Do invoices and delivery notes belong in the AI knowledge base?

No. Records such as invoices, delivery notes and bookings stay in the leading system, i.e. in the ERP or the accounts – that is where they are reliably kept, totalled and retained. At this point, the knowledge base only adds the context, for example the service report or the correspondence that explains why a case was resolved the way it was.

How many documents does a knowledge base need to get started?

There is no minimum. What has proven effective is a basic stock of documents covering the most frequent everyday questions – often that is a few dozen to a few hundred documents. After that, the collection keeps growing in day-to-day operation, because every unanswered question shows which document is missing next.

Do the documents have to be perfectly tidied up before uploading?

No, perfect filing is not a prerequisite. It is enough to sort out clearly outdated versions or mark them as superseded. Because every answer names the source document, old material quickly stands out in day-to-day work and can be cleaned up in a targeted way.

What happens to scanned documents and paper records?

Scanned documents are converted into searchable text by text recognition (OCR) and can then be queried as normal. Paper records have to be scanned first. There are limits with handwriting and very poor scans – cleanly printed service reports and manuals usually work well.

May confidential documents such as personnel records go into the knowledge base?

In principle yes, but only with an access concept that defines who may find which content. It is also important to have a provider that keeps the data on EU servers and does not use customer documents to train third-party AI models. If in doubt, companies start with non-critical collections first.

Which documents belong in an AI knowledge base?

An AI knowledge base should hold documents that record knowledge: instructions, service reports, problem solutions, checklists, standards and internal rules. Records such as invoices, delivery notes or bookings do not belong in it – their home remains the ERP system or the accounts. The simplest rule of thumb for choosing is: the ERP knows what happened – the knowledge base knows how it is done.

In brief

Documented know-how belongs in the knowledge base: instructions, reports, problem solutions, standards and internal rules.
Records such as invoices, delivery notes and bookings stay in the leading system, i.e. in the ERP or the accounts.
The test question is: would someone in the team want to know how something is done or how something is regulated? Then the document belongs in it.
Inconspicuous formats count too: email threads with problem solutions, meeting notes and scanned service reports.
A manageable basic stock of documents is enough to start; the collection then keeps growing in day-to-day operation.

How can you tell whether a document belongs in the knowledge base?

The most reliable test question is simple: could someone in the team look for this document because they want to know how something is done, why something was decided the way it was, or what to do in a particular case? If so, it belongs in the knowledge base. If, on the other hand, the document only records that a transaction took place – delivered, invoiced, booked – then it is a record and stays in the system that was built for it.

This distinction between know-how and transaction data sounds trivial, but in practice it avoids the most common disappointment: a knowledge base that has been filled with thousands of copies of records and still cannot answer a single question about the right way to proceed. The value lies not in the number of documents but in the knowledge they contain.

An AI knowledge base plays to its strengths where knowledge exists in text form and needs to be found with a question in everyday language. How this works technically, and why uploading individual files to ChatGPT is not a long-term solution for this, is described in the guide Using ChatGPT with your own data.

Which documents are well suited?

The most valuable documents are the ones people ask about most often in everyday work. In most businesses, these fall into five groups.

Instructions and manuals. Installation instructions, operating instructions, set-up and maintenance instructions – whether written in-house or taken over from the manufacturer. Manufacturer documents in particular are often stored as PDFs somewhere on the drive and are still searched for afresh on every job.

Service reports and problem solutions. The report on how a fault at a customer’s site was fixed last year is one of the most valuable documents in the business. Exactly this experience is lost if it exists only in one employee’s head. Anyone who records solutions – even as a short note – builds up real company knowledge.

Standards, checklists and templates. Quality requirements, acceptance checklists, safety briefings, standard text blocks for recurring letters: everything that describes how things are done properly in your business.

Internal rules. Holiday rules, expenses rules, key allocation, cover arrangements. Otherwise, questions about these keep landing with the same two people in the office – the knowledge base answers them quickly and names the valid document.

Project and customer knowledge. Meeting minutes, project notes and the correspondence that explains why something was agreed the way it was. A sense of proportion is worthwhile here: not every email needs to go in, but the message explaining the reasons for a special solution certainly does.

Which documents are better kept in the leading system?

The leading system is the program that is formally responsible for a type of data. For records and stock, this is usually the ERP system (enterprise resource planning – the business software for orders, stock and invoicing) or the accounts. These systems can do something a knowledge base does not even try to do: keep figures reliably, form totals, monitor deadlines and fulfil statutory retention obligations.

Belongs in the knowledge base	Stays in the leading system
Installation, operating and maintenance instructions	Invoices and payment records (accounts)
Service reports and documented fault solutions	Delivery notes and orders (ERP)
Internal rules such as holiday or expenses rules	Stock levels and item master data (ERP)
Checklists, standards and quality requirements	Open items and dunning (accounts)
Meeting minutes and project notes	Payroll accounts and payroll (payroll software)
Explanatory correspondence about a case	Order and quote data as a record (ERP/CRM)

Of course there are borderline cases. A warranty case includes an invoice – that is a record and stays in the ERP. But the same case also includes the service report and the email explaining the reasons for the goodwill decision – that is knowledge and belongs in the knowledge base. So the knowledge base does not replace the ERP; it adds the why and the how that no record contains.

How do you make the selection in practice?

Instead of sorting for months, a pragmatic approach in five steps has proven effective:

Start with the most frequent questions. For one to two weeks, collect the questions that are actually asked in the business – on the phone, in the workshop, in the office. This list is your selection criterion.
Gather the relevant documents. For each frequent question, collect the documents that answer it: from folders, drives and inboxes.
Sort out what is obviously outdated. No big clear-out needed – it is enough to leave out clearly outdated versions or mark them clearly as superseded.
Include scans. Scanned service reports and old manuals are made readable by text recognition (OCR – the automatic conversion of images into searchable text) and can then be queried just as well.
Test with real questions and add more. If the test reveals a gap, a document is usually missing. That is not a fault of the system but the most valuable indication of which knowledge should be documented next.

What about confidential documents?

Confidentiality is not a reason for exclusion, but it is a planning issue. Two questions should be settled before documents are added: who may find this content – and where is the data stored? An access concept ensures that, for example, rules are visible to everyone, but internal project matters only to the team concerned.

On the provider side, three commitments count: data storage on EU servers, no use of customer documents to train third-party AI models, and a source for every answer so that every statement can be checked in the original document. For particularly sensitive collections such as personnel files, the rule is: only add them once access rights have been set up properly – if in doubt, the knowledge base starts with the non-critical collections and then grows in a controlled way.

How do you get started in concrete terms?

The proven route starts small: a basic stock of documents covering the twenty most frequent questions, then a testing phase with real requests from everyday work, then step-by-step expansion. How this introduction works step by step – from selecting documents through the testing phase to live operation – is shown on the page Process of the introduction. The product page gives an overview of the features and data storage of Goma-it’s AI knowledge base.