Analysis

AI document management: how to tidy a company archive and get answers that show their source

published · Author: Dariusz Wtykło

AI document management tidies a company archive in three steps: it reads scans, describes and sorts documents, and then answers questions about them in plain language. Analysing written text is now the most common use of AI among EU enterprises, at 11.8%. One condition always applies: every answer must show its source document, because even professional tools built on their own document base gave incorrect information in 17% to 33% of answers in one study.24

Almost every company knows this scene. Someone asks for a supplier contract from a few years back and the search begins: in binders, on the shared drive, in mailboxes and in the memory of the one person who “always knew where it was”. The documents exist, but nobody can work with them. That is not a technology problem; it is a tidiness problem — and tidying is exactly where AI can help today.

Before a company buys an AI assistant for documents, it should be able to answer three questions: what do we keep, for how long, and who has access? AI sorts and finds things well, but it will not decide for the owner what can be thrown away. That is why we start with one cabinet or one drive, not the whole company.

— Dariusz Wtykło, owner and CEO of Primessio

How much company knowledge still sits on paper

Paper has not left the office. In a Bitkom survey of 1,103 German companies, 38% ran about half of their office and administrative processes on paper, and only 15% worked fully paperless1. In one company in five paper still dominated: 14% ran about three quarters of their processes on paper and 6% nearly all of them1.

At the same time, companies increasingly ask AI to do precisely this: read text. According to Eurostat, 20.0% of EU enterprises with 10 or more employees used AI in 2025, most commonly to analyse written language — 11.8%2. For a managing director the message is simple: putting documents in order with AI is not an experiment, it is a common, proven use.

What AI can do with an archive

Here are five things AI does with documents, from simple to more demanding. Each can be checked on a small sample before the company commits to anything bigger.

Reading scans

To a computer, a scanned contract is a picture, not text. Text recognition (OCR) turns that picture into text you can search. Newer AI tools also cope with tables, stamps and poorer scan quality. Handwriting and faded printouts still need a manual check.

AI document classification: naming and sorting

AI reads a document and proposes what it is: contract, invoice, minutes, official letter. It adds basic details — who it concerns, which period it covers, which matter it belongs to. Hundreds of files named something like “scan_final_new” become a list you can filter.

Pulling out the key facts

From a contract: parties, signing date, term, notice period. From an invoice: counterparty, amount, due date. These facts go into a table, so the company can see, for instance, which contracts end next quarter — without opening each one.

Semantic search: by meaning, not by word

A classic search finds a file only if you type the exact word that appears in it. Semantic search looks for meaning: asked about a “penalty for late delivery”, it also finds the clause on “damages for delay”. This is usually the first change employees notice in daily work.

An AI knowledge base: answering questions from company documents

An AI knowledge base works like this: the assistant first retrieves the passages that match the question and only then writes an answer based on them. The method was described in a 2020 research paper as retrieval-augmented generation, or RAG5. A well-built assistant shows, next to each answer, which document and page it came from.

New employees gain most from organised knowledge. In a study of 5,179 customer support agents, an AI assistant that suggested responses raised the number of issues resolved per hour by 14% on average and by 34% for novices6. The authors’ explanation: the tool passes the working practices of experienced colleagues on to newer ones6. That is the result from one company, not a promise for every company — but it shows what organising knowledge is for.

Where AI gets it wrong

  • Made-up answers. An assistant built on its own document base is wrong less often than a general chatbot, but it is still wrong sometimes. In a study of professional research tools for lawyers that work this way, 17% to 33% of answers contained incorrect or made-up information4. Treat an answer without a source as a rumour.
  • Poor scans. A crooked page, handwriting or a faded printout leads to reading errors. A person always checks amounts, dates and account numbers taken from such documents.
  • Old versions. AI does not know which version of a contract is in force unless the company has marked it. It will answer correctly — from the wrong file.
  • Access that is too broad. An assistant that “sees everything” will show the payroll to anyone who asks. Document permissions must apply to the assistant too.
  • Data leaving the company. Before uploading documents to a tool, check where they are processed and whether the provider uses them to train its models.

Data retention: what AI does not decide

AI does not decide what may be deleted. The law does, and the periods differ from country to country. Two examples from the EU:

DocumentGermanyPoland
Accounting books10 years9at least 5 years7
Accounting vouchers8 years9at least 5 years7
Commercial letters6 years9depends on the document type
Personnel files (hired from 2019)under separate rulesthe employment period plus 10 years from the end of the year in which employment ended8

Tax and sector rules can extend these periods, and other countries have their own. Confirm your list with an accountant or lawyer. This article is not legal advice.

The other side is personal data. The GDPR requires that it is not kept longer than necessary. In practice a company should have written retention periods for each purpose and a deletion procedure10. Here AI is a helper: it flags documents past their date and prepares a list for decision. A person approves the deletion.

Example: the kind of order AI can work on

In our system for the law firm Halberg & Wierzba (concept project · fictional data), every document belongs to a matter, has a version history and a retention date, and a checklist shows what is still missing from the matter. That order does not itself need AI. It is, however, the condition for an AI assistant to answer from the right version of the right document. See the description: law firm management software.

Where to start: a plan for the first week

A company can take these steps on its own, without buying anything:

  • Pick one area. One cabinet, one drive or one document type, such as supplier contracts. Not the whole company.
  • Count the problem. For a week, note how often someone looks for a document and how long it takes. That is your baseline.
  • Write the rules on one page. Which documents live here, who has access and how long they must be kept.
  • Test on a few dozen documents. Scan them, let an AI tool describe them and check by hand how many descriptions are right. Keep sensitive data out of the test until you know where the tool processes data.
  • Ask a dozen real questions. The kind that come up in the company every week. Check whether the answer is right and whether it shows the source document.

After a week you have numbers from your own company: how much time searching takes and how often AI gets it right. That is the basis for deciding whether to go further. The next steps — from process map to maintenance — are set out in the Primessio AI Implementation Framework, a method developed by Primessio in project practice. If you would rather take that path with a team that does it alongside you, see our AI automation and AI implementation services.

Frequently asked questions

Can AI organise a company archive on its own?

Not on its own. AI reads scans, suggests names, categories and descriptions, and helps people search. Deciding what matters, who has access and what may be deleted stays with the company. Without those rules, AI only moves the mess from the cabinet to the drive faster.

What is AI document classification?

AI reads a document and proposes what it is — a contract, an invoice, minutes, an official letter — and adds basic details such as who it concerns and which period it covers. A person checks a sample. The result is a list you can filter instead of a folder of files with meaningless names.

What is semantic search?

Search by meaning rather than by exact word. A classic search finds a file only if you type the word that appears in it. Semantic search, asked about a “penalty for late delivery”, also finds the contract that talks about “damages for delay”.

Does AI replace a data retention policy?

No. A retention policy says how long each type of document is kept and who approves deletion; the law sets the minimum periods, and they differ by country. In Germany, for example, accounting books are kept for 10 years and commercial letters for 6 years. AI helps apply the policy by flagging documents past their date. Confirm your own list with an accountant or lawyer.9

Can an AI knowledge base give wrong answers?

Yes. An assistant that answers from its own document base is wrong less often than a general chatbot, but it is still wrong sometimes. In a study of professional research tools for lawyers, 17% to 33% of answers contained incorrect or made-up information. The assistant should show the source document with every answer, and a person should check important decisions.4

Sources

  1. Bitkom: 4 von 10 Unternehmen arbeiten papierlos (Digital Office Index 2024) · · tier 2 · checked
  2. Eurostat: 20% of EU enterprises use AI technologies · · tier 1 · checked
  3. Główny Urząd Statystyczny: AI, zakupy i sprawy urzędowe – tak Polacy korzystają z internetu w 2026 roku · · tier 1 · checked
  4. Stanford University / arXiv: Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Magesh, Surani, Dahl, Suzgun, Manning, Ho; arXiv:2405.20362) · · preprint, not peer-reviewed · tier 2 · checked
  5. NeurIPS / arXiv: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., NeurIPS 2020; arXiv:2005.11401) · · tier 1 · checked
  6. National Bureau of Economic Research: Generative AI at Work (Brynjolfsson, Li, Raymond; NBER Working Paper 31161) · · preprint, not peer-reviewed · tier 2 · checked
  7. Sejm RP (Internetowy System Aktów Prawnych): Ustawa z dnia 29 września 1994 r. o rachunkowości (tekst jednolity, Dz. U. 2023 poz. 120), art. 74 · · tier 1 · checked
  8. Ministerstwo Rodziny, Pracy i Polityki Społecznej: Dokumentacja pracownicza · · tier 1 · checked
  9. Bundesministerium der Justiz (gesetze-im-internet.de): Abgabenordnung (AO) § 147 Ordnungsvorschriften für die Aufbewahrung von Unterlagen · · tier 1 · checked
  10. European Data Protection Board: Data protection guide for small business — Data protection basics · · tier 1 · checked

Related