A customer lands on your website at 9pm and asks whether you do same-day returns. Nobody is online. A chatbot could answer in seconds. It could also answer wrongly, and then the wrong answer is yours.
That is the real decision with an AI support chatbot. Not whether it can chat, because they all can. It is whether you can stand behind what it says, hand a customer to a person when it should, and predict the bill.
We list three tools built for this: Fin by Intercom, Decagon and Sierra. All three are aimed at larger teams. This post is honest about that, and about what a small business should check either way.
Quick answer: If you already run a helpdesk and get steady support volume, start with Fin, because it works on top of the helpdesks businesses already use. Decagon and Sierra are built for large companies and sell through demos and quotes, so a small business will usually find them a poor fit. Whichever you try, test it on your own hardest questions, confirm how it hands off to a person, and work out the bill at your real volume before you switch it on.
Why this is different from a website FAQ
An FAQ page says the same thing to everyone and you wrote every word. An AI agent composes a fresh answer each time. That is the benefit and the risk.
In 2024 a tribunal in British Columbia looked at exactly this. A customer asked an airline's chatbot about bereavement fares and got an answer that did not match the airline's actual policy. The airline argued it was not responsible for what the chatbot said. The tribunal rejected that, treating the chatbot as part of the airline's website, and found the airline had not taken reasonable care to make sure it was accurate. It also said the correct policy sitting elsewhere on the site did not excuse the wrong answer.
The decision is from Canada and is not binding in Australia. We are not lawyers and this is not legal advice. The point for a business owner is practical: customers will treat the chatbot as you. If you want the Australian position, ask your own adviser and read the guidance from the ACCC and the OAIC.
Fin by Intercom
Fin is the one a small business is most likely to be able to use. The vendor says it handles conversations across chat, email, voice, Slack and social, can read and write to other systems (for example updating an account or processing a refund), and passes the conversation to your human team with the full context so a person does not start again.
It is built on Intercom's helpdesk but, the vendor says, also works with others including Zendesk, HubSpot, Freshdesk and Salesforce. That matters if you already have a helpdesk and do not want to move.
In our review, Fin is good for businesses that want an agent resolving customer questions, especially on Intercom. The catch is billing. It is charged per outcome, a conversation can count as resolved when the customer simply leaves, and seats, minimums and add-ons come on top. If you cannot predict your support volume, you cannot predict the bill.
Decagon
Decagon builds customer-facing agents for voice, chat and email, and lets teams define how an agent should behave in plain language. Its customers are large companies in sectors such as finance, travel and retail.
The vendor's page does not publish pricing and does not explain how handoff to a person works. In our review the catch is that pricing is quoted per deal, and per-conversation billing can charge for conversations that still end with a person. For a small business that wants a self-serve tool with a public price, skip it.
Sierra
Sierra builds and runs agents for large enterprises across chat, SMS, WhatsApp, email and voice. The vendor says it is priced on outcomes rather than a flat fee. It reports why conversations are transferred to human agents, but its site does not describe the handoff itself.
In our review, the catch is that rates are not published and "resolved" is the term to pin down in the contract. Like Decagon, it is not a quick, low-cost start.
If none of those fit
Most small businesses do not need an enterprise agent. They need something that answers the same ten questions at night and passes everything else to a person.
There are two other routes on the site:
- Build it yourself on an agent platform. Relevance AI builds agents for support, sales and operations. The catch from our review is that usage is metered, so the bill grows with how much your agents do.
- Use a voice agent for calls. If the pain is the phone, not the website, Retell AI builds voice and chat agents billed per minute, with several charges stacked in each minute. Our post on AI voice agents for after-hours calls covers what to ask.
You can also browse everything we list under customer support chatbots.
The checks that matter before launch
1. Test your hardest questions, not the easy ones. Write down twenty real questions your customers ask, including the awkward ones: refunds outside the policy, warranty edge cases, anything involving a deadline. Run them all and read every answer. Any answer you would not sign off should be fixed in the source material or taken off the agent's list of topics.
2. Check what it learns from. The vendor pages say these agents are trained on your policies and help content. That makes your own documents the weak point. If your returns page is out of date, the agent will be confidently out of date. Fix the source first.
3. Check the handoff. Ask three things. When does it pass to a person? Does the person see the whole conversation? What does the customer see if nobody is online? A chatbot that traps a customer with no way out is worse than no chatbot.
4. Make it obvious it is AI. Tell customers they are talking to an AI and how to reach a person. This is good practice and follows the tribunal's lesson that businesses should warn customers about the limits of their AI tools.
5. Price it at your real volume. For the outcome-billed tools, ask in writing what counts as resolved, what happens when a customer leaves without confirming, and whether a conversation that ends with a person is still charged. Then multiply by your last three months of conversations. Our reviews flag this for Fin, Decagon and Sierra for a reason.
6. Decide who owns it. Someone on your team should read a sample of conversations every week for the first month. Agents drift as your products and policies change. If nobody owns it, it will get stale.
7. Keep a switch. You should be able to turn it off in one click and fall back to a contact form. Ask the vendor how.
What we would do
If you have a helpdesk, a steady flow of repeat questions and someone to review conversations, trial Fin on a narrow set of topics for a month and measure how many it settles without a person. If you have a handful of enquiries a week, a good FAQ page, a clear contact form and a fast reply will do more than any chatbot. If you are not sure which side you are on, that is a normal place to be.
An AI policy helps too. Our guide to writing an AI use policy for a small team covers who is allowed to put what in front of customers.
Where to start