# How to build AI agents for a business: a practical guide

> How to build an AI agent for a business: pick one job, give it knowledge and tools, add limits and handover, test in draft mode, then grow its autonomy.

Leonard Loo · CEO & Co-Founder, Volty

Updated Nov 5, 2026

6 min read

To build an AI agent for a business, choose one job with a clear source of truth, give the agent your information and the tools it needs, write its rules and a path to a person, test it while someone approves every output, and widen its autonomy only as it proves reliable. You can assemble the parts yourself from a language model and code, or configure an agent on a platform that provides them. This guide explains the parts, compares the two routes and gives a launch plan you can follow.

## Start with the job

A project that begins with "let's add AI" and no job has nothing to measure. A usable job has four properties.

- It repeats often, so there is volume to learn from.
- The right answer lives somewhere you can point to: a price list, a calendar, a policy page, a record.
- A mistake is recoverable, or a person can check before it matters.
- You can measure it: reply time, bookings, resolved chats, hours saved.

Good first jobs are answering enquiries after hours, qualifying leads, booking appointments and sorting a shared mailbox. Poor first jobs are anything that spends money or gives advice that a regulator expects a qualified person to give. The AI agent examples article lists real jobs and what they needed.

## The parts of an agent

| Part | What it does | What goes wrong without it |
| --- | --- | --- |
| Language model | Reads the message and writes the reply | Slow, costly or weak replies if the model is the wrong size for the job |
| Knowledge | Your documents, website and records the agent can look up | Invented answers |
| Tools | Actions in other systems: book, update, create, send | An agent that can talk but not do |
| Memory | What the agent knows about this customer and what your team has taught it | Customers repeat themselves, mistakes repeat |
| Rules and limits | What it must, may and must never do | Promises you cannot keep |
| Handover | Passing the chat to a person with context | Stuck customers |
| Logs and review | A record of what it saw and did | No way to fix it |

Knowledge usually matters most. Retrieval, meaning the agent searches your documents and answers from the passages it finds, is the standard way to keep answers accurate. How retrieval makes WhatsApp chatbots accurate explains the mechanism.

## Route 1: configure an agent on a platform

A platform supplies the model, the channels, the inbox, the memory and the guardrails, and you supply the knowledge, instructions and connections. This route fits teams whose goal is the outcome, and it is how most small and mid-sized businesses start.

- Connect a channel, such as your WhatsApp number, web chat or email.
- Give it knowledge: your website, files and documents you keep current.
- Write instructions in plain language: tone, questions to ask first, limits, when to hand over.
- Connect the systems it needs to act in, such as a calendar, customer records or an order system.
- Set the reply mode to draft and have your team review.
- Automate topic by topic.

## Route 2: build from a model API

If you have developers and need behaviour a platform does not offer, you can build the agent loop yourself. The outline is the same in any framework.

- Take the incoming message and the conversation so far.
- Retrieve relevant passages from your knowledge and attach them.
- Send the prompt, with your rules and a list of tools, to the model.
- If the model asks to call a tool, run it, return the result and ask the model again. Repeat until it produces a reply.
- Send the reply, or queue it for approval.
- Log the whole exchange, and store anything worth remembering.

That loop is a few hundred lines. The harder work is everything around it: authentication per customer, approvals, retries when a tool fails, rate limits, evaluation, an inbox where staff see and take over, and the connections to messaging channels with their own rules. Some teams prototype the loop themselves and then run it on a platform, a pattern covered in why Volty works with Claude Code.

## Write the rules before you test

A useful rule set is short and concrete.

- Who the agent is and whom it serves, in a sentence.
- Questions it asks before it quotes or books.
- Facts it may state, and the source it must take them from.
- Things it never does alone: refunds, discounts, medical or legal advice.
- When it hands over: upset customers, a request for a person, two failed attempts to answer.
- How it signs off after a handover.

If the agent talks to customers on WhatsApp, add Meta's requirements: free-form automated replies only inside the 24-hour window, templates only to opted-in people, and a prompt, clear path to a person. See WhatsApp AI agents for the rules in full.

## Grade the actions by impact

An agent that can act needs a ladder of trust. A practical grouping:

| Impact | Examples | Setting |
| --- | --- | --- |
| Low | Fill a contact field, add a note, save a lesson | Applies at once, with an undo |
| Medium | Book or reschedule an appointment | Automatic once drafts have been clean |
| High | Refunds, spending, mass sends, deleting records | Waits for a person's approval |

Volty works this way: low-impact changes apply at once with a Revert, while high-impact ones such as spend, sends at scale and anything irreversible wait for a person to approve or reject. The framework in hire your AI agents, don't install them argues for earning autonomy in the same order.

## Test like an employee on probation

- Gather 50 real past messages, with the answers your team gave. Replay them through the agent and compare.
- Add edge cases: an angry customer, a question outside your knowledge, a language you do not speak, a photo, a request for a person.
- Run it live in draft mode. A person approves, edits or rejects each reply, and each edit is a lesson.
- Count how many drafts went out unchanged. When a topic passes 90% for a week, automate that topic.
- Keep sampling. Read 20 conversations a week even after launch.

## Measure what matters

- First-reply time, and the share of messages answered outside office hours.
- Share of conversations resolved without a person, and the share handed over.
- Corrections per week, which should fall as the agent learns.
- The business outcome: bookings, qualified leads, quotes sent.
- Cost per conversation, including model usage and message fees.

Cake Inspiration measured its agent this way: 91% of conversations resolved by Volty without a human across 800+ messages in one week, and a median reply of 5 seconds. Kai Yang, Founder, Cake Inspiration, said: "I can fix the agent myself now, without waiting on an engineer. I just tell it what I didn't like about a reply, and it learns. Sometimes it will automatically suggest improvements and it will ask me to approve it. I think that's pretty cool."

Cake Inspiration. Read the case study

## Common mistakes

- Starting with five jobs. One job reaches reliable quality sooner.
- Feeding the agent stale documents. When prices change, update the source the same day.
- Skipping draft mode because the demo looked good.
- Giving broad permissions early. Grant tools one at a time.
- No owner. Someone on your team must read conversations and own corrections.

## Building an agent in Volty

Volty is an AI helpdesk for builders, so the channels, shared inbox, knowledge, memory, approvals and logs are already there. You connect a channel, point Volty at your website and documents, and tell it what to do by chatting with it. Each inbox chooses between auto-reply, drafts and off, and staff teach Volty by correcting replies, with every lesson visible and reversible. Volty connects to HubSpot, Google Calendar, Notion, Shopify, Slack and any MCP server, and acts in them through those connections. Start with the WhatsApp AI page, and compare plans on pricing.

## Frequently asked questions

How do I build an AI agent?

Pick one job with a clear source of truth, give the agent your information and the tools it needs, write its rules and a handover path, test it where a person approves every output, then widen its autonomy as it earns trust. You can assemble it from a language model and code, or configure one on a platform that supplies the model, memory and channels.

Do I need to code to build an AI agent?

Not for a customer-facing agent on a platform that provides the channels, knowledge and tools as settings. You need code if you build from a language model API: you write the loop that calls the model, runs tools, stores memory and logs what happened.

What are the parts of an AI agent?

A language model that reads and writes, knowledge it can look up, tools it can call to take actions, memory of the customer and the conversation, rules and limits, a way to hand over to a person, and logs so you can see what it did. Weak parts show up as wrong answers or unsafe actions.

How long does it take to build a business AI agent?

A first version on a platform can run in an afternoon. A trustworthy one takes weeks: a week of review in draft mode, then gradual automation of topics that came out clean. A custom build adds engineering time for tools, logging, evaluation and staff screens.

How do I keep an AI agent from making costly mistakes?

Limit what it can do alone. Let it apply low-impact changes and require a person's approval for refunds, spending and anything hard to undo. Keep its answers tied to your own sources, log every action, and review conversations it could not handle.

## Recommended for you

How to create a WhatsApp chatbot: two routes, step by step · How to train an AI agent on your business, step by step · HubSpot WhatsApp Integration: Options, Steps and Limits · Instagram CRM: Turn DMs Into Customer Records · Omnichannel CRM: How It Works and How to Choose One · Salesforce WhatsApp Integration: Routes, Setup and Limits · Customer service automation: what to automate first and how · How to Change WhatsApp to a Business Account (Step by Step) · Instagram Customer Service: How to Set It Up and Run It

## Start your Sandbox Trial

Email us questions · Start your Sandbox Trial

Sign up for our 7-day trial. No credit card needed.

Future-proof your business with AI. Hear from our CEO

The Leonard and Volty Show · October 8, 2026

Meta now charges for WhatsApp replies

The Leonard and Volty Show · October 1, 2026

Too busy to hire?

The Leonard and Volty Show · September 21, 2026

From $500 chatbot jobs to agentic AI

The Leonard and Volty Show · September 10, 2026

Your AI is a new employee

The Leonard and Volty Show · September 4, 2026

Trust in the AI world

The Leonard and Volty Show · August 29, 2026

Does the AI model matter?

The Leonard and Volty Show · August 19, 2026

What are AI agents?

Meta now charges for WhatsApp replies. Here's what it means for your business, October 8, 2026

## Links

- [AI agent examples](https://voltade.com/blog/ai-agent-examples)
- [How retrieval makes WhatsApp chatbots accurate](https://voltade.com/blog/improving-whatsapp-chatbot-understanding-with-rag)
- [why Volty works with Claude Code](https://voltade.com/blog/why-volty-works-with-claude-code)
- [WhatsApp AI agents](https://voltade.com/blog/whatsapp-ai-agent)
- [hire your AI agents, don't install them](https://voltade.com/blog/hire-your-ai-agents)
- [Cake Inspiration](https://voltade.com/case-studies/cake-inspiration)
- [WhatsApp AI page](https://voltade.com/product/whatsapp-agent)
- [pricing](https://voltade.com/pricing)
- [How to create a WhatsApp chatbot: two routes, step by step](https://voltade.com/blog/how-to-create-a-whatsapp-chatbot)
- [How to train an AI agent on your business, step by step](https://voltade.com/blog/how-to-train-ai-agent)
- [HubSpot WhatsApp Integration: Options, Steps and Limits](https://voltade.com/blog/hubspot-whatsapp-integration)
- [Instagram CRM: Turn DMs Into Customer Records](https://voltade.com/blog/instagram-crm)
- [Omnichannel CRM: How It Works and How to Choose One](https://voltade.com/blog/omnichannel-crm)
- [Salesforce WhatsApp Integration: Routes, Setup and Limits](https://voltade.com/blog/salesforce-whatsapp-integration)
- [Customer service automation: what to automate first and how](https://voltade.com/blog/customer-service-automation)
- [How to Change WhatsApp to a Business Account (Step by Step)](https://voltade.com/blog/how-to-change-whatsapp-to-business-account)
- [Instagram Customer Service: How to Set It Up and Run It](https://voltade.com/blog/instagram-customer-service)
