Field notes for teams handing work to AI.

A custom AI agent is software that does a job, not just a conversation. It reads what comes in, decides what to do using your rules, uses your tools to do it, and asks a person when the stakes are high. If a chatbot answers the question, an agent finishes the task the question was about.

Chatbot, assistant, agent: the real difference

The words get used loosely, so here is how we use them on every scope document we write.

Chatbot

Talks to your customers on one or more channels. It answers, qualifies and books, and it works inside the conversation.

Assistant

Helps one person inside a tool, such as drafting an email or summarising a call. It waits to be asked.

Agent

Runs a workflow end to end. An event triggers it (a new lead, an invoice, a form), it works across several tools, and it reports what it did.

Most businesses that ask us for "an AI" need one of the first two. The ones that need an agent usually describe a person on their team who spends half the day copying information from one screen to another.

Work a custom agent can take off your plate

Good first candidates are repetitive, follow rules you can write down, and touch more than one system. A few we have shipped:

• Lead intake. Read the inquiry, look up the company, score it hot, warm or cold, write it into the CRM and book the call.
• Order and invoice checks. Match a purchase order to an invoice, flag the mismatches and draft the supplier email.
• Weekly reporting. Pull numbers from four tools every Monday at 07:00 and send leadership one page with the changes that matter.
• Research and monitoring. Watch competitor pricing or new listings and summarise only what changed.
• Content operations. Turn one long video into clips, captions and a posting schedule, then wait for approval before anything goes live.

Where agents go wrong, and the guardrails we build in

Agents fail in predictable ways. They act on bad input, they act too confidently, or nobody can see what they did. We design for all three before we write a single prompt.

• Approval steps. Anything that spends money, sends a contract or speaks for your brand in a new way waits for a human click.
• Narrow permissions. The agent gets the access this workflow needs and nothing more.
• A visible log. Every run is recorded in your client portal with its inputs, decisions and outputs, so you can audit any result in a minute.
• Fallbacks. If a tool times out or the data looks wrong, the agent stops, retries once and then hands the case to a person.

Start with one workflow, not a platform

The fastest route to value is one workflow that already costs you hours every week. Our route to a first agent has four steps.

1. Map it. Write down the trigger, every step, every tool and every decision a person makes today.
2. Measure it. Count how often it runs and how long it takes. That becomes your baseline.
3. Run it in shadow mode. For the first one to two weeks the agent drafts every action and a person approves it. You see the error rate before you trust it.
4. Hand over the routine. Once the drafts are reliably right, approvals shrink to the risky cases only.

If the workflow fails step one because nobody can describe the rules, an agent is the wrong tool. Fix the process first. We will tell you that on the call rather than sell you a build.

What it costs and how long it takes

A single-workflow agent on our Starter AI Agent package is a one-time $2,997 and goes live in 7 to 10 days. Multi-workflow agents with CRM integration, and agent swarms that coordinate across your systems, sit in the higher tiers. Model and hosting costs are passed through at cost.

See what each tier includes on the Custom AI Agent page, or compare the packages side by side in Custom AI Agent pricing.

Every missed call is a customer who tried to give you money and couldn't. Voicemail rarely saves it, because many callers hang up rather than leave a message and simply ring the next business on the list. An AI voice agent picks up on the first ring, at 2 pm or 2 am, and handles the call the way your best receptionist would.

What a voice agent actually does on a call

It listens, understands the request and acts on it inside your systems. In practice that covers most of what reaches a front desk:

• Answers questions about hours, prices, location, insurance or service areas from your approved answers.
• Books, moves and cancels appointments straight into your calendar.
• Qualifies new leads with your questions and scores them hot, warm or cold.
• Routes urgent or complex calls to the right person, with a summary so nobody asks the caller to start again.
• Takes a proper message: name, number, reason and urgency, written into your CRM.

After hours is where the money hides

For many service businesses the calls that matter most arrive when nobody is there: the burst pipe at night, the worried pet owner on a Sunday, the patient who can only call at lunch. A voice agent treats those calls exactly like daytime ones. It books the next free slot, dispatches the on-call technician for a real emergency, or sends a confirmation text so the customer stops shopping around.

Outbound matters too. Reminder calls the day before an appointment cut no-shows, and a follow-up call to a quote that went quiet often brings it back.

What makes a voice agent sound right

Callers forgive a lot, but not silence and not a robot that talks over them. These are the details we tune on every build:

• Response speed. Replies start in under a second, so the conversation keeps its rhythm.
• Interruptions. The agent stops talking when the caller does, and picks up the new question.
• A clear identity. It says it is an assistant for your business. It never pretends to be a person.
• A fast way out. Saying "a person, please" reaches a human or books a callback, every time.
• Your words. Persona, tone and vocabulary are agreed in discovery and written down, not improvised.

Measure it like a receptionist

After launch, judge the agent on the numbers you would use for a person: answer rate, bookings per hundred calls, calls escalated and complaints. Every call has a recording and a searchable transcript, so when something sounds off you can find the exact moment, and we fix it the same week.

Cost, timeline and where to start

A Starter Voice Agent begins at $1,497 setup and goes live in 7 to 10 days. It covers inbound calls on one line, up to 20 FAQs with call routing and simple booking. Premium adds outbound follow-up and reminders plus a lead qualification script. Telephony and model usage are billed at cost.

Read what each tier includes on the AI Voice Agent page, or go straight to AI Voice Agent pricing.

Most websites lose buyers quietly. Someone lands on your pricing page at 11 pm with one question, finds no answer and leaves. A contact form asks them to wait a day for a reply they will no longer need. An AI chatbot answers while they are still interested, then helps them take the next step.

Why speed decides who wins the lead

Interest fades fast. The business that replies first usually gets the conversation, and the one that replies tomorrow gets a polite no. A chatbot replies in seconds on your website, Messenger, WhatsApp or Instagram, so the first reply is always yours, including at weekends.

From question to booked call

A good chatbot is not an FAQ page with a text box. It moves the visitor forward in four steps:

1. Answer. It responds to the actual question using your knowledge base, prices and policies.
2. Qualify. It asks two or three of your questions, such as budget, timeline or location, and scores the lead hot, warm or cold.
3. Convert. Hot leads get a calendar slot inside the chat. Warm leads get the right resource and a follow-up.
4. Hand off. Anything sensitive or complex goes to a person with the full conversation attached.

Each step writes to your CRM, so your team starts every sales call knowing what the customer already asked. Nobody re-asks the basics, and the customer notices.

Ecommerce: order tracking and recommendations

For online stores the biggest win is often support, not sales. "Where is my order?" can be a large share of inbound messages, and a chatbot connected to Shopify or WooCommerce answers it with live tracking in seconds. The same connection lets it recommend products based on what the shopper is viewing, and bring back the visitor who was about to abandon a full cart.

It also catches the questions people would rather not ask a salesperson, such as whether the cheaper plan is enough or what happens if they cancel. Answered honestly and instantly, those questions keep buyers moving instead of bouncing.

What to prepare before you build one

• Your top questions. Collect the 30 to 75 questions your team answers most, with the answers you want given.
• Your rules. When should it book, when should it hand off, and what must it never promise?
• Your first channel. Start where most conversations already happen, then add the others.
• Your tone. Three adjectives and two replies you like are enough to start.

Cost and timeline

A Starter Chatbot begins at $997 setup on one channel with up to 30 FAQs, and goes live in 7 to 10 days. Premium adds lead qualification flows, appointment booking and CRM integration. Diamond adds WhatsApp and Instagram, ecommerce integration and custom training on your knowledge base. Track conversations started, questions resolved without a person, leads captured and bookings made from week one.

See the full feature list on the AI Chatbot page, or compare tiers in AI Chatbot pricing.