Connecting documents and systems: how your company knowledge gets into the AI

For an AI knowledge base to answer questions from your company, the knowledge first has to get into it. And the most valuable knowledge is rarely in invoices or delivery notes – it is in what your people can do and have decided: how the fitter fixed the recurring fault on the system, what was agreed with the customer by email, why a project was carried out this way and not another. This know-how is spread across voice notes, mailboxes, service records, instructions, project folders – and to a considerable extent across people’s heads. Connecting works differently for each source: much of it runs automatically, and in some places deliberate manual work remains. What matters is not to tackle everything at once, but to start with the sources richest in knowledge and expand step by step. The basics are explained in the article What is an AI knowledge base?.

In brief

• The greatest added value lies in documented company know-how: voice notes and dictations, emails with decisions, service records, instructions and project documentation – not in records.
• Records remain leading in the ERP: the ERP knows what was sold – the knowledge base knows how it is done.
• Digital documents can largely be taken over automatically; for paper and photos, text recognition (OCR) converts them into searchable text.
• The system sets attributes such as customer, project, system, date and topic itself; a person confirms the document type, because it determines who may see the document.
• You can get knowledge that exists only in people’s heads out with a dictation routine, short interviews and handover documentation – this is a matter of habit, not of technology.
• Connect sources according to knowledge density and effort: first network folders and dictations, then emails, then the prioritised paper archive, finally specialist systems.
• The originals stay where they are: the knowledge base works with copies, and your existing filing continues unchanged.

Which document sources are there in a typical SME?

A company’s knowledge is almost never in one place. It is spread across the file server, thousands of emails, files in the cupboard, photos on the fitter’s mobile phone, data in specialist programs – and the experience of long-serving employees that was never written down. This fragmentation is exactly why employees lose time every day searching and asking; the article Time lost searching for documents looks at where this time is lost in everyday work.

When connecting to an AI knowledge base, it is worth making a distinction that is often overlooked: not every source is worth the same. Sources with documented know-how – service records, instructions, project documentation, emails with agreements, dictations – will later answer the questions that really cost time in everyday work: “How did we solve that last time?” Sources with records – invoices, delivery notes, accounting data – on the other hand belong in the systems where they already are. This results in the following overview:

Source	What it contains	Connection	Degree of automation
Voice notes / dictations	Experience from jobs and projects – often documented nowhere else	Recording is automatically converted into text	High – dictating and confirming the document type remain with people
Email mailbox (Outlook / Microsoft 365)	Agreements, decisions, commitments including attachments	Relevant emails and attachments are taken over according to rules	High – define rules once
Network folders / file server	Instructions, service records, project documentation, contracts	Folders are read and continuously monitored	High – largely runs by itself after setup
Smartphone photos	Rating plates, completed records, condition documentation	Upload in the browser, then text recognition (OCR)	Medium – taking the photo remains manual work
Paper archive	Old instructions, maintenance logbooks, signed records	Scanning, then text recognition (OCR)	Medium – reviewing and scanning remain manual work
SharePoint / DMS	Documents already filed in a structured way	Interface or regular export	Varies – depends on the system
ERP	Orders, records, master data – deliberately remains leading	Usually not at all, or only for accompanying documents	Deliberate separation instead of connection

The order in this table is no coincidence: know-how at the top, records at the bottom. The following sections go through the sources one by one – each with what happens automatically, where manual work remains and which stumbling blocks appear in practice.

How do you capture knowledge that exists only in people’s heads?

Before we get to folders and mailboxes, the most uncomfortable truth needs to be put on the table: the most valuable knowledge in many companies is not in a single document. The service technician who knows every system in the installed base. The office manager who knows which customer has which special agreement. As long as this knowledge exists only in people’s heads, no software can find it – and with every resignation, every sick leave and every retirement, a piece of it is irretrievably lost.

The good news: documenting this knowledge is much easier today than it used to be, because nobody has to type any more. Three routes have proved their worth.

Voice notes and dictations are the route with the lowest hurdle. After the job, the fitter dictates for two or three minutes on his mobile phone: what the problem was, what the cause was, how he solved it, what to watch out for next time. The recording is automatically converted into text (transcription), tagged and assigned to the customer or the system; afterwards the fitter only has to confirm what kind of note it is. Two minutes of speaking become a permanently findable building block of knowledge – without writing, without a form, without searching for the right folder.

Short interviews are suitable for established knowledge that is not tied to individual jobs. Half an hour to three quarters of an hour with an experienced employee, guided by a few key questions – which faults keep recurring on which types of system, which customers have which peculiarities, what would a successor need to know first – is recorded and transcribed. The result is not a novel, but a searchable foundation that simply did not exist before.

Handover documentation is the emergency version of the same idea: when a knowledge holder leaves the company or retires, the last weeks are used to dictate and capture systematically what would otherwise leave with them. If you only start during the notice period, you save only a fraction – which is why the dictation routine belongs in everyday work, not in the farewell.

This is how to introduce capturing knowledge from people’s heads in practice:

1. Name the knowledge holders. Who in the company is asked most often? These two to four people are the starting point – not the entire workforce.
2. Define the occasions. Dictation does not happen “at some point”, but at defined points: after every service job with something special about it, after every project completion, after every complaint that has been resolved.
3. Set guiding questions. Three fixed questions are enough: What was the problem? How was it solved? What do you need to know next time? This lowers the threshold and makes the notes comparable.
4. Support for two weeks. At the beginning, everyone forgets. A short reminder from the job manager in the first weeks decides whether the routine sticks.
5. Make the benefit visible. As soon as the first question in the team is answered via an old voice note, the mood tips from “extra work” to “obvious”.
6. Add established knowledge. Only once the routine is running do the short interviews with the most experienced employees follow – as planned appointments, not on the side.

What such routines look like in real companies and what they change in everyday work is shown with specific processes in the article AI knowledge base in practice.

How are network folders and file servers connected?

The network folder is almost always the most productive digital source, because documented know-how is already there: instructions, service records, project documentation, contracts, calculation bases. When connecting, you define which folders are to be taken over – for example “Projects”, “Instructions” and “Contracts”, but not “Private” or “Old_Backup_2011”. The knowledge base reads these folders, takes over the content and then monitors them continuously: new and changed files are recognised automatically, and common formats such as PDF, Word and Excel are processed directly. The system sets keywords and field values such as customer, project or date itself; what remains to be confirmed is the document type, because it determines who will later get to see the file.

Manual work also remains in the selection. Someone from the company has to decide which folders contain relevant knowledge and which only data junk – no software can make this decision, and it is worth it: if you take over everything unfiltered, you dilute the quality of the answers with outdated file versions.

Three stumbling blocks come up regularly. First, version chaos: files such as “Quote_final_v3_NEW_really.docx” exist in five versions, and without a prior decision all five end up in the stock. The rule that has proved its worth is to clarify briefly for each folder whether old versions come along or only the latest version. Second, orphaned folder structures: areas that nobody has maintained for years often contain outdated instructions that would resurface as supposedly valid answers – better to exclude them or mark them as “Archive”. Third, permissions: if the “Personnel” folder on the file server has restricted access, this restriction must be preserved in the knowledge base. This can be solved, but it must be explicitly taken into account during setup.

How do emails and attachments get into the knowledge base?

In many companies, the email mailbox is the secret main archive – and not just because of the attachments. The real treasure is the emails themselves: the supplier’s commitment, the changed agreement with the customer (“as discussed on the phone, we are extending the maintenance by two years”), the reason why a complaint was settled as a goodwill gesture. Such decisions are not written down anywhere else. If you only take over attachments, you leave behind exactly the knowledge that would later answer the question “What was actually agreed with the customer?”

Via a connection to Outlook or Microsoft 365, relevant emails and attachments can be taken over automatically, without anyone having to file every message individually. This is controlled by rules: for example “capture emails from the office@ mailbox including attachments”, “take over correspondence with these senders” or “only messages with attachments”. Duplicate detection ensures that the same file does not end up several times when it has been sent back and forth three times in an email thread.

Manual work remains in defining these rules – and in honestly refining them after the first few weeks, when it becomes clear that a rule captures too much noise (newsletters, automatic confirmations) or too little substance. The most important stumbling block, however, is not a technical one: personal mailboxes regularly contain private messages, job applications or sick notes that have no place in any knowledge base. In practice, shared mailboxes such as office@ or service@ are therefore connected completely, while personal mailboxes are not connected at all or only via narrow rules. What needs to be considered legally – from employee data to erasure obligations – is covered in detail in the article AI and the GDPR in Austria.

What happens to the paper archive?

Paper has to take the detour via the scanner. After that, text recognition – OCR in technical jargon, short for optical character recognition – does the actual work: it converts the scanned image into machine-readable text. Only this text makes a document searchable and quotable for the AI. This and other technical terms relating to connecting sources are explained concisely in the glossary.

Modern text recognition also copes with older originals: slightly yellowed pages, carbon copies, old PDF files that only exist as images. It reaches its limits with handwriting, very faint copies and damaged pages – there, recognition remains patchy, and important individual items are better checked. A maintenance logbook filled in by hand in the nineties will therefore not be captured perfectly; the question then is whether the system in question is even still in the installed base.

That is where the greatest manual work lies: not in scanning, but in reviewing beforehand. With paper, too, look through the know-how lens: old maintenance logbooks, system documentation and signed acceptance records for current properties are valuable – folders full of copies of records whose content has long since been captured in accounting are not. What has proved its worth is starting with the active and frequently used stock and only adding the rest of the archive when needed. A pragmatic test: has a folder been taken out of the cupboard even once in the last two years? If not, it goes to the end of the list. And if you want to dispose of old records while reviewing, you should first check the statutory retention obligations – the Austrian Economic Chamber WKO offers an overview.

How do smartphone photos help?

A lot of company knowledge is created where there is no scanner: on the building site, at the customer’s premises, in the warehouse. A photo of the rating plate, of the acceptance record filled in by hand or of the condition of a system before conversion is quickly taken. Via an upload in the browser, the photo goes straight from the smartphone into the knowledge base, where text recognition extracts the readable content and tagging assigns the image to the right customer or project.

Conversion and assignment run automatically; what remains with people is the photo itself and a quick look at whether the suggested document type fits – both so small that they realistically happen in everyday work. The typical stumbling block is the same as with voice notes: habit. In the first few weeks, only those who are reminded take photos. It helps to establish photo and dictation as a pair – after the job, two photos and a two-minute voice note, always in the same order.

How are existing systems such as SharePoint, DMS or ERP integrated?

If you already use SharePoint or a document management system (DMS, that is, software for structured document filing), you do not have to replace it. The knowledge base sees itself as a layer on top: it fetches the documents from the existing system and makes them queryable in plain language, while the source system remains leading. How DMS and AI knowledge base fundamentally differ – and why one does not replace the other – is examined in the comparison of document management and AI.

Technically, there are two routes. Either a direct interface through which documents are synchronised continuously – with SharePoint and Microsoft 365, this is the usual route. Or a regular export, in which the source system places documents in a transfer folder that is monitored like a network folder. Which route fits depends on the system in question and is a topic for the analysis phase; blanket promises would be dubious here. To be honest, it is also true that very old or exotic specialist programs with no export option sometimes cannot be connected sensibly at all – then the route via printed or saved reports remains.

With the ERP system (the central company software for orders, stock and accounting), a deliberate separation is recommended instead of a connection. Records, bookings, order and master data are in the right place in the ERP and should remain leading there – copying them into the knowledge base would not bring better answers, only noise. The rule of thumb: the ERP knows what was sold – the knowledge base knows how it is done. At most, accompanying documents with knowledge content that are only attached in the ERP can sensibly be taken over, such as technical documentation for an order. The question “What did customer Huber order last year?” is still answered by the ERP; the question “How did we solve the problem on Huber’s system?” is answered by the knowledge base.

In which order do you connect the sources?

The most common mistake in introduction projects is trying to connect all sources at the same time. The project becomes big, drags on, and the benefit only shows late. A staggered order according to two criteria makes more sense: how much usable know-how is in the source (knowledge density), and how quickly can it be connected (effort)? Sources with high knowledge density and low effort come first – this results in a prioritisation framework that has proved its worth in practice:

Stage	Source	Why at this point	Effort until benefit
1	Network folders (instructions, records, project documentation)	Highest knowledge density with the least effort – the team can ask its first questions straight away	Days
2 (parallel)	Voice notes / dictation routine	Captures knowledge from now on that would otherwise be lost; technology trivial, habit needs support	Weeks (habit)
3	Email mailboxes (shared addresses)	Agreements and decisions that exist only there; rules are quickly defined	Days to weeks
4	Paper archive, prioritised	Valuable, but reviewing and scanning take time – only active stock, the rest when needed	Weeks
5 (parallel)	Smartphone photos in everyday work	Complements the dictation routine; purely a matter of habit	Weeks (habit)
6	SharePoint / DMS	Needs coordination on the interface; worthwhile once the basis is in place	Weeks
–	ERP	Deliberately separate instead of connecting; at most take over accompanying documents	Largely not applicable

Two things stand out in this grid. First, the habit topics – dictations and photos – run in parallel with the technical connections, because they do not block any technology but need lead time: a routine that starts in week one already delivers a usable stock in week eight. Second, the ERP stands outside the staggering, because the aim is not to take it over at all.

This order has another advantage: with each stage you learn which questions your team actually asks. This sharpens the selection for the next source – it often turns out that part of the planned stock is not needed at all.

Which sources are not worthwhile?

Just as important as the question of what goes in is the question of what stays out. Every superfluous document dilutes the quality of the answers, lengthens the introduction and increases the amount of data that needs to be protected. Five candidates regularly fail in practice:

Record and accounting archives. Invoices, bank statements and accounting journals are recorded as leading in the ERP and with the tax adviser. The knowledge base would not gain better answers from them – if you want to know what was invoiced, you ask the ERP.

Version graveyards and old backups. Folders such as “Backup_old” or “Server_backup_2015” contain the same documents in outdated versions. Connecting them produces contradictory answers, because the AI cannot know that “Price_list_2016” no longer applies.

Personal mailboxes as a whole. Private messages, job applications and personnel matters do not belong in a shared knowledge stock – shared mailboxes and narrow rules are the right route here.

Raw data without text content. CAD drawing files, machine log data or measurement series contain hardly any usable content for language-based search. It makes more sense to take over the accompanying documents instead: the description of the drawing, the test report on the measurement series.

Dead archive material. Stock relating to customers, systems or products that have not mattered for many years costs review effort without any return. It can be added at any time if an old case unexpectedly becomes relevant again – usually that never happens.

The basic rule behind all five points: what gets connected is what answers future questions. What only exists because it once existed stays out.

Why is automatic tagging the underlying principle?

Classic filing systems only work as long as everyone sticks to the same folder logic – and that is exactly what fails in practice, because every employee files differently. Automatic tagging reverses the principle: instead of a person typing in categories, the software reads the content and derives the attributes from it – for example “customer Huber, system 3, topic maintenance, term until 2027”. The system fills in these field values itself. One piece of information deliberately remains with people: whether it is a maintenance contract, a service record or a quote is confirmed by someone in the review area, because the document type determines which group of people the document is released to. For unambiguous documents this is one click – but it cannot be skipped for any document. This applies to a scanned record just as much as to a transcribed voice note or an email that has been taken over.

This has two practical consequences. First, it hardly matters any more which source a document comes from or how untidily it was filed: after it has been taken over, everything is indexed uniformly – the fitter’s dictation stands on an equal footing with the contract from the network folder. Second, the search becomes independent of the file name: if you ask for the maintenance contract for system 3, you will find it even if the file is called “Scan_0047.pdf”. An AI knowledge base such as junova, Goma-it’s AI knowledge base, is built on exactly this principle: all sources flow into a shared stock that is tagged throughout, from which answers always come with a source – based on your own company documents.

How do you get started with connecting in practice?

At the beginning there is no technology, but taking stock with two questions: where are the documents your team searches for most often today? And who are the people who are asked most often? The first question determines the first technical source, the second the starting point of the dictation routine. Employees can usually answer both in half an hour. In a structured introduction process, this is exactly the purpose of the analysis phase before setup; what the steps from the initial conversation to training look like is shown in detail in the Process of an introduction.

A realistic expectation is important: connecting is not a one-off feat of strength, but a process over a few weeks in which automation carries most of the load and manual work is limited to a few clearly defined decisions – which folders, which rules, which occasions for dictating and, when capturing, the brief confirmation of the document type. Companies that start with one source and expand consistently soon have a searchable company memory that does not just find documents again, but preserves the company’s skills – without anyone having had to give up their familiar filing.