You train an AI voice agent on your business by writing a knowledge base (hours, prices, services, policies, FAQs), rules for what it may and may not do, and handoff instructions, then testing with real calls. Running cost is about $0.07 to $0.31 a minute plus telephony; 150 to 400 short calls a month comes to roughly $50 to $150.
"Trained" is a loose word here. In nearly every small-business setup the underlying model is untouched; the agent reads your documents on each call. That is good news, because you can improve it by editing a document rather than commissioning anything, and because the quality of the result depends on how well you write down what you already know.
What the agent is made of
A voice agent is a chain of four parts. Speech-to-text turns the caller's words into text. A language model reads that text, your instructions and your knowledge base, and decides what to say or do. Text-to-speech turns the reply into a voice. A phone line connects the caller. Platforms such as Vapi, Retell AI and Bland AI bundle some or all of these, and their price pages show which parts are inside the headline rate and which are billed separately. That distinction is where most surprises in the monthly bill come from.
Your job is the middle part: instructions plus knowledge. Everything else is configuration.
Step 1: write the knowledge base
Start with the questions callers really ask. Pull the last 50 voicemails or call notes, or ask the person who answers the phone to list the ten questions they hear daily. Group them into five sections and write each in short, plain sentences the agent can read aloud without sounding like a leaflet.
- Who we are and where calls go. Business name as spoken, what you do in one sentence, which services you do not offer, opening hours including public-holiday exceptions.
- Prices and what they include. Call-out fees, hourly rates, minimum charges, deposits, and what you will never quote by phone.
- Policies. Cancellations, warranties, payment methods, areas covered described as a list of towns or districts you choose.
- Process. What happens after a booking: confirmation, arrival window, what the customer should prepare.
- Handoff and emergencies. The calls the agent must not handle, who takes them, and the wording it uses.
A filled-in extract
For an illustrative five-person plumbing firm, a section of the knowledge base might read as follows.
HOURS: Phones answered by a person Mon-Fri 8:00-17:30. Outside those hours the agent takes messages and books non-urgent visits.
CALL-OUT: Standard call-out visit $95, includes the first 30 minutes. After that $70 per hour, billed in 15-minute steps.
WE DO: leaks, taps, toilets, radiators, boiler servicing, water heaters.
WE DO NOT: gas appliance installation, drain surveys, commercial kitchens.
NEVER QUOTE a fixed price for a repair. Say: "A plumber needs to see it first. I can book a visit for $95."
EMERGENCY: if the caller mentions flooding, a burst pipe, or smell of gas, say: "Please turn off the water at the stop tap if you can, and if you smell gas leave the building and call the gas emergency number. I'm passing this to our on-call plumber now." Then transfer to the on-call number.Notice what the extract does. Each fact has a number or a condition. Each "never" has a replacement sentence so the agent is not left improvising. The emergency instruction is the longest, which is correct.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWrite it like speech
Read each entry aloud. Long lists of prices sound dreadful on a call, so give the agent the three prices callers ask about and let it offer to text the rest. Spell out how to say addresses, times and numbers ("half past five", "ninety-five dollars"). Keep each item under about 60 words, because long blocks are the ones agents blur together. If you already have a written FAQ for a chatbot, adapt it; the guide on building the FAQ your chatbot needs covers the question-gathering.
Step 2: write the instructions
Separate from the facts, the instruction prompt tells the agent how to behave: who it is, what it may do (answer questions, take messages, book visits, transfer), what it may not, and how to talk. A compact, working structure:
You are the phone assistant for [business]. You answer callers politely in short sentences.
You may: answer from the knowledge base, take a message, collect name, number and address, book a non-urgent visit using the booking tool, transfer to a person.
You must not: quote repair prices, give advice that could cause harm or damage, discuss other customers, promise arrival times that are not in the booking tool.
If you do not know the answer, say so and offer a message or a call back. Never guess.
If the caller is angry, asks for a manager, or mentions an emergency, transfer at once.
At the start say: "You've reached [business]. I'm an automated assistant. How can I help?"The last line matters. Telling callers they are talking to an automated assistant is courteous, avoids complaints and is a legal requirement in some places, so check the rules for your callers' areas before launch. Also read your vendor's terms on recording, and tell callers if calls are recorded.
Step 3: choose a platform and price it
List prices below come from vendor pages read in November 2026. They change, and the cost per minute depends on which models and voices you pick.
| Platform | Headline price | What is and is not inside |
|---|---|---|
| Retell AI | $0.07 to $0.31 a minute for voice agents; $10 in free credits for new pay-as-you-go accounts | 20 concurrent calls free on every account; phone numbers listed at $2 a month; its own telephony is charged per minute by country (the first rate shown is $0.015) |
| Vapi | Platform fee of $0.05 a minute pre-paid, plus provider costs at pass-through; $5 of free credit to start | Speech-to-text about $0.0095 to $0.0099 a minute, language model $0.0077 to $0.0452, voice $0.0146 to $0.0238, telephony extra (Twilio inbound $0.008) |
| Bland AI | Start plan $0.14 a minute, no monthly fee; Build plan $0.12 a minute plus $299 a month | Rate covers language model, speech-to-text and text-to-speech; telephony billed separately; transfers billed at $0.05 or $0.04 a minute; Start caps calls at 100 a day |
| ElevenLabs | Its API pricing page lists $0.08 a minute for additional minutes on its Speech Engine agent pipeline, beyond each plan's included minutes | Burst minutes above a plan's concurrency limit are $0.16; check what the language model adds and which plan fits your volume |
Vapi's own components give a way to build a rough all-in figure: $0.05 platform + about $0.01 speech-to-text + $0.008 to $0.045 for the language model + about $0.015 to $0.024 for the voice + $0.008 telephony = roughly $0.09 to $0.14 a minute. That is a sum of listed ranges, not a quote, and your real number depends on the models chosen. Synthflow, whose pricing page says it is sold only as an annual enterprise agreement with no free plan, trial or per-minute rate, is the kind of product to skip at small scale.
The knowledge base itself can carry a charge, which matters when the whole point is an agent that answers from your own documents. Retell lists an extra $0.005 a minute when an agent uses a knowledge base, with ten knowledge bases included free and $8 a month for each one after that. Bland's Start plan allows ten knowledge bases and Build allows 50. One well-organised knowledge base per agent is usually enough for a small business, so these limits rarely bite, but the per-minute add-on belongs in your sum.
Example: monthly running cost at three volumes
Assume an average call of 2.5 minutes and an all-in rate of $0.12 a minute, a mid-range figure that you should replace with your vendor's quote. Illustrative.
| Calls a month | Minutes | Usage at $0.12 | Number and extras (assumed $10) | Monthly total |
|---|---|---|---|---|
| 150 | 375 | $45 | $10 | $55 |
| 400 | 1,000 | $120 | $10 | $130 |
| 1,000 | 2,500 | $300 | $10 | $310 |
At 400 calls that is about $0.33 a call. If the agent saves a person eight hours of phone time a month at $25 an hour, that is $200 of time against $130, with the bigger value in calls caught after hours. If it books two extra jobs a month worth $95 call-outs, it has paid for itself again. But a platform with a $299 monthly fee, such as Bland's Build plan, costs more than that on fees alone at 150 calls, so a minimum fee changes which plan is cheap.
Setup costs people forget
Your time to write the knowledge base (a day), a number or porting your existing one, a booking-system connection if the agent books visits, a second language if you need one, and a week of testing. If you hire someone for setup, ask for the knowledge base in a document you own, so you can move platform later without rewriting it.
Step 4: test calls before any real caller
Run at least 25 calls from a phone that is not yours. Cover easy questions, hard ones and awkward callers. A solid test set:
- Opening hours on a public holiday.
- Price for something you do not do.
- A caller who gives a price and demands the agent agrees to it.
- A caller who talks over the agent and changes their mind mid-sentence.
- A caller with a strong accent or a noisy line.
- A flooding emergency, with the transfer checked live.
- A question about another customer's visit.
- A request for a manager.
- A caller who goes silent for ten seconds.
- A call asking something the knowledge base does not cover.
Record each call, mark pass, partial or fail, and write the cause: missing fact, bad instruction, wrong transfer, speech misheard. A sector-specific test set you can adapt is in a 15-call test for plumbers.
A before and after from one failed call
Test call 3: the caller says, "Your website says $60, so you'll do it for $60." The agent replies, "Yes, we can do that." Cause: the knowledge base lists an old price and the instructions never said how to handle price disputes. Fix: correct the price, then add, "If a caller quotes a price that differs from the knowledge base, do not agree. Say the current price from the knowledge base and offer a call back from the office." Re-run the call until it passes twice in a row.
Step 5: handoff rules and what the human sees
Handoffs fail when the person who picks up has no context. Specify what the agent passes on: name, number, address, reason for calling, what was already said, and the time. Specify where: a transfer to a live number when someone is in, a message to a phone when not. And specify what counts: any emergency word, any refund or complaint, any caller who asks twice for a person, any question where the agent has said "I don't know" twice. Hotels have long experience here, summarised in which calls to hand over and which to keep.
A second knowledge base: a cleaning company
The plumbing extract leans on emergencies. A company with a different rhythm needs different content. Take an illustrative eight-person cleaning company serving homes and small offices. Its knowledge base has more about scope and scheduling than about risk.
SERVICES: regular home clean (weekly or fortnightly), one-off deep clean, end-of-tenancy clean, small-office clean.
PRICING: Home clean from $32 an hour per cleaner. Deep clean priced per property after a short form. Never quote a total on the phone; say "I can have our office send a written quote today."
INCLUDED: our own equipment and products. Not included: oven interiors on a regular clean (add-on $35), carpet shampooing, windows above ground floor.
BOOKING: minimum 2 hours. New customers book a first visit within 5 working days. Cancel 24 hours ahead or we charge half a visit.
KEYS AND ACCESS: never take key details by phone; say a team member will arrange access securely.
HANDOFF: any complaint, damage report, request to change a regular cleaner, or request for a refund goes to the office manager.Two things stand out. The knowledge base tells the agent what it must not collect (key codes), which is as important as what it may. And the money talk is deliberately limited: the agent names an hourly rate but does not total a job. Where a company can only quote after seeing a property, an agent that gathers the details and promises a written quote does more good than one that guesses.
Mistakes that appear in the first week
Contradictory sources
The agent is told the call-out fee is $95 in one section and $85 in another, because the website and the price sheet disagree. It picks one at random. Search your documents for each number before you load them, and make one document the master.
Instructions that argue with the knowledge base
"Always be helpful and give the caller an answer" next to "never quote prices" produces a confident invented quote. Keep the rules short, consistent, and phrase refusals as a replacement action: not "do not quote", but "if asked for a price for a repair, offer the visit fee and a booking".
Treating silence as consent
Callers pause while they find a pen. Agents may interpret the pause as the end of the call or talk over the caller. Test pauses specifically, and adjust the platform's response delay or end-of-call settings, which most platforms provide, until it waits a natural beat.
Loops and long calls
A call that loops, for example an agent that keeps asking for a postal address it cannot parse, costs minutes and annoys the caller. Add a rule: after two failed attempts, take a message or transfer. Check the longest calls each week, since they are both the costliest and the most likely to be going wrong.
Contract points that change the real price
Before paying anyone for setup or committing to a plan, ask for the following in writing: what happens to your recordings and transcripts and how long they are kept; whether your calls can be used to train anyone's models; what happens to the knowledge base if you leave; whether minimum monthly fees apply, which matters at low volume; how overage is charged when you exceed a plan's included minutes; and what the price is per concurrent call if two customers ring at once. The Retell page, for example, includes 20 concurrent calls free and lists extra capacity at $8 per concurrent call a month; Bland's plans cap concurrent calls and calls per day by plan. At small volume you will rarely hit these limits, but a seasonal spike might.
Some sectors have further rules about recording and disclosure of automated calls, and health, legal and financial businesses have additional duties around the data a caller shares. For those, read the sector tutorials, such as voice agents for law firms and voice agents for a dentist, and take advice before going live.
Keeping the knowledge base true
A voice agent confidently repeats whatever you wrote last spring. Give the knowledge base an owner and a review date. Put the version and date at the top of the document, and add a line to your price-change routine: when a price, hour or policy changes, edit the knowledge base the same day. Once a month, call the agent and ask five questions whose answers changed recently. A quick way to catch drift is to keep a short "changes log" at the bottom of the document: "3 Nov: call-out fee raised from $85 to $95; winter hours start 1 Dec." When a caller reports a wrong answer, add a test call for it, fix the document and keep the test, so the same mistake cannot return quietly.
Seasonal content needs the same care. Holiday opening hours, promotions and temporary service limits are the facts callers ask about most and the facts most likely to be stale, so schedule their updates a week before they apply and remove them a day after they end.
A two-week rollout plan
Days 1 to 2: collect the top 20 caller questions and write the knowledge base. Day 3: write instructions and load both into the chosen platform with a test number. Days 4 to 6: run the 25-call test and fix the failures. Day 7: have a colleague who did not write it make ten calls and try to break it. Days 8 to 10: switch on the agent only for after-hours calls, with a message handed to the office each morning. Days 11 to 14: review every call, count transfers and missed messages, fix the document and decide whether to extend it to daytime overflow. Treat that second fortnight as the real test, because genuine callers are less predictable than any script.
How to check it is working in the first month
Listen to a sample of ten calls a week for four weeks. Count calls handled without a person, calls transferred, and calls where a caller hung up within 20 seconds, which often signals confusion. Review every transfer for whether it should have happened. Track missed jobs: if a caller left a message that nobody returned, the agent has made the problem worse. Update the knowledge base weekly from what you hear, and date every change. After a month, the number to compare is cost per handled call against the time you spent supervising it.
If the supervision time exceeds the time saved, narrow the agent's job. A voice agent that only takes messages and answers hours and prices is a good first version, and you can widen it as the test results support it. For what a month of calls costs when priced several ways, see one month of voice agent cost priced five ways. If you would rather map the calls with someone, that is what my AI implementation consultation covers.
Questions before you train a voice agent
Do I need to train a model on my business data?
No. Most small-business voice agents do not retrain a model. You give the agent a written knowledge base and instructions, and it reads them during each call. Better documents, not model training, are what improve answers.
How long does setup take?
Allow a day to write the knowledge base, a day to configure and connect a number, and a week of test calls and fixes before real callers. The writing is the slow part, and it is where answers get their accuracy.
What should the agent never answer?
Anything it cannot verify from your documents: quotes that depend on inspection, medical or legal advice, refunds and complaints. Write these as handoff rules so the call goes to a person with a short summary.
Is the per-minute price the full cost?
Rarely. The headline rate may exclude telephony, extra language-model usage, a phone number and any platform fee. Price a month from your own call volume and average call length, and ask the vendor to confirm each line in writing.
Further reads
- AI Voice Agent for Contractors: What It Handles and What It Hands Off — What a contractor's agent handles and what it hands off.
- Logistics AI Voice Agent: Setup, Call Flows and Monthly Cost — Call flows and monthly cost for status calls.
- AI Voice Agent for Salons: Booking Rules to Set Before It Answers — Booking rules to set before the agent answers.
- Can an AI Voice Agent Book Car Dealership Service Appointments? — Can an agent book service appointments?
- What an AI Agent Can Do for a Plumber After the Call Ends — What the agent can do once the call ends.
- AI Voice Assistant for Tree Surgeons: What It Can Book and Quote — What a voice assistant can book and quote.
- Restaurant AI Voice Agent: 12 Test Calls to Run Before Go-Live — Twelve test calls to run before going live.
- AI Voice Agent for Ecommerce: Cost per Order-Status Call — Cost per order-status call from listed per-minute prices, set against a part-time hire, with lookup limits and handover rules.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Vapi, Retell AI, Bland AI and Synthflow pricing pages, ElevenLabs API pricing page, read in November 2026.