What is an AI agent for a small business?
An AI agent for a small business is software that answers your calls, texts and web chats in plain conversation, asks the questions your best salesperson would ask, and books qualified people onto your calendar. It works nights and weekends and writes everything it learns into your CRM.
I've been running ads for businesses since 2013, and the leak I see most often sits after the ad. A lead raises their hand and nobody gets back to them for hours. When Harvard Business Review researchers audited 2,241 US companies, only 37% responded to a web lead within an hour and the average response time was 42 hours. Firms that reached out within an hour were nearly seven times as likely to qualify the lead as firms that waited one hour longer (summary of the study's figures). An agent closes that gap without you hiring a night shift.
What job should your first AI agent do?
Give your first agent one job: answer new inquiries, qualify them, and book the qualified ones. Leave support tickets, billing questions and upsells for later. A narrow agent is easier to script, easier to test, and easy to judge by one number, booked appointments that show up.
Most owners want the agent to do everything on day one. That's how you end up with a bot explaining your refund policy while a hot lead waits on hold. Pick the channel where you lose the most leads today. For a roofer or a med spa that's usually missed calls. For a coach running lead ads it's usually form fills that sit untouched until the next morning.
How do you build an AI agent step by step?
You build a small business AI agent in seven steps: pick the job, define a qualified lead, write the script, connect your calendar and CRM, set handoff rules, test it with real scenarios, and launch on one channel. None of it needs code if your platform has a visual builder.
- Pick the job and the channel. Inbound calls, missed-call text back, web chat, or new form leads. One to start.
- Define a qualified lead. Write down three to five facts that make someone a fit: the service they need, where they are, their timeline, a budget range, and whether they make the decision.
- Write the script. The next section covers how.
- Connect your tools. Calendar, CRM, phone number and text messaging, so the agent can book and confirm instead of only talking.
- Set handoff rules. Decide when a human takes over and how the agent alerts that person.
- Test it. Run 20 to 30 scripted conversations, including rude, confused and off-topic callers.
- Launch on one channel. Read every transcript for two weeks, fix what breaks, then add the next channel.
How do you write the prompt and script for an AI agent?
Write the prompt like a job description for a new hire: who they are, which business they work for, what a good outcome looks like, the exact questions to ask in order, what they must never say, and when to hand off. Use short, plain sentences, because the agent copies your tone.
A prompt that books well usually has these parts:
- Role. "You are Maya, the front desk for Smith Roofing in Tulsa."
- Goal. "Find out if the caller needs a roof inspection, and if they qualify, book one."
- Questions in order. One at a time, with what to do after each answer. A caller outside your service area gets a polite no and a referral.
- Facts it may share. Service area, hours, and price ranges you're comfortable quoting on the phone.
- Hard limits. No legal or medical advice, no discounts, no promises about insurance claims.
- Confirmation. Repeat back the name, phone number, address, date and time, in the caller's time zone, before booking.
- Handoff. "If the caller is upset or asks for a person, say the owner will call back within 15 minutes, and create a task for the owner."
Write the opening line yourself and keep it under ten words. "Smith Roofing, this is Maya, how can I help?" sounds like a front desk. A long intro packed with disclaimers sounds like a robocall, and people hang up on robocalls. If someone asks whether they're talking to AI, the agent should say yes. Outbound calling has its own consent rules, which I cover in what the FCC allows when AI agents call your leads.
The qualifying questions are where most of the value sits. For the full method, read how AI lead qualification works.
What does an AI agent need to connect to?
An AI agent needs four connections at minimum: a phone number or chat widget to talk through, your live calendar to book into, a CRM to store each contact and conversation, and text or email to send confirmations. Without the calendar and CRM, it can hold a conversation but it can't finish the job.
| Connection | What it does | What breaks without it |
|---|---|---|
| Phone number or chat widget | Lets the agent answer calls, texts and site chats | Leads still hit voicemail or an empty form |
| Live calendar | Books into real open slots and blocks double bookings | The agent promises times you don't have |
| CRM | Saves the contact, answers, transcript and tags, and starts follow-up | Leads disappear after the conversation ends |
| Text and email | Sends the confirmation, calendar invite and reminders | More people forget and no-show |
| Human alert | Texts or assigns a task to a person when the agent hands off | Upset callers wait for a callback that never comes |
Keep one calendar as the source of truth. If your team books in two places, the agent will double-book you within a week.
How do you test an AI agent before customers hear it?
Test an AI agent by calling and texting it yourself from a written list of 20 to 30 scenarios: the ideal buyer, the caller who only wants a price, someone outside your service area, and someone who gets annoyed. Read every transcript and fix the prompt before a single real lead reaches it.
Scenarios I always include:
- A perfect lead who wants to book today
- Someone outside your service area
- A price shopper who won't answer questions
- A caller who changes the appointment time twice
- A lead in a different time zone
- Someone who asks, "Am I talking to a robot?"
- An existing customer calling with a complaint
On each test, check three things. Did it ask every qualifying question? Did the booking land on the right day and time in the calendar? Does the CRM record match what was said? Run the time zone test every time you change the script. In our own calendar data, one no-show traced back to an appointment booked in the wrong time zone.
How much does it cost to build and run an AI agent?
Running a voice AI agent costs cents per minute in usage, plus your platform fee and setup time. Retell AI, a voice agent provider, publicly lists $0.07 to $0.31 per minute for AI voice agents on pay-as-you-go. Text and chat agents usually cost less because no voice is generated.
Usage is rarely the big line item (see Retell's pricing page for the per-minute breakdown). The real cost is setup: writing the script, connecting tools and reading transcripts. Build it yourself and budget a couple of weekends. Have it done for you and you're paying for someone else's weekends. I go through real monthly numbers and the break-even math in what AI voice agents cost and whether they're worth it.
What mistakes do businesses make with their first AI agent?
The common mistakes are giving the agent too many jobs, letting it book without the live calendar, hiding the way to reach a human, skipping time zone confirmation, and never reading transcripts after launch. Each one shows up as lost bookings or no-shows within the first two weeks.
- Too many jobs. An agent that books, troubleshoots and sells add-ons does all three badly. Add jobs one at a time.
- No way to reach a person. A Gartner survey of 5,728 customers found 64% would prefer companies didn't use AI for customer service, and their top concern was that it would get harder to reach a person. Give callers a clear path to a human and you answer that concern directly.
- Booking days out. Have the agent offer the soonest open slot first. In our own calendar data, people who showed were mostly booked within 24 hours, while no-shows were mostly booked 2 to 5 days out.
- A stiff opening. If the first ten seconds sound like a recording, the caller is gone before the agent asks a single question.
- Set it and forget it. Read transcripts weekly. Every odd answer is a line to add to the prompt.
This is the setup we build at KenjiAI. Voice AI agents answer and call your leads, Conversation AI handles texts and chat, and both qualify and book into the same calendar and CRM, with confirmations and reminders sent automatically. If you'd rather skip the weekends, here's how our voice agents work.
Start this week with step two. Write down the three to five facts that make someone a good fit for your business, and the exact words you'd use to ask about each one. That page becomes your agent's script, whether you build it yourself or have us build it.