AI Chatbots for Business: When Are They Worth It?

The short answer

A chatbot is worth building when it answers questions people genuinely have, faster and more accurately than they could find the answer themselves, and gets out of the way when they need a person. That sounds obvious. Most disappointing chatbots fail one of those three tests: they answer questions nobody asked, they answer wrongly, or they trap people.

This article is about the decision: whether a customer-facing assistant makes sense for your organization. If you have already decided yes, the work that determines quality is the knowledge base behind it, which is covered in building an AI knowledge base for customer-facing teams.

Start with your question volume

Before talking to any vendor, spend two or three weeks counting. Pull the questions that reach your team by email, phone, social messages, contact forms and the front desk. Categorize each one:

• Answerable from published information. Hours, locations, eligibility, how to book, what is included, directions, deadlines.
• Answerable from internal information. Process questions your staff can answer from procedures that are not public.
• Needs a person. Account-specific, complaints, quotes, anything that depends on someone's circumstances.

Then look at the shape. If a large share of the volume sits in the first category and repeats endlessly, a chatbot has real work to do. If most questions need a person, a chatbot will mostly be a step people have to get past.

An illustrative example. Say a ski hill in New Brunswick finds that in peak season its small team spends much of each day answering the same handful of questions about conditions, lift hours, rentals, lessons and parking, often after hours. The answers exist and change on a predictable schedule. That is a strong case. Now say a boutique law firm gets a few enquiries a week, nearly all about someone's specific situation. That is a weak case: a person replying promptly is better service and costs less.

Seasonality matters here in Atlantic Canada. Tourism operators, recreation facilities and event organizers often see question volume spike for a few months a year, exactly when staff are stretched. An assistant that absorbs the repetitive peak can be worth more than its annual volume suggests, provided someone keeps the seasonal information current.

Use cases that work, and ones that don't

Usually worth it

• Information lookup. "What's on this weekend?", "Do you take walk-ins?", "Which form do I need?" Questions with a factual answer in your content.
• Guided navigation. Helping someone find the right page, program or service on a large site, such as a municipality or institution.
• Triage. Asking a few questions, then routing the person to the right team with a summary attached.
• Staff-facing assistants. Helping front-line staff find answers in internal procedures. Often the best first step, because a staff member checks every answer before passing it on.

Usually not worth it

• Lead capture disguised as help. A bot that exists mainly to collect an email address before answering frustrates people and rarely improves lead quality.
• Personalized advice. Legal, financial, clinical or employment questions where the right answer depends on facts the bot does not have and should not collect.
• Thin content. If your website does not contain the answers, the chatbot will not either. Fix the content first.
• Low volume. When a person can reply within the hour and questions are few, a chatbot adds a layer without saving much.

Customer experience: better than search, faster than waiting

The bar a chatbot has to clear is simple: it must be better than searching your site and faster than waiting for a reply. A few principles keep it on the right side of that bar.

• Say it is an AI. People are generally fine talking to an assistant. They are not fine discovering later that it was not a person.
• Show sources. A link to the page the answer came from lets people check, and builds trust when the answer is right.
• Answer in the customer's language. In New Brunswick, an assistant that handles French and English equally well is a basic expectation, not a feature. Test both.
• Make the exit obvious. A visible way to reach a person at every point, not after three failed attempts.
• Keep it accessible. The chat interface must work with keyboards and screen readers, like the rest of the site.

Consistency is the other half. On Hey Freddy, the studio's own bilingual city guide, Freddy has one personality and one set of guardrails and answers in whichever official language someone uses, whether they type into Ask Freddy on the site or call Freddy the Radio Wonder Phone. Whatever channels you offer, customers should meet the same organization, with the same facts, at every door.

Data privacy: what people will type

Whatever you intend the chatbot for, people will type things into it: account numbers, health details, complaints about named staff, their home address. Plan for it.

• Do not ask for personal information the bot does not need. Collect the minimum to route the conversation.
• Know where transcripts go. Which vendor stores them, for how long, in which country, and whether they are used to train models. Read the business terms, not the marketing page.
• Set retention. Keep transcripts as long as you need them for quality review and no longer.
• Tell people. Your privacy notice should explain that the assistant exists, what happens to conversations and how to reach a person instead.

Canadian privacy law is built on principles such as meaningful consent, limiting collection to what is needed, safeguarding what you hold and being accountable for what your vendors do with it. Public bodies and health organizations have their own legislation on top. None of this is legal advice; if your chatbot will handle personal information, involve your privacy officer or counsel early. The AI governance checklist for Canadian marketing teams lists the vendor questions to ask.

Handoff to people

The handoff is where most chatbots are judged, because it is where the frustrated customers end up. Design it as carefully as the answers:

• Clear triggers. A request for a person, low confidence, a complaint, an account-specific question or repeated confusion.
• Context carried over. The person who picks up sees the conversation. Nobody should have to repeat themselves.
• Honest timing. Outside staffed hours, say when a person will reply, and make sure someone does.
• Feedback loop. Track what gets escalated and why. Repeated escalations on one topic usually mean a missing or unclear document.

What drives the cost, and how to decide

The chat window is the cheap part. What drives the cost of a credible customer-facing assistant is:

• Preparing and cleaning the knowledge base, usually the largest piece of work.
• Integration with your website, booking system, CRM or ticketing tool.
• Building and running a test set before launch and after changes.
• Usage-based model costs, which grow with conversation volume and length.
• Staff time to review conversations, update content and handle escalations, permanently.

Any credible estimate needs scoping against your actual question volume and content. A quote that does not ask about either is pricing a widget, not a service.

Red flags in a chatbot proposal: no question about where your answers will come from, a promise that the bot will "learn your business" on its own, no plan for testing before launch, no mention of how conversations reach a person, and vague answers about where transcripts are stored.

Before you commit, check:

1. Do we have enough repetitive, answerable questions to justify it?
2. Do accurate written answers already exist, in both official languages if we serve both?
3. Who owns the content and keeps it current?
4. How will people reach a person, and who answers?
5. Do we know what happens to the data people type in?
6. How will we measure whether it worked? Measuring ROI from AI marketing workflows covers the baseline you need.

If several answers are "not yet", start with a staff-facing assistant or with fixing the content. Both make a future chatbot better, and both pay off on their own. Our AI and automation work often starts there for exactly that reason.