In this guide 6 sections
Your documents are ready when each policy has one current version, every file has an owner, scans carry readable text and the rules people follow are written down. If two or more fail, tidy first. Otherwise it quotes the wrong page, confidently.
Does each policy exist in one current version?
Search your shared drive for the leave policy. You may find Leave_Policy_final.docx, Leave_Policy_final_v2.docx and a PDF from two years ago, and an assistant will find all three. It cannot tell which one HR signed off.
Retrieval tools read whatever they are given and do not judge, so duplicates confuse them and the answer depends on which copy ranks highest in the search.
Pick the policies staff ask about most. Refunds, warranty claims, expense limits and shift swaps lead, so give each one a single current file and move drafts into an archive folder that the assistant never indexes.
- One current file per policy, with an issue date on page one
- Superseded versions moved out of the indexed folder
- A change note saying what moved between versions
- Filenames that match the title staff use in conversation
Where do your real rules live?
Rules often live outside documents. The discount ceiling for a loyal dealer sits in a WhatsApp group, and the returns exception for damaged packaging lives in one supervisor’s memory.
An assistant built on the official handbook will answer from the handbook, and staff will ignore it because they already know the handbook is out of date.
Ask three people in different roles which questions they answer repeatedly every week, such as a dispatcher, a service engineer and an accounts clerk. Write those answers down before anything is indexed. A short note per question will do.
Can a machine read your files at all?
A scanned supplier contract is often a photograph inside a PDF, so select the text with your cursor and see whether anything highlights. No highlight means no text layer. A search tool sees a blank page.
Text recognition can add that layer. Handwritten margin notes, stamps and skewed pages still come out wrong, and tables suffer most because a price grid can flatten into a single line of numbers.
Sample twenty files across your formats: Word, Excel, PDF scans, email exports and slide decks, then count how many return clean text and which come back as garbage. That count tells you the real cleanup effort.
- Image-only PDFs that need text recognition
- Spreadsheets where meaning sits in colour or cell comments
- Password-protected archives and zipped folders
- Email chains saved as screenshots
Who owns each file?
A knowledge base decays the day it launches. Someone must answer for each page. Ownership is the dull part of this work, yet it decides whether staff still trust the assistant a year on.
Add three fields to your document register, namely owner, last reviewed and review due, so that nothing goes stale without someone noticing. A quality manager owns the inspection procedure. A finance lead owns the expense policy. Leavers hand their files over within the week.
Ownerless files are a finding. Archive them, and do not index them.
Fix first, or build now?
Build now if most of your core documents are single-version, readable and owned, and start with a pilot over one department, such as support or the service desk. It will expose the remaining gaps quickly.
Fix first if you counted duplicates in more than a few policies, or found that key rules exist only in chat. Building on that base produces confident wrong answers. Confident wrong answers are worse than none.
Score your top thirty documents this week. Bring the sheet to the first planning call, where the gaps it shows will set the order of cleanup before any build starts.
Document readiness scorecard
Up to thirty rows fill one sheet, one per document, with columns for title, folder, live versions, text layer, named owner and last reviewed date. The last column marks each row fix first or build now, and a summary box at the foot counts the marks and states the verdict.
Open a blank worksheet to printQuestions owners ask
How many documents do we need before an AI knowledge base makes sense?
Quantity matters less than repeat questions. A hundred well kept files that staff ask about daily will serve you better than ten thousand stale ones, so start where the questions pile up.
Do we have to digitise every old paper record first?
No. Begin with the documents people consult weekly. Archive boxes of old job cards, paper registers and filed delivery notes can wait until the assistant has proved itself on current material.
Can the assistant keep confidential files away from some staff?
It should. Access rules from your drive or HR system can carry over, so a technician sees service manuals but not salary letters. Confirm this before any pilot opens.
What if our documents are in more than one language?
Mixed languages work, with testing. Run sample questions in every language your staff use, then check each answer against the source page and note any that drift into the wrong language.