Contents7
In November 2022, an Air Canada customer asked the airline’s chatbot whether he could claim the reduced bereavement fare after the fact. The chatbot said yes, within 90 days. The airline’s policy, published on another page of the same website, said the opposite. The customer was reimbursed by court order, and the airline lost the argument it had put forward: that the chatbot was responsible for its own answers.
This dispute comes down to one lesson: what your AI assistant says commits the company as much as a page on your website. And an assistant answers from what it is given to read, which, most of the time, is the knowledge base. It therefore now has three readers, your customers, your teams and your AI, and it has to be written for all three.
In short
An AI assistant answers from your knowledge base: an outdated or contradictory page becomes a wrong answer, given with confidence.
A company is answerable for what its chatbot says, just as for the rest of its website: that is what a Canadian tribunal ruled in 2024 in the Air Canada case.
An article that works for AI covers a single question, gives the answer up front, states its scope, and carries a date and an owner.
The assistant cites its source, hands over when the base is silent, and every question left unanswered becomes the next article.
What happened with Air Canada’s chatbot?
The customer had taken a screenshot of the chatbot’s answer, which invited him to request the reduced fare within 90 days of his ticket being issued. Air Canada did not accept these requests after travel. Before British Columbia’s Civil Resolution Tribunal, the airline argued that the chatbot was, in effect, a separate entity. The tribunal called this a “remarkable submission”, then dismissed it:
It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.
The airline was ordered to pay 812.02 Canadian dollars, including damages, interest and fees. The amount is modest; the reasoning, however, applies to any business: the tribunal noted that Air Canada had not taken reasonable care to ensure its chatbot was accurate, and that a customer has no reason to guess that one part of the website is accurate and another is not. It is a Canadian decision, not a rule that binds you everywhere, but its common sense transfers easily.
Why does your knowledge base have three readers?
Your customers, who look for the answer on their own before writing to support. They read quickly, on their phone, and stop at the first useful sentence.
Your teams, who use it to answer, train a new colleague or check a rule. They need the exceptions, the ones usually kept in someone’s head.
Your AI assistants, which look in it for the passages closest to the question asked and answer from them. That is the principle of RAG, which we explain in RAG for business: an AI that answers from your documents. They read everything, without perspective: a 2023 page forgotten in a corner carries as much weight as the current policy.
AI vendors say so themselves. Microsoft asks, for its agents, that you ensure the content is current and accurate, since the agent answers with it, and that you periodically review sources to remove anything outdated. Anthropic recommends having the assistant cite its sources for each claim, so the answer can be checked.
How do you write an article an AI will quote correctly?
Seven rules, which we apply to any knowledge base before handing it to an assistant:
- One question per article.
The title is the question, asked in the customer’s words: “How do I change the delivery address of an order?”, not “Logistics policy”.
- The answer first.
The first two sentences answer; details, steps and special cases follow.
- The scope written down.
Who the rule applies to, for which products, in which countries, since when. An AI cannot guess a condition nobody wrote down.
- Exceptions in the article.
What support knows “off the top of their head” does not exist for the assistant.
- One source per topic.
Two pages that contradict each other produce a random answer. The second points to the first, or disappears.
- A date and an owner.
Each article carries its last update date and the name of the person responsible for it.
- An access level.
Public, internal or confidential: the assistant open to customers only reads the public level.
An article written to be skimmed
A vague title: “Deliveries”
The rule in the middle of the fourth paragraph
“Except in special cases” without saying which
No date, no owner
An article written to be quoted
The title is the customer’s question
The answer in two sentences, then the detail
Each exception named, with its condition
Updated on 3 October, by the logistics manager
Access rights deserve particular attention: an assistant that sees everything can repeat everything. Our guide Enterprise AI governance: who accesses what explains how to split these rights by department.
What should the assistant do with this knowledge base?
A good knowledge base is not enough: the assistant needs clear instructions on how to use it. Four rules, always the same:
- answer only from the knowledge base, without filling gaps from memory;
- cite the article the answer comes from, so the customer and the colleague alike can check;
- hand over to a person when the base has no answer, rather than improvising;
- log every unanswered question, which becomes the week’s to-do list.
The loop of a living knowledge base
- The question
A customer or a colleague asks the assistant.
- The search
The assistant finds the articles that answer it.
- The answer
Cited with its article, or passed to a person if nothing answers it.
- The log
Each unanswered question is recorded, with its date.
- The article
The topic owner writes or corrects it, and the loop starts again.
It is the same split as in our guide on support: AI sorts and prepares, humans review. Anything involving money, a refund, a discount, a commercial gesture, remains a human decision, however good the knowledge base.
Should you let AI write the knowledge base?
It helps a lot, provided it stays at draft stage. HubSpot actually charges for this service: its knowledge base agent, in beta, uses 200 credits per generated article, or $2 at $0.010 per credit. We converted these rate cards in AI credits in software: what “included” AI really costs.
The right use: AI writes a first draft from resolved tickets and answers approved by the team, and the topic owner reviews, corrects and publishes. The wrong use: publishing without review. The knowledge base then becomes the assistant’s own source of truth, and a published mistake comes back in every answer, looking like a rule.
Watch out
An outdated page is more dangerous than a missing one. With no page, the assistant hands over; with a wrong page, it answers with confidence. When a rule changes, the old page is corrected the same day, or withdrawn.
Where to start?
Start from the questions, not the documents. Export the fifty most frequent questions from your emails and tickets, then check for each one that an article answers it, according to the seven rules. Gaps and contradictions show up at this stage. Then name an owner per topic, give each article a date, and only then connect the assistant.
Our AI assistant on your documents, at €6,500, is installed on a knowledge base prepared in this way: it answers with the source cited, respects access rights by department, and can run entirely on a model hosted on your infrastructure. Through an API, the cost of each answer can be measured and capped: that is the purpose of our AI at the right cost pack. To place this project among the others, our guide Integrating AI in your business: private and at the right cost gives the overall approach.
Frequently asked questions
What is the difference between a knowledge base and an FAQ?
An FAQ gathers a few common questions on one page. A knowledge base covers every topic, one article per question, with its details, its exceptions, its date and its owner. For an AI, the difference matters: an FAQ that is too short pushes it to fill in the gaps itself, whereas a complete knowledge base lets it quote an approved answer, or admit it has none.
Which tool should you choose for your knowledge base?
The tool matters less than the structure. A help centre, a wiki or plain files will do, if they allow one article per question, a visible date and owner, access levels, and reading by the assistant, through an API or an export. Check that last point before choosing: a knowledge base the AI cannot read cleanly is of no use to it.
Is a company responsible for what its chatbot says?
In the Air Canada case, decided in Canada in 2024, yes: the tribunal held that the airline was answerable for all the information on its website, chatbot included. It is not a rule that applies everywhere, but the logic is the same: for a customer, the assistant’s answer is the company’s answer. Hence the value of an accurate knowledge base and of answers that cite their source.
How often should articles be reviewed?
The same day for any article affected by a change of rule, price or product. For the rest, a quarterly review, organised by owner, is generally enough. The list of unanswered questions sets the real pace: if it grows, the knowledge base has fallen behind, and that is what needs dealing with before widening the assistant’s scope.
Sources
- Microsoft, best practices for Microsoft 365 Copilot declarative agents (new tab)learn.microsoft.com
- Microsoft, editing knowledge sources in Copilot Studio (new tab)learn.microsoft.com
- Anthropic, reduce hallucinations (new tab)platform.claude.com
- HubSpot, product and services catalog (new tab)legal.hubspot.com
Microsoft, best practices for Microsoft 365 Copilot declarative agents
Read on learn.microsoft.com(new tab)Jim · founder of Brain Plus AI
Also in business ai
Understand9 min
Enterprise RAG: an AI that answers from your documents
RAG connects an AI model to a company's documents so that it answers by citing its sources. How it works, when it is useful, why it sometimes misses the point, and what to set up before opening it to the team.
Practical guide12 min
Automate customer support: AI sorts, humans review
Incoming tickets and emails read, sorted and prepared by AI, then reviewed by your team when the reply commits the company: how to split the work, what can go out on its own, and why sorting costs a few cents when each step has the right model.
Practical guide12 min
Enterprise AI governance: who can access what
An AI assistant reads whatever your access rights let it read, sharing mistakes included. Here is how to split your knowledge bases by department, keep a common core and decide who writes to the shared memory.