AI chatbot for your website: 24/7 help for ecommerce and B2B

In short: an AI chatbot for your website is an assistant that answers visitors in plain language, at any hour, using your company's own data: product catalogue, terms of sale, order status, technical sheets. It works when it reads from verified sources, hands the conversation to a person when needed and complies with the GDPR and the EU AI Act.

Key points

The value is not in the language model but in the data the assistant can reach: catalogue, orders, CRM, technical documentation.
An assistant without clear rules for handing over to a human loses exactly the conversations worth the most.
A company is responsible for what its chat says, just as it is for what its web pages say.
Customers' top fear is no longer being able to reach a person: keep that option visible at all times.
Start with a narrow scope and an accuracy threshold agreed before launch.

An online store gets the same questions around the clock: delivery times, returns, sizes, stock. A B2B company gets harder ones: compatibility between parts, certifications, minimum order quantities, lead times for different volumes. A well-designed AI assistant covers that gap and routes the request to the right system; a badly designed one gives wrong answers with total confidence. For the bigger picture on how business buyers now purchase online, see the opportunities and limits of B2B ecommerce.

What is an AI chatbot for a website, and how is it different from a rule-based bot?

An AI chatbot for a website is conversational software built on a large language model (LLM) that understands free-text questions and writes answers, instead of following a fixed decision tree. A rule-based chatbot offers buttons and predefined paths: it is predictable, but it gets stuck as soon as a question falls outside the script.

But the difference that matters is where the assistant gets its information. A generic model answers from what it learned in training, and it knows nothing about your prices or your terms. An assistant connected to company data uses RAG (retrieval-augmented generation): it first retrieves the relevant documents from your sources, then writes the answer based on them.

Type	How it answers	Strength	Limit	Best for
Rule-based chatbot	Predefined paths and buttons	Predictable and cheap	Rigid, frustrating for off-script questions	Simple FAQs, routing requests
Generic AI assistant	Language model with no company data	Natural conversation	Can make up prices, policies and specs	Not advisable on a business website
AI assistant connected to your data	RAG over catalogue and documents, plus actions via API	Specific, up-to-date answers	Needs clean data, testing and upkeep	Ecommerce and B2B companies with structured catalogues and processes

What can a virtual assistant answer for ecommerce and B2B?

It can answer anything written down in a reliable, current source, and nothing that is not. A practical rule: if one of your staff would need to open a document or a back-office system to answer, the assistant must be able to read the same document or system.

Ecommerce examples

Availability of a variant (size, colour, format) read live from the catalogue.
Order status, after a check with order number and email address.
Returns and shipping terms by country, taken from the official page rather than rewritten separately.
Product suggestions based on catalogue attributes, such as "I need a washable rug, 160×230 cm".

For the in-store side, see how AI can strengthen retail.

B2B examples

Technical specs and part compatibility, taken from product sheets and manuals.
Certifications, applicable standards, downloadable documentation.
Lead qualification (industry, volumes, timing, region) before passing the request to sales.
Answers in the customer's language, useful for exporters: in that case the test set must be repeated for every language you serve, and technical terminology must be checked (see AI translation software).

What it must not do: invent discounts, confirm delivery dates nobody has checked, or give legal or technical opinions that bind the company. These limits go into the assistant's instructions and are checked in testing, question by question.

Can it recommend products and offer personalised deals?

Yes, but it helps to separate two levels of personalisation, because they come with different requirements.

Personalisation based on the conversation. The assistant uses what the visitor writes and the page they are viewing: it asks two or three questions (use, size, quantity), filters the catalogue by attribute and suggests compatible alternatives or the accessories needed. In an online store it is the shop assistant asking "where will you use it?"; in a B2B company it is the engineer pointing to the right mounting kit for a given part number.
Personalisation based on behaviour. Showing content or offers based on earlier visits and pages viewed is profiling. If it relies on cookies or other tracking tools, consent rules apply: the Italian data protection authority (Garante), for example, in its guidelines of 10 June 2021 allows them to be used to tailor the service only after the user's consent has been obtained.

Either way, offers must come from rules the company has set: active promotions, price lists by customer type, quantity breaks. The model can explain an offer that exists; it must not create one. In B2B, negotiated prices stay behind a login or go to sales.

There is also an often overlooked use: conversations are a running analysis of what your website is missing. If many people ask for the same size, the same certification or the same lead time, that information belongs on the product page or the relevant page, not only in the chat. The same questions also feed content and campaigns, as I explain in digital marketing tools for B2B.

What are the benefits, and under what conditions?

The benefits are real, but each depends on a specific condition. Without it, the same benefit turns into a cost.

Benefit	You get it if	You lose it if
Answers outside office hours	Frequent questions have an official, up-to-date source	The chat promises answers nobody can give until the next day
Less repetitive work for the team	Simple requests are closed in the chat and the rest arrive with the details already collected	The customer has to repeat everything to the agent
Help during purchase	The assistant reads real stock, variants and terms while the customer is choosing	It recommends products that are out of stock or incompatible
Better-qualified leads	Qualifying questions are agreed with sales and end up in the CRM	The chat collects contacts with no context, to be called back from scratch
Data on customer questions	Someone reads the conversations and updates the site and catalogue	Logs are archived and nobody opens them

For the lead acquisition side, see also B2B lead generation solutions.

How do you add an AI chatbot to your website: off-the-shelf platform or custom build?

There are three routes, and the choice depends on where your data lives and how many actions the assistant has to take, more than on the initial budget.

Route	How it works	Fits when	Check
Off-the-shelf platform	A widget added to the site, set up on uploaded pages, FAQs and documents	Informational questions on content that rarely changes	Where data is stored, conversation export, handover to your helpdesk
Platform with integrations	A widget connected to catalogue, orders and CRM through APIs and automation tools	Ecommerce and B2B companies that want answers on live data and leads in the CRM	Integration permissions, index refreshes, error monitoring
Custom build	Interface, orchestration and data access built for the specific case	Unusual processes, customer portals, strict security constraints	Maintenance costs and the skills to run it over time

Before choosing a platform, check these points:

Can you restrict the assistant to your sources and see which document each answer comes from?
Is there a data processing agreement (GDPR Article 28), and does it state the countries where data is processed?
Can conversations be exported and deleted in line with your retention periods?
Does the handover to a human work with your helpdesk or CRM, with the transcript attached?
Can the widget load only on click or after consent, without slowing your pages down?
Is the usage-based cost predictable at your conversation volumes?

If your site runs on WordPress, the same assessment applies, as I describe in AI in CMSs and WordPress.

How does it connect to your catalogue, CRM and other systems?

It connects in two ways: reading, to look things up, and writing, to take actions. Both go through APIs and an orchestration layer that decides which source to query and which action is allowed.

Reading: the catalogue (Shopify, WooCommerce, a PIM or a database), policy pages, technical PDFs. Documents are indexed and refreshed whenever they change, so the chat never quotes an outdated price list.
Writing: creating a contact or a deal in the CRM, opening a ticket, booking a call, sending a summary to sales. Here the assistant uses tools (also called function calling): specific functions it can call with controlled parameters.

For an assistant on my own site I used a four-step pattern: the widget sends the message to an n8n webhook; an AI agent with session memory decides whether to answer directly or call a tool; the tool queries the database (in that case, the blog articles stored in Supabase); the answer goes back to the widget. For an online store the tool queries catalogue and orders; for a B2B company, technical sheets and the CRM. The same pattern works with Make or custom code: what matters is keeping model, sources and actions separate.

The CRM connection is where the chat stops being only a support cost and becomes part of lead generation: every qualified conversation reaches sales with the context already written down. It is the same principle behind business process automation: data is entered once and flows between systems without retyping. More cases like this are collected in the automation section of the blog.

When should the conversation be handed to a person?

Whenever the answer commits the company, whenever the customer asks, and whenever the assistant cannot find a reliable source. The handover to a human (handoff) has to be designed before launch, not bolted on after the first complaint.

The criteria I start from, to be adapted case by case:

The visitor asks to speak to a person: no attempt to keep them in the bot.
The question is about negotiated prices, discounts, contracts, warranties or complaints.
The potential value is above a threshold: a large B2B order, a custom quote.
The search returns no relevant documents, or returns documents that contradict each other.
The message signals frustration or urgency.

Outside working hours, the handover means collecting the details properly (name, company, request, contact) with a realistic reply time, not promising "an agent will be with you shortly" at 11 pm on a Saturday.

Do customers trust an AI assistant?

Not by default. In a Gartner survey of 5,728 customers, carried out in December 2023 and published in July 2024, 64% said they would prefer companies not to use AI in customer service, and 53% said they would consider switching to a competitor if a company started using it. Their top concern was that it would become harder to reach a person, followed by AI displacing jobs and AI giving wrong answers.

For a website, this translates into four design choices:

State straight away, in one line, that it is an AI assistant and what it can do.
Keep the option to talk to a person visible at all times.
When the conversation moves to an agent, let them pick up where the chat left off, with the transcript: Gartner makes the same recommendation in its release.
Write short, natural answers without repeated pleasantries: in B2B, people writing in want a technical fact, not small talk.

In practice, trust is built with correct answers and a clear way out, as in any B2B business relationship.

How far can you trust an AI chatbot's answers?

As far as you have tested it. A language model always produces a plausible answer, even without the information: you need verified sources, tight instructions and regular spot checks.

The best-known case is Air Canada. In November 2022 the airline's website chatbot told a customer he could claim a bereavement fare after travelling, which the airline's policy did not allow. On 14 February 2024 the British Columbia Civil Resolution Tribunal sided with the customer: the company is responsible for all the information on its website, whether it comes from a static page or a chatbot. Air Canada had to pay CAD 812.02 in damages, interest and fees.

Then there is the McDonald's drive-thru case. From October 2021 McDonald's tested automated voice ordering with IBM in about 100 drive-thrus, and in 2024 decided to end the test. According to Nation's Restaurant News, the company wanted order accuracy above 95% before a wider rollout, while a 2022 BTIG report put it in the low 80% range; the system's mistakes had also ended up in videos shared on TikTok. The lesson applies to websites too: set the quality bar first, and only scale when the numbers reach it.

Positive results need careful reading too. In February 2024 Klarna reported that in its first month its AI assistant had handled 2.3 million conversations, two thirds of its customer service chats, with resolution times down from 11 minutes to under 2. In May 2025, however, its CEO told Bloomberg that focusing too much on cost had led to lower quality, and that Klarna was hiring human agents again so customers can always reach a person. AI pays off on frequent, well-documented questions; the rest needs a route to a human.

Finally, security. Prompt injection is an attack in which text typed by a user, or read from an external source, changes how the model behaves; OWASP ranks it first among the risks of LLM applications. In practice the assistant should have only the permissions it strictly needs, and every write action should go through server-side checks. I discuss the wider topic in AI and cybersecurity and, on control and explainability, in governance and explainability of AI in business.

What do the GDPR and the EU AI Act require from a website chatbot?

They require transparency and control over data. A chat collects personal data (names, emails, sometimes order numbers or company details), so the GDPR applies in full; since 2 August 2026 the AI Act transparency rules apply as well.

Say it is AI. Article 50 of the AI Act requires systems meant to interact with people to be designed so that those people know they are dealing with an AI system, unless it is obvious. The duty falls on the provider, meaning whoever builds the assistant (possibly your own company): either way, add a visible line when the chat opens.
Updated privacy notice. Article 13 of the GDPR lists what to tell people when you collect their data: purposes, legal basis, recipients, retention periods, any transfers outside the EU. The chat belongs in the privacy notice like any other form.
Contracts with vendors. The chat platform, the model provider and the automation tools all process data on your behalf; Article 28 requires a written agreement with each processor.
Data minimisation. Ask only for the data needed to answer, set an expiry for conversation logs, and check in the contracts that your data is not used to train the vendors' models.
Consent for third-party scripts. If the widget uses non-essential cookies or tracking, load it after consent or when the user clicks, like the rest of your tracking.

How to put it into practice: first steps

Start with a narrow, measurable scope, then widen it.

Analyse real requests. Group by topic the last 200-300 questions received by email, phone and forms: the 10-15 most frequent are your first scope.
Prepare the sources. For each topic, name the official document or system the answer must come from. If a source is missing or contradictory, fix it before building the chat.
Set thresholds and handoff rules. Agree the minimum share of correct answers on a set of real questions (for example 90-95%), the handover criteria, and reply hours and channels.
Connect the CRM. Every qualified conversation should create or update a contact with the transcript and the collected details, so sales never start from scratch.
Measure for 60-90 days. Track the indicators in the table below, then decide whether to widen the scope.

Metric	What it measures	Where to find it
Resolution rate	Conversations closed without human help	Chat logs
Accuracy	Correct answers in a manually reviewed sample	Weekly review
Handoff	Conversations passed to a person and time to pick them up	CRM or helpdesk
Qualified leads	Contacts from the chat that reach a quote	CRM
Repeat enquiries	Customers writing again about the same question	Helpdesk and email

A hypothetical example: a B2B distributor of plumbing components with a catalogue of 5,000 SKUs. The first scope is not "answer everything" but three questions: part compatibility, stock availability, technical documentation. Everything else, prices included, goes to sales already qualified.

To set these metrics consistently with your other AI projects, see how to measure the value of AI with the right KPIs. Designing assistants and automations connected to catalogue and CRM is part of the solutions I build for ecommerce and B2B companies.

Mistakes to avoid

Pasting policies into the instructions instead of reading them from the official source: after the first update, the chat says one thing and the page another.
Hiding the handover to a human, or making it hard to ask for.
Never reading the conversations: they are the best source for spotting what your site and catalogue are missing.
Loading the full widget on every page at page load: Google's Lighthouse documentation lists chat widgets among the resources to load with a facade, a static element that is replaced by the real widget only when the user clicks it.
Letting the model make up discounts or terms: offers must come from rules and price lists set by the company.

Frequently asked questions (FAQ)

What is the difference between a chatbot and an AI virtual assistant?

A rule-based chatbot follows predefined paths with buttons and fixed replies. An AI virtual assistant understands free-text questions through a language model and, when connected to company data, answers with specific information about products, orders and terms. On a business website it only makes sense if there are reliable sources for it to read from.

Can an AI chatbot replace customer service?

No. It can handle frequent, well-documented requests and collect the details of complex ones, but anything that commits the company, such as discounts, complaints and contracts, must reach a person.

How much does an AI assistant for a website cost?

The cost has three parts: the platform or development, language model usage (billed on the volume of text processed) and the integration work with catalogue and CRM. Integration is usually the most variable item, because it depends on how clean the starting data is.

Does an AI chatbot have to disclose that it is AI?

Yes. Since 2 August 2026, Article 50 of the EU AI Act requires that people be informed when they interact with an AI system, unless it is obvious from the context. A clear sentence when the chat opens is the simplest way to comply.

Who is liable if the chatbot gives wrong information?

The company that runs it on its website. In Moffatt v. Air Canada (2024) the tribunal held the airline responsible for its chatbot's statements just as for its static pages. To reduce the risk you need official sources, explicit limits and a handover to a human on sensitive topics.

Can an AI chatbot connect to Shopify, WooCommerce or a CRM?

Yes, through APIs. The catalogue is read to answer questions about products and stock, and the CRM is updated to log contacts and qualified requests. Automation tools such as n8n or Make orchestrate these steps without rebuilding the website.

Is an AI chatbot or a live chat with human agents better?

They are not alternatives. The AI assistant covers frequent questions at any hour and collects the details of the others; human agents handle anything that commits the company and complex cases. In the Gartner survey published in 2024, customers' top fear was precisely no longer being able to reach a person: the most robust setup is an assistant that hands the conversation to an agent without making the customer repeat anything.

Sources

If you want to know which of your website's requests an AI assistant could handle, and with which data, start with a free assessment.