AI knowledge base readiness is whether the documents an AI answer tool will read have a list, an owner, a date and a version mark, with access only for the right people and text a tool can read. This checklist covers business documents such as policies, contracts, manuals and procedures, and it leaves a support-desk knowledge base out of scope. It holds 15 checks. Each lists what to bring and the point at which it passes, and then a score closes the sheet. The score shows whether records exist, not whether answers will be right. It sits under the Signisys knowledge search use case.
Last reviewed 2026-11-02 by Manish Sharma.
AI Knowledge Base Readiness: The Five Groups and Their Checks
This checklist serves a leader whose firm is about to point an answer tool at its own documents. The team scores itself, and nobody signs the sheet off as an audit. In Signisys’ working method, the checks sit in five groups, each built around a single question:
| Group | Question the group settles | Check numbers |
|---|---|---|
| Scope | Which documents are in the set? | 1 to 3 |
| Owners and dates | Who owns each document, and when was it last read? | 4 to 6 |
| Versions | Are old versions and duplicates marked? | 7 to 9 |
| Access | Who is allowed to open each document group? | 10 to 12 |
| Reading and review | Can the files be read, and is someone named to test? | 13 to 15 |
The groups and their pass tests are our own and follow no published standard. They score the documents only. In addition, this page does not explain why searches fail, how to build a search tool or which tool to buy. For why searches fail, see five causes, from question wording to access rules.
AI Knowledge Base Readiness: The Three Marks
The sheet gives every check one of three marks, and each mark has a short rule.
Ready (2): the pass test holds, with a dated record behind it. Almost (1): the test holds for part of the set, or the record has no date on it. Not yet (0): the pass test fails.
Without a dated record, a check stays below Ready, whatever the team believes. Use N/A only when a check does not apply to your set; remove its two points from the maximum and note why.
Pick a sample and write it down, such as the documents behind the last three staff questions, and do not assume the rest of the knowledge base matches it. If nothing is N/A, the top score is 30. The checklist sets no target or band.
For example, take check 5, the review date. If each group in the sample has a review date on file, the mark is Ready. If two groups have dates and a third has none, it is Almost. Finally, where nobody has recorded a date at all, it is Not yet.
Which Documents Are in the Set?
The first group settles which documents go into the knowledge base, and each later check works from that list.
| Check | Bring | Ready when |
|---|---|---|
| 1. The team has listed the document set | One line for each document group | Someone outside the team can find each group from the list |
| 2. The team has named where each group lives, such as a shared drive, an email folder, a chat tool or a paper file | The main location for each group | Each group has one main location |
| 3. The team has named the documents left out of the set | The left-out list | A written reason sits beside each one |
Check 3 matters as much as the first two. A document left out on purpose is a decision, while a document nobody thought of is a gap. The list then shows which is which.
Paper files also count here. If the answer to a staff question sits only in a cabinet, the list should say so, even though no tool will read it.
Key Takeaway Name what stays out of the knowledge base as well as what goes in.
Who Owns Each Document, and When Was It Last Read?
Because documents change over time, someone has to answer for each group. This group asks for names and dates written down.
| Check | Bring | Ready when |
|---|---|---|
| 4. Each document group has a named owner | One name for each group | When asked, the named person can say which version is current |
| 5. The team has recorded a review date for each group | The list of dates | Each date is on file |
| 6. The owner has written how old is too old for their documents | One sentence for each group | The sentence exists, and the team flags any document past it |
Check 6 gives no number; the right age differs between a holiday calendar and a supply contract. The owner knows the documents, so the owner writes the sentence.
An owner who has left the firm counts as no owner. In that case, check 4 stays at Not yet until someone else takes the group on.
Are Old Versions and Duplicates Marked?
Whatever sits in the knowledge base can end up in an answer. The tool can then quote an old, unmarked copy as if it were current.
| Check | Bring | Ready when |
|---|---|---|
| 7. The team has marked the current version in the place people open it | One marked example | A person who is not the owner picks the current version at first look |
| 8. The team has marked or removed old versions and duplicates | The list of old copies | Each copy on the list has a mark or is gone |
| 9. A short list pairs the firm’s internal names with the words staff use | The list of pairs | It covers the documents behind the last three staff questions |
Check 7 is a test of what a stranger sees. If only the owner can tell the current file from last year’s, the mark is Almost at best.
Check 9 deals with wording. Staff may ask about “maternity leave” while the policy title says “parental absence”, and a short list of pairs lets the owner see that gap. A short list is enough if it covers the questions people really ask.
Who Is Allowed to Open Each Document Group?
Access is the part of knowledge readiness that protects people as well as answers. If a confidential file sits in the set without a rule, its contents can surface in an answer to someone who should not see them. OWASP’s RAG Security Cheat Sheet lists the same failure: a user’s query retrieving a passage from a classified document that was stored without access rules on each passage.
| Check | Bring | Ready when |
|---|---|---|
| 10. The team has recorded an access rule for each group | The rule for each group | Each rule names who may read the group |
| 11. The team has named the confidential documents | The named list | The owner of each document has agreed to its place on the list |
| 12. A second person, not the owner, has checked one rule against what a named colleague can open | The dated note of that check | The note matches the rule |
Check 12 turns a rule on paper into something someone has seen. The person checking picks one colleague, looks at what that colleague can open, and then writes down what came up.
The wider policy on who may use AI tools, and for what, is a separate subject. It sits under governing enterprise AI use, so this checklist does not repeat it.
Key Takeaway Someone other than the owner should see an access rule hold for one real colleague.
Can the Files Be Read, and Is Someone Named to Test?
A scanned contract may look complete to a person and still give the tool nothing to read. The last group looks at that, and at whether the team has named anyone to watch the answers once they start.
| Check | Bring | Ready when |
|---|---|---|
| 13. The team has listed the files that are not readable as text | A list of scans, images and locked files | Each file on the list has a decision: read as text, replace or leave out |
| 14. A named person holds the list of test questions | The name and the list | Both are on file |
| 15. A dated record exists of the last review of questions the tool could not answer | The record | It has a date and a decision for each item |
Turning scans and images into text is a job of its own, covered under extracting data from documents.
Checks 14 and 15 test only that a name and a record exist. How to run the test or the review depends on the tool and the team, and this page does not set it out. For one way to run that test with real questions, see a six-step method for putting AI search over company documents.
AI Knowledge Base Readiness: Group Totals and What to Clear First
Total the marks in each group, then note the number of Not yet checks beside the total. In Signisys’ working method, you read the groups in order, with Scope and Owners and dates first. The later groups rest on those two. A version mark or an access rule means little while nobody knows which documents are in the set, or who owns them.
Clear the Not yet checks before you polish an Almost. Then date the sheet, name the sample and score again after each change. The total is for your own firm and sets no band, target or league table against other firms.
Put the scoring date and the sample on the sheet before you file it. A later score can then be set against this one.
A Score for Records, Not for Right Answers
This sheet checks records only. A full score shows that records exist, and it does not predict whether a tool’s answers will be right. It covers business documents, not a support-desk knowledge base, and it is neither an audit nor legal advice. After the score, the choice between a search tool and an assistant that writes replies is set out in a guide to the routes by which staff get answers from documents.
A second view helps in two cases: when the owners are unable to agree on the sample behind a group, or when the access group scores Not yet on checks 10 and 12.
Signisys provides data and AI services through its enterprise AI solution and its data and AI service page. The box below describes a Signisys service, not neutral advice, and it makes no promise. Bring your scored sheet and the sample behind it.
AI Knowledge Base Readiness: FAQ for Document Owners
What does AI knowledge base readiness cover for business documents?
Who should score AI knowledge base readiness in a mid-size firm?
Can you score AI knowledge base readiness without a tool?
What if some of our documents are scans?
When should we score again?
Security Insights, Delivered.
Practical, vendor-neutral analysis of threats, tools, and architecture decisions.