An assistant on your own site

Someone finds your company late on a Sunday evening. They have one question. Does this fit my case, how long does delivery take, can I send it back, what does it cost for a team of twelve. Your website has a contact form, and the form waits until Monday.

By Monday that buyer has asked an assistant somewhere else. The question for a company is no longer whether people want to ask. It is whether they can ask you, on your own site, and get an answer you would sign.

Why it is worth having

The habit is formed elsewhere. Forrester reported in January that 94 percent of business buyers use AI during their buying process, and G2 found in a March survey that half of software buyers now start their research with an AI chatbot more often than with Google. People who are used to asking look for a place to ask.

Where companies have put an assistant in front of their own customers, the numbers they report are large. Amazon said in October 2025 that 250 million customers had used Rufus, its shopping assistant, that year, and that shoppers using it were 60 percent more likely to complete a purchase. Klarna said in 2024 that its assistant handled 2.3 million conversations in its first month, two thirds of its customer service chats, with errands resolved in under two minutes instead of 11 and 25 percent fewer repeat enquiries.

Both are company figures, and the first is a correlation: people who ask questions may be closer to buying anyway. But the direction is consistent. An answer at the moment of the question keeps a buyer, and a good answer saves a ticket.

Answers from your own sources

An assistant on your site speaks for you, and a tribunal has already said so. In February 2024 a tribunal in Canada ordered Air Canada to pay a passenger after the chatbot on its website described a refund rule the airline did not have. Air Canada argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal disagreed: the airline was “responsible for all the information on its website,” whether it came from a static page or a chatbot.

The opposite failure is an assistant that answers from general knowledge. When The Markup tested New York City’s chatbot for businesses in 2024, it told employers they could take a cut of their workers’ tips and landlords that they could turn away tenants with housing vouchers, answers that contradicted the city’s own rules.

So the first rule is simple. A good assistant answers from your sources: the product data, the price list, the terms, the manual, the help pages and the policies actually in force. It shows where an answer comes from. When the sources do not cover a question, it says so instead of improvising.

It knows your products and your rules

Knowing the catalogue is the easy half. A useful assistant also knows the rules around it: what may be promised and what may not, which discounts exist and who may give them, which questions touch money, health or law and belong with a person, and how returns work in your shop, not in shops in general.

Those rules have to be written down before the assistant goes live, in plain sentences, by the people who own them. Most companies discover in this step that some of their rules were never written anywhere. That alone is worth the exercise.

The assistant also has to say what it is. Under the EU AI Act, from 2 August 2026, people must be told when they are dealing with an AI system unless it is obvious. A clear label costs nothing and earns more trust than a human name on a bot.

It hands over to a person

The most important feature of a good assistant is the way out. A buyer with a complaint, a large order or a question the sources cannot answer should reach a person, with the conversation handed over, so nobody has to explain it twice.

The person on the other side gains too. In a study of 5'179 customer support agents, the economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised the issues resolved per hour by 14 percent on average and by 34 percent for novice workers, and that customer sentiment improved.

That is the division of work we believe in. The machine answers the routine question at once, prepares the case and finds the passage. People keep the conversations that need judgement, care or a decision, and a customer who reaches a real person at that moment remembers it.

What to avoid

An assistant without sources, which sounds confident and invents. A bot that pretends to be a person. An endless loop with no way to a human. A chat window that wants an email address before it answers anything. An assistant that can be talked out of its rules because text on a page or in a message tells it to. And a launch without testing: before going live, put the hundred questions your team hears most often to it and read every answer.

How to start

Start small and specific. Collect the questions customers really ask, from the inbox, the phone and the sales calls. Gather the sources that answer them and write down the rules. Decide where a handover goes, who picks it up and at what hours. Then launch on one part of the site, read the conversations every week and fix the sources, not only the answers.

This is the work behind Twin, the brain we build for a company: its knowledge, rules and integrations in one place, so an answer to a customer comes with the passage from the manual, a return follows the company’s policy and a question it is not sure about goes to a person. Studio, where a company runs its website, connects to Twin as its brain. On the phone the same idea is Timbre, our voice assistant for calls, reception, booking and messages, around the clock. And when the data must not leave the company, Twin can run privately on Eigna, built by our partner Eigna Labs: a model trained on the company’s own work, in its own cloud or on its own servers.

The best assistant is the one that knows where its knowledge ends. Build that part first.