In November 2022, Jake Moffatt asked the chatbot on Air Canada's website how the discount for travel after a death in the family worked. The chatbot told him he could apply for that discount afterwards. Another page on the same website said he could not. When Air Canada refused to pay, the airline argued that the chatbot was responsible for its own answers. In February 2024 the Civil Resolution Tribunal of British Columbia called that defence remarkable and ordered Air Canada to pay compensation. The amount was small, the lesson is big: when your assistant says something, you say it. And when two of your own sources contradict each other, the assistant picks one, with exactly the same confidence.
An assistant that answers questions based on your own documents is at the top of the list for many organisations. Think of an HR assistant that answers questions about leave and expense claims, a knowledge base in which advisers find earlier advice, or an assistant that prepares quotes based on previous quotes. How the technology behind it works is explained in What is RAG?. This article is about the part that decides most projects and is most often underestimated: the state of the documents underneath.
Why the assistant gets no better than your sources
An assistant on your own sources looks up, for every question, the passages that seem to fit and builds an answer from them. If there are three versions of the leave policy in the folder, it may quote the outdated one. If two documents contradict each other, it picks one. And if an exception only lives in the head of an experienced colleague, that exception does not exist for the assistant. That is not a fault of the model. A new colleague who only gets the folder gives the same wrong answers. The difference is that a new colleague starts to doubt and checks, while an assistant only does that if you have explicitly taught it to.
New York showed how this plays out at scale. In October 2023 the city launched the MyCity chatbot for business owners. Journalists from The Markup tested the chatbot in March 2024 and got answers that went against the law: employers would be allowed to keep part of their staff's tips, and landlords would be allowed to refuse tenants with a housing voucher. A housing expert called the answers incomplete and, at worst, dangerously inaccurate. The city kept the chatbot online anyway, until the new mayor announced in January 2026 that he would shut it down.
Step 1: decide which documents count
Don't start by building, start with an inventory. Which documents may the assistant use, and which should it not? Remove old versions, drafts and duplicate files, or move them to an archive the assistant cannot reach. Give every source an owner who keeps it up to date, and agree on what happens when a policy changes: who replaces the document, and when. This is often the most work in the whole project, and it pays off straight away, even without AI. An HR department that keeps its policies in one place, with one valid version per policy, also answers questions faster itself.
Look at permissions as well. Microsoft Copilot follows the permissions your organisation has already set, Microsoft writes. That is good news and a risk at the same time. If a folder with salary data or sick-leave files is accidentally open to the whole department, the assistant can draw on it too. What went unnoticed for years because nobody found that folder is suddenly one question away with an assistant. So check who can see which folders before you start.
Step 2: capture the knowledge that isn't written down yet
In many processes the real knowledge sits in the exceptions. When do you deviate from the standard rate? Which client always gets an extra check? For which question do you call the lawyer first? An assistant can only work with those exceptions if they are written down somewhere. A simple decision tree or a list of fixed considerations makes the difference between an assistant that is usually right and an assistant you can trust. Collect that knowledge with the people who do the work now: have them go through a pile of difficult questions from the past month and note for each question why they chose what they chose.
Stichting DEN faced the same question in the cultural sector. Knowledge about funding schemes, toolkits and websites was spread across many different places. NewWorks built a web app for DEN with several assistants, each with its own role, tone and scope, all drawing from one centrally managed knowledge base. The most important design decision was not in the technology but in the management: DEN fills and updates that knowledge base itself, without a developer. That way the answers stay correct when the schemes change.
Step 3: test with real questions before you go live
Collect, in advance, questions that employees actually ask, with the right answer and the source. Take them from the HR mailbox, from the ticketing system or from what new colleagues ask in their first weeks. Have the assistant answer those questions before it goes live, and judge every answer on three points: is it correct, does it cite the right source, and does it honestly say it doesn't know when the answer isn't in the documents? That shows you exactly where it goes wrong. Keep that set. After every change to the documents or the instructions you run it again, and you see whether things get better or worse.
Also include questions the assistant should not answer. An HR assistant that gets a question about a colleague's sick leave should refer the person on instead of guessing. Those are exactly the questions that show whether the boundaries you have set actually work.
Step 4: build in a safety net
Decide for which questions the assistant refers to a person, what it says when the answer isn't in the sources and who may see which documents. Have it cite the source with every answer, with the excerpt from the document. At Bouwend Nederland that is the heart of the solution: the tool queries dozens of documents at once with the same questions and shows, with every answer, the passage it is based on, so employees can check for themselves whether the interpretation is right. A user who sees the source catches mistakes themselves. A user who only gets a confident answer has to trust blindly.
Choose deliberately where the assistant stops, too. An assistant that prepares a draft for an employee to check and complete suits most processes better than an assistant that replies to customers by itself. Air Canada had made that choice the other way round. Only when your test set shows that the answers are correct do you move the boundary.
How NewWorks approaches this
Every engagement at NewWorks starts with a validation phase: which problem are we solving, which data is available, who is the internal owner and how do we measure success? For an assistant on your own documents that means: first the sources, the owner and the test questions, then the build. In an AI Build Sprint one working solution is in production within four to eight weeks, built with your own team, so that someone inside your organisation can manage it once we are gone. Why that preparation matters so much is also explained in Why your AI pilot never reaches production.
Want to take a step this week? Pick one question colleagues often ask each other, look up which documents contain the answer and count how many versions you come across. That tells you more about how ready you are for an assistant than any demo.
Curious how this would work for you?
Plan an introduction
