Almost every company has its knowledge scattered around. SOPs sit in Google Drive, the price list lives in a sales spreadsheet, the leave policy is a PDF someone shared in a WhatsApp group two years ago, and the rest is in the heads of long-serving staff. When a new hire asks a question, the answer depends on who happens to be free.
Many companies now want AI to handle these questions. The technology is available, but the results depend heavily on the material it is given. A well-organised internal knowledge base is the main requirement for AI to answer correctly. This article covers how to organise documents, assign owners, keep content current and control access per team.
How AI answers from an internal knowledge base
Question-and-answer systems over company documents usually use an approach called RAG (retrieval-augmented generation). Before answering, the system searches for the parts of your documents most relevant to the question. The AI then writes an answer from those parts and cites the source, so the user can open the original document.
This means the AI is only as good as the documents it finds. If there are two versions of the returns SOP with different rules, it may pick the old one. If the overtime rule was only ever announced verbally, it will not know about it. Most of the work in building a knowledge base happens in the documents, well before the technology comes in.
Organising SOPs and documents
Start with the questions that come up most often. Trying to gather every document at once usually stalls the project halfway. Ask each team lead to list the ten questions their staff asked most in the past month. At a clinic, these might be the registration procedure for BPJS patients, the instrument sterilisation schedule or how to handle complaints about queues. The list shows which documents need tidying first.
A few habits make documents easier to read for both AI and people:
- One document, one topic. The shipping SOP and the returns SOP should be separate.
- Use clear headings and subheadings, since the system usually splits documents along their structure.
- Put the effective date and the owner's name at the top.
- Archive old versions. Do not leave files called "SOP final", "SOP final revised" and "SOP latest" in the same folder.
- For scanned PDFs, make sure the text can be read through OCR, including tables.
- Write down rules that have so far only been passed on verbally or in chat.
Every document needs an owner
A knowledge base without owners goes stale within months. Assign one person or one role to each group of documents. HR owns the people policies, the head of sales owns the price list, and the warehouse lead owns warehouse procedures.
The owner's job is not heavy, but it has to be clear. They make sure the content matches how work is actually done, update it when things change, and follow up when someone reports that the AI gave a wrong answer from their document. Showing the owner's name on the document also tells staff who to ask when the AI's answer is not enough.
For companies with several branches, such as a distributor with warehouses in a few cities, you also need to mark which documents apply everywhere and which apply to one location. Without that, the AI may answer a question from the Surabaya team with the Medan warehouse rules.
Keeping content current
The most common problem with a knowledge base is that it falls behind reality. Prices have gone up and the approval flow has changed, but the document is still the old version. With AI this is riskier, because a wrong answer comes back sounding confident.
A few practices help:
- Schedule regular reviews, for example every three months for SOPs and every month for price lists.
- Tie document updates to policy changes. Every new memo or announcement should be followed by a revision of the related document.
- Make sure the AI system re-reads documents after they are updated, so the search index does not keep old versions.
- Give people an easy way to report wrong answers, and route those reports to the document owner.
- Log questions the system could not answer. That list shows you which documents do not exist yet.
Access per team
Not all knowledge is meant for everyone. Payroll data, supplier contracts and performance reviews belong to specific teams. When AI is added, these access rules have to come along.
The principle is that the AI may only search documents the person asking is allowed to read. This restriction has to be applied at the search stage. Hiding the answer in the interface is not enough, because the document content has already been sent to the model. The most practical approach is to follow the access structure you already have, such as folders per division or user groups per team.
Pay attention to personal data as well. HR and customer documents often contain NIK (national ID numbers), NPWP (tax IDs) or phone numbers. Data like this should be masked before the AI processes it, in line with your company's privacy policy and the principles of Indonesia's Personal Data Protection Law (UU PDP).
Start with one team
Pick one team with clear, recurring questions, such as customer service or HR, and run a trial for a few weeks. Check how often the AI answers correctly and cites the right source, then fix the documents before expanding to other teams. Averoa helps companies build search with citations and per-team access through its Company AI service.
