AI chatbot for business: build, buy or hire, with real cost drivers

An AI chatbot for business is a chat interface, on your website, in an app or inside a messaging channel, that uses a language model to answer questions and sometimes take actions. Decide first what it must do: answer from your content, look up a customer's data, or change something in a system. That choice, more than the vendor, sets the cost and the risk.

#Start with what the bot must do

Most chatbot projects go wrong because the team picked a tool before agreeing on the job. There are three levels, and each one asks for more work and more safeguards.

Most small businesses need only the first level, and many gain a lot from it. Look-up often removes more repetitive tickets, because "where is my order" is one of the most common questions, but it adds a security question: how does the bot know who it's talking to? Action is the most useful and the most dangerous. We rarely turn it on for money, contracts or account closures without a human approval step.

#Widget, platform or custom build

A common path is to start with a widget or platform for answers, then build custom when the bot needs to read or change data in your own systems. If you're also comparing the agent approach more broadly, our AI agent development cost guide covers that side.

#What drives the cost

We don't quote prices here, because they vary too much to be useful. These are the drivers that move a quote, in the order they usually matter.

• Scope: answering from a help center is a smaller project than reading orders, and far smaller than changing them.
• Number of systems: each integration (CRM, help desk, order system, identity) adds build and testing work.
• Content readiness: scattered, outdated or contradictory documents need cleanup before a bot can use them.
• Channels: a web widget, WhatsApp and voice each have their own setup and rules.
• Languages: every extra language needs testing with native speakers.
• Security and data rules: where data is processed, what is stored and for how long, and who can read transcripts.
• Handoff design: routing to a human inbox with context takes real design time.
• Evaluation: building a test set of real questions and running it after every change.
• Running costs: model usage, hosting, platform fees and someone to review failures weekly.

Ask every vendor for the one-time and the ongoing costs separately. The ongoing part is the one buyers forget.

#How a good chatbot answers

A bot that answers from your content works in four steps. It takes the question, retrieves the relevant passages from your approved documents, writes an answer using only those passages and shows its sources. If nothing relevant comes back, it says so and offers a person. Our guide to RAG knowledge assistants explains the retrieval step.

Say a customer asks whether a product can be returned after 45 days. If your policy page says 30 days, the bot should quote the policy and offer a handoff for exceptions. It shouldn't guess, and it shouldn't invent an exception to be helpful.

Key point: A chatbot speaks for your company. Limit it to approved sources, show where answers come from and give people an easy route to a human.

#Accuracy, liability and disclosure

The business owns what its chatbot says. In the 2024 British Columbia case Moffatt v. Air Canada, a tribunal held the airline responsible for wrong refund information its website chatbot gave a customer, and rejected the airline's argument that it wasn't liable for what the chatbot said. The decision is cited as 2024 BCCRT 149, and law firm summaries such as McCarthy Tétrault's (opens in a new tab) cover it. It's a single decision in one jurisdiction, but the lesson is practical: assume you're accountable for answers.

Disclosure is becoming a rule in some places. Article 50 of the EU AI Act says providers must design AI systems that interact directly with people so those people are informed they're dealing with an AI, unless that's obvious to a reasonably well-informed person. See the Article 50 text (opens in a new tab). Check the application dates and how they apply to you, and take legal advice for the countries you serve. Telling users they're talking to an AI is good practice anywhere.

Privacy needs its own thought. Decide whether chat transcripts are stored, where they're processed, how long you keep them and who can read them. Tell users not to paste sensitive information, and mask card numbers and ID numbers before anything is logged.

#Where a human steps in

Design the handoff before the bot. Typical triggers:

1. The user asks for a person.
2. The bot has no relevant source, or its confidence is low.
3. The topic is on a restricted list: complaints, legal threats, refunds above a limit, medical or safety issues.
4. The user is upset, or repeats the same question twice.
5. The action requested needs authorization.

When it hands off, it passes the transcript, the customer's identity and a one-line summary, so the person doesn't start from scratch. Our AI customer service guide goes deeper on routing, and the customer support page shows how we build it into a help desk.

#Evaluate before launch

Don't judge a chatbot on the demo. Collect 100 to 200 real questions from your inbox and help desk, including the awkward ones. Write the right answer for each, or mark "should hand off". Run the bot against the set, have someone read the results and track three numbers: correct answers, wrong answers stated confidently and correct handoffs. Rerun the set every time you change the content, the prompt or the model.

After launch, read a sample of conversations weekly, and review every conversation that ended in a complaint.

#A pre-launch checklist

• The bot's job is written down as answer, look up or act
• Answers come only from sources you've approved
• Every answer can show its source
• A person is one click away, with context passed along
• A restricted topic list exists and is tested
• Users are told they're talking to an AI
• Transcript storage, retention and access are decided
• A test set of real questions is passing
• A named person reviews conversations weekly
• You can switch the bot off in minutes

#When a chatbot isn't the answer

If you get twenty inquiries a week, a good FAQ page and a clear contact form may do more. If your content is outdated or contradicts itself, a chatbot will repeat the mess faster. And if the real problem is slow human follow-up, fix that first. A chatbot that promises a callback nobody makes damages trust. For a first conversation, see our use cases for examples of what teams automate and where the human stays.

If you're deciding between a widget, a platform and a custom build, book a strategy call. We'll look at your content, systems and support volume and tell you which option fits.

#Frequently asked questions

#What is an AI chatbot for business?

It is a chat interface powered by a language model that answers customer or staff questions, and in more advanced setups looks up data or takes actions in your systems. Good ones answer from approved content, cite it and hand off to a person when unsure.

#How much does an AI chatbot cost?

It depends on scope, the number of systems connected, content readiness, channels, languages and security needs, plus running costs for model usage and review. A widget answering FAQs costs far less than a bot that reads orders or changes records. Ask for one-time and ongoing costs separately.

#Should I build a custom chatbot or buy one?

Buy a widget or platform if the job is answering common questions from your own content. Build custom when the bot must read or change data in your systems, or when data rules limit where information can go.

#Can a chatbot give wrong answers?

Yes. Models can state incorrect things confidently. Limit the bot to approved sources, show citations, test it on real questions and keep a handoff to a person. Businesses can be held responsible for what their chatbots say.

#Do I have to tell customers they're talking to an AI?

In some places, yes. The EU AI Act includes a disclosure rule for AI systems that interact directly with people, with an exception for cases where it's obvious. It's good practice everywhere, so state it plainly in the chat window.