AI assistants

AI assistant on company knowledge: when and how to build one

An assistant that answers customers and staff from your own documents. When a ChatGPT Project is enough, when you need a dedicated one, and how to roll it out.

Bartosz CzerniawskiFounder of dajtoAI
Published: 6 min read
A person at a desk reading an AI assistant chat, with a binder of documents beside the laptop
Contents

An AI assistant on company knowledge is a chat that answers questions from customers or staff using your documents: the offer, terms, price list, procedures. It doesn’t guess from the model’s general knowledge. It looks for the answer in what you give it, and if the answer isn’t there, it says so and passes the case to a person.

Below: when an assistant like this pays off, how it differs from a ChatGPT Project, and how to build one in five steps. Each step has checklist questions. Legal status as of 23 October 2026.

When does an assistant on company knowledge make sense?

When the same questions come back every day and the answers are already written down somewhere. A customer asks about delivery times that are in your terms. A new hire asks how to issue a credit note, and the procedure sits on the shared drive. Someone answers by hand, for the hundredth time.

If you get a few questions a week, it’s not worth the effort. If you get dozens a day, or they arrive in the evening when nobody replies, the assistant starts paying for itself.

A Project in ChatGPT or Claude, or a dedicated assistant?

These are two different things, and it’s worth separating them from the start.

A Project in ChatGPT or Claude (you upload documents, add instructions, the team asks questions) is enough when a few people inside the company use the knowledge. Everyone has an account, sees the sources and judges the answer themselves. The cost is a subscription, and setup takes an afternoon.

A dedicated assistant is needed when customers ask. It runs on your website or in a messenger, has its own boundaries, hands conversations to a person, collects contacts and can be measured. Uploading files isn’t enough here. It has to be designed.

Step 1: Gather the knowledge. Which documents are the source of truth?

An assistant is only as good as the documents it gets. Collect whatever you answer from today: the offer, price list, terms, FAQ, procedures, the most common emails. Then check them for contradictions. If the price list on the website says something different from a PDF from 2024, the assistant will quote one of them and won’t know which is current.

Checklist:

  • Which documents are current, and which should go?
  • Do the documents contain anything customers shouldn’t see (margins, staff details)?
  • Who in the company keeps the knowledge up to date?

Step 2: Set boundaries. What won’t the assistant say, and when does it hand over?

This is the most important step and the one most often skipped. Write down which topics the assistant covers and which it doesn’t. For example: it answers about the offer and lead times, it doesn’t price custom jobs, it doesn’t give legal advice, it doesn’t promise discounts.

Add a handover rule: when the assistant passes the conversation to a person, and how (form, email, phone, Calendly). A good assistant knows what it doesn’t know. A bad one makes up an answer that sounds convincing.

Checklist:

  • Which topics should the assistant stay away from?
  • What happens when a customer asks about something outside the documents?
  • Who receives handed-over conversations, and how fast do they reply?

Step 3: Say it’s AI. What does the AI Act require?

Article 50(1) of the AI Act has applied since 2 August 2026. An AI system that talks directly to people must be designed so they know they’re dealing with AI, unless that’s obvious. In practice: a clear note at the start of the chat, for example “I’m Company X’s AI assistant. If you’d rather talk to a person, type ‘agent’.”

The second topic is personal data. If the assistant collects names, emails or order numbers, describe this in your privacy policy and check the data processing agreement with the tool provider. We cover AI labelling in more detail in our post on Article 50 of the AI Act.

Checklist:

  • Does the customer know from the first message that they’re talking to AI?
  • What personal data reaches the assistant, and where is it stored?
  • Does the privacy policy describe the assistant?

Step 4: Test on real questions. Who is liable for a mistake?

The business. In 2024, a tribunal in British Columbia (Canada) heard the case of an Air Canada customer who was given the wrong discount rules by the airline’s website chat. The airline argued it wasn’t responsible for what the chat said. The tribunal rejected that: the chat is part of the company’s website, and the company is responsible for what’s on its website. It’s a Canadian decision, but the logic travels well.

So before launch, collect 30–50 real customer questions (from emails, chat, phone calls) and run them through the assistant. Check every answer. Add tricky ones too: about discounts, competitors, things that aren’t in the documents.

Checklist:

  • Does the assistant answer the most common questions correctly?
  • What does it do when it doesn’t know: admit it, or make something up?
  • Can it be talked into a promise the company wouldn’t make?

Step 5: Update it and read the conversations. What happens after launch?

An assistant isn’t a one-off project. The offer changes, and so does the price list. Decide who updates the documents and how quickly a change reaches the assistant.

Read the conversations too, at least weekly at the start. You’ll see what customers really ask, where the assistant gets lost and which information is missing from your website. That’s often the most valuable part of the whole rollout.

Checklist:

  • Who reviews conversations, and how often?
  • How quickly does a change in the offer reach the assistant?
  • Which questions can’t the assistant handle, and does it need more knowledge?

On your own or with help?

A ChatGPT or Claude Project for your team is something you can set up yourself in an afternoon. An assistant for customers is a proper rollout: boundaries, handover to a person, testing, AI Act and GDPR compliance. That’s what we do in AI assistants. A custom build starts at PLN 3,900 net (+23% VAT), and a website assistant on subscription starts at PLN 990 net per month (+23% VAT).

Not sure which tool to start with? See our comparison of ChatGPT, Claude, Copilot and Gemini for business.

Summary

An assistant on company knowledge makes sense when the same questions come back every day and the answers are already written down. For your team, a ChatGPT or Claude Project is often enough. An assistant for customers needs organised knowledge, clear boundaries and a note that it’s AI. Test it on real questions before launch, and read the conversations regularly after.

One thing to do today: write down the 10 questions customers ask most often, and check whether each answer is written down somewhere.

Sources

  1. Regulation (EU) 2024/1689 (AI Act), Article 50, EUR-Lex.
  2. Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal (British Columbia), 14 February 2024.
  • AI assistant
  • chatbot
  • knowledge base
  • AI Act

Written with the help of AI, edited and checked by a human.

Bartosz CzerniawskiFor 10 years he made films, campaigns and content for businesses. Today he builds AI tools and processes.More about me
Call · 30 min · free

Facing a similar problem?

Book a 30-minute call.

Book a free 30-minute call(opens in a new tab)
All posts
Newsletter

AI in business, straight to the point

What works, what it costs and what to avoid. Straight to your inbox, no spam. The newsletter is in Polish.

The data controller is Moon Lion Media Bartosz Czerniawski. Sign-ups are handled by Kit. More in the privacy policy.

Manage cookies

Essential · always on

They remember your choice and keep the site secure. The site won't work without them.