The agent is a script with a voice, not an employee
ElevenLabs Agents (the product formerly called Conversational AI) is a platform for building voice agents: you pick a voice, write a system prompt, attach a knowledge base, and connect a phone number. When someone calls, the agent listens, thinks with a large language model, and speaks back in near real time.
It can transfer to a human, take a message, answer questions from documents you give it, and trigger tools like booking a slot or sending a text.
What it is not: a mind. It follows the rules you wrote and reads the documents you gave it. If your hours, service area, and pricing exist only in your head, the agent has nothing to work from. Writing those rules down is the actual project. The platform is the easy 20 percent.
For a small business, the use case that actually pays off is inbound reception: answering the calls you already miss. Missed calls are lost jobs, and an agent that answers at 11pm, takes a clean intake, and texts you a summary earns its keep without ever pretending to be a person.
500 calls a month costs about $135-150, all in
ElevenLabs publishes plan pricing, and it held up when we checked it in July 2026. The Pro plan is $99/mo with 1,238 call minutes included; extra minutes cost $0.08 each. The LLM and telephony are billed separately on top, based on usage.
Say you take 500 calls/mo averaging 3 minutes: 500 x 3 = 1,500 min/mo. Telephony through Twilio adds a local number at $1.15/mo and inbound minutes at $0.0085/min.
That is roughly $135-150/mo, or about $1,600-1,800/yr, before anyone builds anything. It matches what we see in production: a private-lender receptionist built this way runs about $2,050/yr all-in at 500 calls/mo, after a 35-vendor evaluation landed on this stack.
Lighter volume is cheaper. At 100 calls/mo (about 300 min), the $22/mo Creator plan plus overage and telephony comes in under $30/mo.
Two cautions: minutes are billed on conversation duration, so a rambling caller costs the same as a productive one (ElevenLabs does discount long silences), and if simultaneous calls exceed your plan's concurrency limit, burst pricing doubles the rate to $0.16/min.
The build is mostly writing down how you handle calls
A working receptionist build has 5 parts, and only 1 of them is technical:
- Call rules. What the agent may say, what it must never guess at, when it transfers, when it takes a message. This is a written document, and it is where most of the effort goes.
- A knowledge base. Hours, services, pricing boundaries, service area, FAQs. The agent answers from these files, not from the open internet. This is the same discipline as a business knowledge base: if it is not written down, the agent cannot know it.
- Telephony wiring. A number (new, or your existing line forwarded), connected through Twilio or ElevenLabs' native telephony.
- Escape hatches. Transfer to your cell, voicemail fallback, and a text or email summary of every call so nothing disappears.
- Test calls. Dozens of them, including hostile ones: callers who interrupt, mumble, ask for a discount, or try to get the agent to promise something. You fix the rules until it stops failing.
Budget the effort accordingly. The platform setup is an afternoon. Getting the rules right so the agent never quotes a price it should not is 1-2 weeks of drafting, testing, and revising.
This is exactly the kind of grounding work Bamboo Digital does in month 1, which is why our voice builds start from a knowledge base rather than a prompt.
Inbound and outbound are different legal worlds
In February 2024 the FCC ruled that AI-generated voices count as "artificial" under the TCPA. That means calls you place with an AI voice need the recipient's prior express consent, and marketing calls need prior express written consent.
Violations run $500-$1,500 per call, and plaintiffs' firms actively hunt for them. This is why the lender build linked above is inbound-only by design.
Inbound is much safer ground: when a customer dials your number, the TCPA robocall restrictions on artificial voices do not apply to that customer-initiated call. But 3 obligations remain:
- Disclose the AI. Have the agent say it is an AI assistant in its greeting. California's AB 2905 (effective January 2025) requires artificial-voice disclosure on autodialed calls, and several states are following. Disclosure also just works better: callers forgive a named robot and punish a fake human.
- Handle recording consent. If calls are recorded or transcribed (they usually are, for the summaries), all-party consent states like California, Florida, and Washington require telling the caller. A one-line notice in the greeting covers it.
- Never let the agent dial out. Keep outbound calling disabled until you have consent records and legal review. One "quick follow-up call" feature is how an inbound receptionist becomes a TCPA defendant.
Skip it if the math or the trust is not there
Do not build this if any of these are true:
- You get under ~5 missed calls a week. Voicemail plus a fast callback is free.
- Your calls are high-stakes conversations, not intake. If every call is a negotiation or an emergency, a wrong answer costs more than a missed one.
- Nothing about your business is written down yet. The agent will guess, and it will guess in front of customers. Fix the knowledge base first; see when a business is ready for agents.
- You cannot review transcripts weekly. An unwatched agent drifts, and you will not know until a customer tells you.
If you do 2-3x that call volume, miss a real share of it, and can write down your rules, the math is hard to argue with: about $150/mo against the first recovered job usually paying for the quarter.
