AI chatbot for your website: an off-the-shelf tool or a custom build?

For most websites, an off-the-shelf AI chatbot is the right start. Tools such as Intercom's Fin, Tidio's Lyro and Chatbase answer from your own help content, hand difficult questions to a person and can connect to common systems, for a monthly fee or a price per conversation. Build a custom chatbot when one of three things is true: a wrong word carries legal or safety risk, the bot has to work with a system no tool connects to, or answers depend on live data that must be shown with its source. We have built one of each, and they are described below.

Key takeaways

  • Off-the-shelf tools now do more than answer FAQs: as of November 2026 they answer from your content, hand off to people and, in some cases, look up orders or create tickets.
  • They price in three different ways: per resolved conversation, per message credit or per AI conversation. Compare them on your own monthly volume, not on the headline price.
  • Build custom when your wording is regulated, when the bot must work with a system no tool connects to, or when answers come from live data and must show where they came from.
  • A custom chatbot is a small product: content pipeline, retrieval, guards on what it may say, the integrations, an accessible chat widget, tests and someone to keep the content current.
  • Whichever route you take, decide first what the bot must never say and when it must hand over to a person.

What can an off-the-shelf website chatbot do?

The tools that rank for "AI chatbot for website" fall into two groups: chatbot builders you add to a site with a snippet of code, and AI agents built into a help-desk product. Three examples, each described from its own vendor page:

Tool How it answers Hand-off and actions Price model, as of November 2026
Intercom Fin Help-desk AI agent Can be used with Intercom or "tools like Salesforce and others" "$0.99 per outcome", charged once per conversation
Tidio Lyro "Lyro works only with the content and guidance you provide" Forwards questions it cannot answer to a person; flexible hand-off rules Lyro add-on "starts at $32.50 /mo." for 50 AI conversations (Tidio pricing)
Chatbase Answers grounded in data sources you connect "Look up orders, update records, create tickets, and trigger workflows" through integrations $120 or $400 a month billed yearly ($150 or $500 billed monthly), with 4,000 or 15,000 message credits; Enterprise priced on request (Chatbase pricing)

These are vendors' descriptions of their own products, not our test results. What they show is that the gap between "a simple FAQ bot" and "a custom build" has narrowed: answering from your content, handing off to people and calling common tools are now standard features.

When is an off-the-shelf tool enough?

An off-the-shelf tool is enough when:

  • The answers live in your public content. Help articles, policies, product pages and FAQs that you can keep up to date.
  • The actions are ones the tool already supports. Creating a ticket, looking up an order in a supported platform, booking through a supported calendar.
  • A wrong answer is an inconvenience, not a liability. A customer gets a slightly off delivery estimate and asks a person.
  • You are happy for the vendor to host the conversations under its terms and in its regions.

Start there, measure what the bot cannot answer, and revisit the decision after a few months of real conversations. Our guide to building a customer-support chatbot covers the setup steps, the hand-off and the disclosure rules that apply either way.

When does a website chatbot need a custom build?

In our projects, three situations have pushed clients to a custom build. Each of the following is a chatbot we built.

1. A wrong word carries legal or safety risk

A supplement retailer, Saba, wanted visitors to get quick answers about products, orders, the company and policies. Supplement wording carries regulatory risk, so a bot that writes freely from page text was not acceptable. We built:

  • a curated set of approved questions and answers, each tagged with how it may be served: as written, with a disclaimer, as a link only, held for review or hedged;
  • a stored answer returned word for word only when it is clearly the right one, and otherwise a reply drafted by Azure OpenAI only from retrieved passages, or a plain "I cannot answer that";
  • an output guard that removes disease claims, quoted prices, foreign links and emails before anything reaches the visitor;
  • safety-topic tracking across a conversation, with a caring reply and a crisis line when someone mentions self-harm.

Tools offer instructions and guardrail settings. What this client needed went further: its own serving rules for every approved answer, a tested output guard written for its products, and the whole system deployed in Azure from its own templates.

2. The bot has to work with a system no tool connects to

Mammoth Self Storage, an Australian self-storage operator, wanted customers to work out how much space they need and reserve it in one conversation. The chatbot we built:

  • estimates item volumes from a photo or video the customer uploads, using the Gemini API, and lists the items with a total for the customer to confirm;
  • finds the nearest site and reads unit availability, discounts and quotes from the operator's storage management system;
  • offers alternative unit sizes when the first choice is not free, and sends the customer to a pre-filled booking page;
  • gives staff an admin screen for the chatbot's prompts and questions, with SMS and email notifications.

Estimating volume from a video and quoting live units from the operator's own system were the point of the bot, and both had to be built. When the value of the bot is a task in your own systems, the integration is the project.

3. Answers depend on live data and must show their source

A US city wanted residents, contractors and businesses to get answers about their property or project without reading ordinances. Our city services assistant answers from the city's own codes and documents with citations, looks up an address across 17 live map layers (zoning, flood zones, council districts and more), shows the result on a map in the chat, and routes people to the right department or form. A resident can also upload a photo of a property issue and get the matching service request form.

Here a fluent answer is worth nothing without the source and the current map data behind it, and both live in the city's systems.

What does a custom build involve?

Expect these parts, whatever the model:

  1. Content and knowledge: which pages, documents and approved answers the bot may use, how they are cleaned and indexed, and how often. A bad crawl must not wipe the index.
  2. Retrieval and the model: search over that content, and a model that answers only from what was retrieved. Our RAG buyer's guide explains how this works and what decides quality.
  3. Guards: fixed replies for topics the bot must not improvise on (prices, orders, abuse, self-harm), and an output check before anything is shown.
  4. Integrations: the systems the bot reads from or acts in, with the narrowest access that works.
  5. The chat widget: fast to load, usable by keyboard and screen reader, and safe with the links it shows.
  6. Tests and monitoring: a set of real and adversarial questions run on every change, and a log of what the bot could not answer.
  7. Ownership: someone who updates the content and reviews the unanswered questions every week.

What drives the cost?

We do not publish prices, because they depend on the parts above. The drivers are the same on every project: how many integrations the bot needs, how strict the wording rules are, how much content there is and how often it changes, model usage per conversation, and who maintains it afterwards. For off-the-shelf tools, the price model matters more than the headline figure: per resolved conversation, per message credit and per AI conversation give very different monthly bills at the same traffic, so price them on last quarter's chat volume.

Privacy rules apply to both routes. If customers in Canada or India will use the bot, our guide to privacy for AI features in Canada and India covers what to check. This is general information, not legal advice.

Questions to answer before you choose

  1. What must the bot never say, and how will you stop it?
  2. Which questions should go straight to a person?
  3. Which systems does it need to read from or act in, and does any tool connect to them?
  4. Where must conversations be stored, and for how long?
  5. Who keeps its content current, and who reviews what it could not answer?

If the answers point to a tool, start with one. If they point to a custom build, our AI assistants service covers chatbots like the three above.

FAQ

Can I add an AI chatbot to my website without a developer?

Usually, yes. Most tools install with a snippet of code and learn from content you give them. You still need someone to choose the content, set the hand-off rules, and review what the bot gets wrong.

Is a custom AI chatbot more accurate than ChatGPT-style tools?

Not automatically. Accuracy comes from answering only from your own approved content, testing with real questions and guarding what the bot may say. A custom build lets you control each of those; a well-set-up tool can do the first two.

Which AI model should a website chatbot use?

The one that answers your test questions best within your data and cost limits. We have built chatbots on Azure OpenAI and on the Gemini API; the content, guards and integrations around the model decide more of the result than the model itself.

Can a chatbot take bookings or orders?

Yes, if it can reach the system that holds them. Some tools connect to common platforms; for anything else, the integration has to be built, as it was for the storage chatbot above.

How do we know the chatbot is working?

Track the share of conversations it resolves without a person, the questions it could not answer, and a sample of answers reviewed every week against your content. Run your test questions again after every content or model change.

Sources (5 links)
Share this article