Data Readiness & Company Knowledge

Most AI projects stall on data, not on the model. The documents are scattered, out of date or locked behind permissions. This lesson shows how a strategist checks data readiness before money is spent, and how to fix the common problems.

Why data comes first

A model is only as useful as the information it can use. If your return policy lives in three different files and two are old, the AI will quote the wrong one. No clever prompt fixes that. Good data work is often the biggest part of a successful project.

What kinds of data matter

• Documents: policies, manuals, contracts, help articles and email templates.
• Records: customers, orders, tickets and products held in business systems.
• Conversations: chats, calls and emails, which show how work really happens.
• Examples: past cases with the right outcome, which become your test set.

The data readiness check

Run this check for each use case before approving a build. Score each line as ready, partly ready or not ready.

USE CASE: {{one sentence}}

FINDABLE     Do we know where the data lives, in which systems and folders?
OWNED        Is there a named person who owns each source and keeps it right?
CURRENT      Is it up to date? Is there a date or version on every document?
CLEAN        Few duplicates, few contradictions, readable text (not only scanned images)?
ALLOWED      Do privacy law, contracts or policy permit this use? Is consent in place?
PERMISSIONED Do we know who may see what, and can the system respect it?
REACHABLE    Can software read it through an API or export, without heavy manual work?
EXAMPLES     Do we have 30 to 100 past cases with the right answers for testing?

DECISION: go / fix first / stop       FIX PLAN AND OWNER: {{...}}

Common problems and fixes

• Scattered sources: make a simple inventory. List where each type of document lives and who owns it.
• Out-of-date content: set a review date for every document the AI will use. Remove retired ones.
• Contradictions: pick the source of truth and mark the others as replaced.
• Scans and images: use text extraction, and check the result by eye on a sample.
• Private data: remove or mask what the AI does not need. Keep personal data out of prompts and logs when you can.
• Permissions: carry each document's access rules into the search, so people only get answers from files they may read.

Who does the work

Data work needs owners in the business, not only engineers. Name a content owner for each knowledge area. Give them time. Make updating the source part of their job, so quality lasts after launch.

A small first step

You do not need to fix all data first. Pick one use case and prepare only the data it needs. Show the result. Then widen. This approach delivers value early and teaches you where the real problems are.

Privacy basics for a strategist

• Collect and use only what you need for a stated purpose.
• Know which data counts as personal data in your region, and which is more sensitive.
• Keep records of what data you use, why and for how long.
• Ask your privacy officer or lawyer early. Do not wait until launch.