Adding an AI chat assistant to your website: options and risks

First, decide what it's for

Write down the job in one sentence before you look at products. Good examples:

• "Answer questions about opening hours, services, and booking, and pass everything else to our team."
• "Help shoppers find products and answer shipping and returns questions from our policy pages."
• "Collect project enquiries outside business hours and book a discovery call."

A narrow job is easier to set up, test, and trust. "Answer anything about our business" is where problems start.

The three main options

1. AI features in a help desk or chat tool

Many customer support and live chat platforms now include an AI assistant that answers from your help articles and hands off to your team inside the same tool. If you already use one for email support or live chat, this is often the quickest route.

Good for: businesses with existing support content and a team that answers chats. Watch for: pricing models that charge per AI answer or resolution, which can be hard to predict. Check the current plans.

2. Website platform add-ons

Website builders and e-commerce platforms such as Shopify, Wix, Squarespace, and WordPress have AI chat apps and plugins, some from the platform itself and many from third parties. They can often read your existing pages and product catalog.

Good for: small sites and stores that want something running quickly. Watch for: the quality and privacy terms of third-party apps vary widely. Apply the questions in how to evaluate an AI tool. For WordPress, treat a chat plugin like any other plugin that touches visitor data: keep it updated and remove it if it's abandoned.

3. A custom assistant

A developer builds an assistant using an AI provider's API, connected to your content and, sometimes, your systems (booking, order lookup, CRM). This gives you the most control over what it knows, how it answers, and where data goes.

Good for: businesses with specific workflows, higher volumes, or stricter data requirements. Watch for: ongoing usage costs, maintenance, and the extra security work involved in connecting it to live systems. This is the kind of build Chykalophia does, and it's worth getting right rather than fast.

The options side by side

The risks to plan for

Wrong answers

AI assistants can invent answers, including prices, policies, and promises you don't offer. Courts and tribunals may hold a business responsible for what its chatbot says. In a widely reported 2024 case, a Canadian tribunal held Air Canada responsible for inaccurate refund information its website chatbot gave a customer.

How to reduce it:

• Restrict the assistant to answering from your own approved content
• Tell it to say "I don't know" and offer a person when the answer isn't in that content
• Keep policy pages current, because the assistant will repeat whatever they say
• Exclude topics where a wrong answer is costly (pricing exceptions, legal, medical, financial advice)

Manipulation (prompt injection)

Visitors can try to trick an assistant into ignoring its instructions, revealing its setup, or saying something embarrassing. If the assistant can take actions (look up orders, issue discounts, send emails), the stakes rise. The OWASP Top 10 for large language model applications lists prompt injection as a leading risk.

How to reduce it: give the assistant the fewest permissions it needs, require confirmation for any action, never put secrets in its instructions, and test it with deliberately awkward questions.

Privacy

Visitors type all sorts of things into chat boxes, including personal and sensitive details. You need to know where those conversations go, how long they're kept, and whether the provider uses them to train AI models.

• Update your privacy notice to describe the chat and the provider
• Check the provider's data terms and sign a data processing agreement if you're subject to GDPR or similar laws
• Discourage sensitive information in the chat, and don't use a general chat assistant for health or payment details. See HIPAA basics for websites handling health info if you're in healthcare.

AI and your customer data covers this in more depth.

Being clear it's an AI

Visitors should know they're talking to an AI. It's honest, it sets expectations, and some laws require it in certain situations. For example, the EU AI Act includes transparency rules requiring people to be told when they're interacting with an AI system. Label it clearly ("AI assistant") and avoid giving it a human name and photo that suggest otherwise.

Accessibility

A chat widget must work with a keyboard and screen readers, have enough color contrast, and not cover important content on small screens. Test it with the same care as the rest of your site. See ADA & WCAG: web accessibility compliance.

Cost and abuse

Usage-based AI pricing means a spike in traffic, or someone deliberately spamming your chat, can raise your bill. Set usage limits and alerts where the tool allows.

Before you launch

Frequently asked questions

Related guides

• AI for your businessAI for customer supportHow to use AI tools to handle common customer questions faster, and the rules for keeping responses accurate, honest, and human where it counts.
• AI for your businessAI and your customer data: privacy basicsWhat happens to customer information you put into AI tools, how to choose settings and plans that protect it, and simple rules your team can follow.
• AI for your businessHow to evaluate an AI tool before you pay for itA practical process for trialling and judging an AI tool for your business, covering real-task testing, data and security terms, pricing models, and your exit plan.
• ComplianceADA & WCAG: web accessibility compliance in the USWhat the Americans with Disabilities Act and the Web Content Accessibility Guidelines mean for your website, and how we build with accessibility in mind.
• Security for your companyVendor and third-party riskHow to check whether a software provider or contractor will keep your data safe, with the questions to ask and how to judge the answers.

Learn more

• OWASP: Top 10 for Large Language Model Applicationsowasp.org
• FTC: Artificial intelligence guidance for businessesftc.gov
• NIST: AI Risk Management Frameworknist.gov