Person using an app on a phone next to a computer

An AI assistant with your company's data: what it can and can't do

An assistant that knows your prices, procedures and catalogue can save your team and your customers a lot of hours. It can also be wrong with complete confidence. What to expect from it, which documents to give it and how to test it safely.

One of the requests we hear most is "we want a ChatGPT that knows our company". It's a sensible idea: an assistant the team can ask about a price, a procedure or the status of something, or that handles customers' first questions with the right information. It works, and set up properly it saves a lot of hours. But it's worth knowing exactly what it is, what it does well and where it goes wrong before putting it in front of anyone.

What it is, without the jargon

An assistant like this combines two things: a language model, which understands the question and writes the answer, and your documents, which are where it gets the information. When someone asks something, the system first searches your documentation for the relevant passages and then writes the answer based on them. It doesn't "learn" your company by heart: it checks your documents every time, just as a new employee would with a good manual to hand.

What it does well

Answering internal questions. How a return is processed, what each price plan includes, who an incident is escalated to, where the template for a given contract is. This is where it pays off most, because the team stops interrupting each other and new people get up to speed sooner.

Handling customers' first line of questions. Opening hours, general availability, which paperwork is needed, how a service works. Always with a clear route to a person when the question gets complicated or the customer asks for one.

Finding scattered information. If your knowledge is spread across PDFs, old emails and shared documents, the assistant puts it all one question away.

What it doesn't do, or doesn't do alone

It doesn't know what isn't written down. If a price changed and nobody updated the document, it will quote the old one with total confidence. An assistant is only as good as the documentation behind it, and keeping that up to date becomes another task.

It can be wrong while sounding certain. That's why it should cite the source of every answer, so whoever reads it can check, and why it shouldn't promise prices, deadlines or terms to a customer without a person confirming them.

It doesn't replace judgement. Exceptions, angry customers, decisions involving money: that's where someone from the team has to step in, and the assistant must know when to hand over.

Which documents to give it, and which not

Start with what the team looks up most often: catalogue, current prices, internal procedures, frequently asked questions, terms of service. Leave out, at least at first, anything containing personal data about customers or employees. And before uploading anything, find out where those documents will be processed, whether the provider can use them to train its models and whether it signs a data processing agreement. With third parties' data, the GDPR applies just as it does to any other tool.

How to test it without risk

Choose a single department and a single use, for example the internal questions of the customer service team. Gather fifty real questions from recent weeks and check how many it answers correctly, how many wrongly and how many it admits it doesn't know. Use it for four weeks inside the team only before putting it in front of customers. If it gets the vast majority right and knows how to say "I don't know", you have a basis for expanding it; if not, the problem is almost always in the documents, not the AI.

If you're still not sure which tasks to tackle first, start with where to start with AI in your business. And if you want an assistant built on your own data with these safeguards from the outset, it's one of the AI systems we implement.

Sources: OpenAI: file search over your own documents · General Data Protection Regulation (GDPR, EUR-Lex) · European Data Protection Board: data protection guide for small business

Take the free GEO test Contact

Keep reading

Related articles

Streaks of blue light suggesting a flow of data
AI & GEO·November 4, 2026 Ranking first on Google is worth less every quarter: the 2026 data The top Google result gets 58% fewer clicks when an AI answer appears, and ChatGPT now has 900 million weekly users. The data, what it means for a Mallorca business and where to start.
Small business team planning tasks on a whiteboard
AI & GEO·October 8, 2026 Where to start with AI in your business: three tasks you can automate in a month Most small businesses start with AI at the wrong end: by buying a tool. How to choose the first task, three that almost always deliver results within weeks, and how to tell whether it was worth it.
Small business team going through paperwork at a table
AI & GEO·September 22, 2026 How much time does your company lose to repetitive work? Take the two-minute test In almost every small company someone spends half their day copying data from one place to another. Nobody decided it: it just happened. Ten questions to put a number on it, and the result on this same page.
Cover: GEO, ranking in AI search engines
AI & GEO·June 18, 2026 GEO: how to rank your website in AI search engines (and why SEO is no longer enough) More and more people ask ChatGPT or read Google's AI answer instead of clicking. GEO is the discipline that gets those AIs to recommend you. Here's what it is and where to start.
Cover: getting ChatGPT to recommend your business
AI & GEO·June 14, 2026 How to get ChatGPT, Gemini and Perplexity to recommend your business AIs don't invent who they recommend: they learn it from what's published online about you. These are the concrete levers to get on their radar.
Cover: Google AI Overviews
AI & GEO·June 8, 2026 Google AI Overviews: what they are and how to appear in AI answers Google now answers many searches with an AI‑generated summary, before the links. Here's how it works and what to do so your site is one of the sources it cites.