
Start with voice if you miss calls. A missed call is a lost customer with a measurable dollar value, the failure is obvious to everyone in the business, and the agent proves itself in the first week. Start with chat if your volume arrives by form, web widget or SMS and your problem is response latency rather than capacity. Most businesses end up running both against a single shared knowledge base, which is the argument for building the knowledge base properly the first time. The channel is a surface; the expensive asset underneath is the indexed knowledge and the function contracts, and both are reusable.
What actually goes wrong with AI voice agents in production?
Four things, in order of frequency. Provider outages, which is why we run a three-provider fallback chain rather than a single vendor. Knowledge drift, where the agent quotes a price you changed two months ago, which is why the tuning loop is part of the monthly fee rather than an upsell. Function failures, where the booking API rejects a write and the agent tells the caller it succeeded, which is why we insist on a failure path that escalates rather than reassures. And silence, where the agent has been down for a day and nobody noticed, which is why the health watch runs hourly. None of these are exotic. They are the ordinary failure modes of any production integration, and they are the reason an agent is an operations commitment rather than a purchase.
We tried an AI phone agent and it was embarrassing.
Almost certainly a no-code agent with a prompt and no functions, no evaluation suite and no tuning loop. It could talk but could not do anything, so it stalled the moment a caller asked for something real. The fix is not a better prompt — it is function calling into your actual systems, an evaluation pass before launch, and someone reading the transcripts every week. That is what the monthly fee buys.
$1,997 a month is more than a part-time receptionist.
It is, in some markets. The agent answers every call simultaneously at 2am on a holiday, and it does not resign. If your call volume is low and predictable and your hours are fixed, a person may genuinely be the better purchase, and we will say so on the call. The break-even is usually somewhere around consistent after-hours volume or more than one call arriving at once.
How much does an AI voice agent cost?
TrainYourAgent publishes three prices: $497 for a one-day build session, $4,950 to build plus $1,997 per month to operate a done-for-you agent with 5,000 minutes included, and $9,950 plus $4,997 per month for multi-location deployments with 25,000 minutes included. The monthly fee covers hosting, model spend, monitoring and weekly tuning.
How long does it take to build an AI agent?
Roughly 21 days from kickoff to a live line on the done-for-you tiers: three days of scoping, about nine days of building, five days of evaluation and the remainder for a controlled launch. The one-day session compresses this into four hours by narrowing scope to a single call flow.
Can an AI agent book appointments in our existing calendar?
Yes, provided the calendar or booking tool has an API or a supported integration. Booking is one of the four function classes we implement as standard, alongside outbound dialling, record lookup and writing a call outcome. Booking against a system with no API is possible but adds scope, and we price that in the proposal rather than discovering it mid-build.
What happens if the AI model provider goes down?
The agent falls through a chain of providers: Anthropic first, then Groq Llama 3.3 70B, then Gemini 2.5 Flash. A single provider outage degrades response quality for a few turns instead of dropping the call. The same chain runs in TrainYourAgent's own production stack.