You have a phone agent, or you are about to get one, and the question that keeps you up is not “can it talk.” It can talk. The question is what it does when a real person with a real problem calls, and whether you will find out about it before the customer does.
The vendor showed you a friendly voice and a booked appointment. What they did not show you is the call where the agent quoted a price you never approved, or the one where a caller with a flooded basement was told to leave a message.
AI call intake is a discipline, not a feature. This article gives you the twelve rules that separate an intake agent you can trust from one you will switch off in a month. They are specific enough that your developer or vendor can build to them, and plain enough that you can check whether they did.
What this actually is
Call intake is the first two minutes of a phone call: finding out who is calling, what they want, how to reach them, and what should happen next. A human receptionist does it without thinking. An AI agent does it exactly as well as it was told to, and no better.
The analogy is a new employee’s intake sheet. Every business that takes calls has one, printed or in someone’s head: name, number, what is the problem, when do they need it, how did they hear about us. Whether it actually gets filled in on every call determines whether the call turns into work. An AI intake agent is that sheet, enforced on every call without exception, plus a rulebook for the moments the sheet does not cover.
The technology underneath is a language model (software that reads and writes text, the same family as the current Claude and GPT-class models) wrapped in speech recognition and a synthetic voice. The broader picture is in AI Phone Answering for Small Business: What Works in 2026. This article is about the rules.
The twelve rules of AI call intake
1. Answer within two rings and respond within a second
Answer on the first or second ring. Then keep the response gap, the pause between the caller finishing and the agent speaking, at around one second. Longer than that and callers start talking over the agent. Test this yourself with a stopwatch on a real call, at 5pm when the platform is busiest, not at 10am.
2. Say it is an assistant in the first sentence
The opener discloses, names the business, says what the agent can do, and offers a person. “Thanks for calling Ridgeline Plumbing, this is the automated assistant. I can take your details and get a plumber scheduled, or connect you to the office. What is going on?” Do not let it claim a human name. Do not let it dodge “are you a robot” with a joke. A straight “yes, I am the automated assistant, and I can get you to a person any time” ends that conversation well.
3. Capture the five fields, every call
Name. Callback number. What they need, in one sentence. When they need it. How they found you. Those five are the minimum for any intake, and the agent should not end a call without attempting all of them. In the systems we build, the instructions list the fields explicitly and the agent keeps asking, politely, until each is filled or the caller declines. A summary with a blank callback number is a lead you paid to lose.
Add fields for your business: service address for a trade, date of birth for a clinic, matter type for a law firm. Keep the total under eight. Every extra question is a chance for the caller to hang up.
4. Confirm back the details that matter
Speech recognition is very good and still wrong sometimes, especially on names, phone numbers and street addresses. The agent must read back the callback number and the address before ending the call, digit by digit: “I have 206-555-0147, is that right?” This single rule catches most transcription errors. It costs ten seconds. Skipping it means some of your callbacks go to wrong numbers and nobody knows why the lead went cold.
5. Never quote a price it was not given
A language model will invent a plausible price if asked, because it is built to produce plausible text. Your agent must have a written price sheet, short and explicit, and the instruction to quote nothing outside it. If you do not want prices quoted at all, the rule is “we quote after we see the job; I can book that visit now.” Test this rule by calling and asking for prices on things that are not on the sheet. If the agent produces a number, the build is not done.
This extends to timing promises (“someone will be there by noon”) and outcomes (“that should be an easy fix”). The agent does not know, so it does not say. The general principle of grounding, keeping the model to what it was actually given, is covered in AI Hallucination Guardrails for Business Applications.
6. Write the transfer triggers down, and make each one work
The agent needs a short list of conditions that put a person on the line. Ours usually looks like: caller asks for a person; caller is angry or upset; the call involves a complaint, refund, dispute or legal matter; the call describes an emergency (define it for your business: gas smell, water pouring, chest pain, a court deadline today).
For each trigger there must be a working path. During business hours that is a live transfer to a phone or ring group. After hours it is a callback commitment with the summary pushed to a phone someone actually looks at, and for emergencies, a direct transfer to the on-call cell. A trigger with no path behind it is a promise the business breaks. The mechanics are in AI Chatbot Escalation to Human: When and How the Handoff Works.
7. Fall back to voicemail deliberately, not by accident
Sometimes the agent cannot complete intake: the line is bad, the caller will not give a number, the platform has an outage. The fallback is a plain voicemail with a clear prompt, designed on purpose, not a dead line. Test what callers hear when the agent fails entirely. Many setups drop the call or play silence. Route the failure case to a normal voicemail box and have someone check it daily.
8. Record the transcript for every call
The transcript is the written record of what the caller and the agent each said. It must be stored for every call, searchable, and kept at least as long as a job might be disputed. It is your quality control, your material for improving the instructions, and your evidence when a customer says “your machine told me it was free.”
Check your state’s rules on call recording. Many states require one-party consent (only one participant needs to know) and some require all parties to be told. “This call may be recorded” in the opener handles the second case. Ask your vendor where transcripts are stored and who can see them.
9. Send a summary to the right person within a minute
The summary is what turns intake into action. Four to six lines: caller name, number, what they need, urgency, what the agent did (booked, took a message, transferred), and what a human should do next. It goes to the right person, not to a shared inbox where it dies. New sales leads go to whoever sells. Existing-customer issues go to whoever handles service. Emergencies go to a phone, by text, immediately.
Route by rule, not by hope. A summary can be posted into a CRM as a new lead, texted to a dispatcher, or emailed to the office, and the choice should depend on the intake fields. If your team lives in HubSpot or Jobber, that is where the summary belongs.
10. Book into the real calendar or do not offer booking
If the agent offers appointments, it must read live availability from the real calendar and write the booking back into it. An agent that “requests” a booking someone has to confirm by hand has just created a queue and a delay, which is the thing you were trying to remove. If the integration is not built yet, the honest instruction is “I will take your details and the office will call to schedule.” Details are in AI Appointment Scheduling and Reminders That Cut No-Shows.
11. Keep the instructions short and versioned
The agent’s behavior comes from a written set of instructions, often called a prompt. Short, numbered rules work better than long paragraphs. Every change should be saved with a date and a note about why, so that when the agent starts behaving differently you can see what changed. A vendor that lets anyone edit the live instructions in a text box with no history will eventually give you a Tuesday where nobody knows why the agent started saying you are closed.
12. Review twenty calls a week and update the sheet
Pick twenty transcripts at random each week and read them. Look for wrong answers, questions the agent could not handle, early hang-ups, and blank fields. Each is either a fix to the instructions or a gap in the material the agent was given. This takes about an hour and it is the difference between an AI call intake agent that improves and one that slowly rots as your business changes around it.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: a two-location physical therapy clinic with fourteen staff. The front desk takes around 80 calls a day, mostly scheduling, and the phones go unanswered whenever both receptionists are checking patients in. New patient inquiries go to voicemail and are returned the next morning, by which point many have booked elsewhere.
What gets built:
- An overflow agent that picks up after three rings and all calls after 6pm, opening with a disclosure and the offer of a person.
- Intake fields: name, callback number, existing or new patient, what hurts (one sentence), insurance carrier, preferred location. Callback number and date of birth are read back digit by digit.
- Booking reads live availability from the practice management system and writes the appointment back, with a confirmation text to the patient.
- Hard rules: no clinical advice of any kind, no statements about what insurance covers, no prices beyond the published self-pay rate. Any mention of severe pain, numbness, or a fall in the last day transfers to the clinic line or, after hours, reads out the urgent-care guidance and takes a message flagged urgent.
- Summaries go into the practice system as a note on the patient record, and new-patient summaries are also emailed to the office manager.
- Weekly review of twenty transcripts by the office manager, with instruction changes logged.
What changes: new patients get booked on the first call. The receptionists stop apologizing. And the owner has a weekly list of the questions callers asked that the agent could not answer, which turns out to be the best feedback the clinic has ever had.
What it costs to run
The intake agent itself runs on a voice platform such as Vapi, Retell or Bland, priced per minute of talk time. Expect somewhere between roughly $0.05 and $0.20 per minute all-in depending on the voice and model, plus a small monthly base. Check the current pricing pages. At 500 intake calls a month of two to three minutes each, that is roughly 1,000 to 1,500 minutes, which lands in the low hundreds of dollars per month.
The phone number and telephony are a few dollars a month through Twilio or bundled with the platform. Transcript storage is negligible. The summary routing and calendar integration might need a small server or an automation tool such as n8n or Make, at $10 to $50 a month. The largest ongoing cost is the weekly review hour, and it is the one line item you should not cut.
The mistakes we see most
-
No read-back. Callback numbers with a wrong digit, and nobody understands why callbacks fail.
-
Prices in the model’s imagination. The agent was never given a price sheet, so it made one up, confidently.
-
Summaries to a dead inbox. The agent worked perfectly and the lead sat unread in a shared mailbox for four days.
-
Emergency triggers with no path. “Someone will call you back” at 11pm to a caller describing water through the ceiling.
-
Instructions edited live with no history. The agent changes behavior and nobody can say what was changed or when.
When to bring in help
An owner can get a basic intake agent running on an off-the-shelf AI receptionist product in an afternoon: disclosure, message-taking, a few fixed answers, a summary by email. Try it on after-hours calls and read the transcripts for a month.
A developer earns their fee when the intake fields are specific to your business, when the summary has to land in your CRM or practice system as a proper record, when booking has to read and write a live calendar, when transfer rules depend on who is on call, or when recording rules and data handling need to be right for your industry.
Levelbrook builds AI call intake agents to these twelve rules, fixed price from a written scope, running in accounts you own so the number, the transcripts and the integrations stay yours. The form below is how a conversation starts.