Preparing data before an AI pilot

Prepare usable sources and clear ownership

List the information the pilot requires and identify who can authorize its use. Check accuracy, duplication, access rights, format and update frequency. A collection of documents is not automatically a usable knowledge base. Conflicting policies and obsolete versions can create failures that appear to be model problems but originate in the source material.

Worked example

Illustrative scenario; not a claim of measured client results.

For a support assistant, select current approved procedures and retain their effective dates. Exclude superseded drafts unless a historical use case requires them. Test whether tables, scanned pages and attachments can be extracted correctly. Give each document family an owner who can resolve contradictions. A smaller reliable collection is a better starting point than an indiscriminate upload of every shared file.

Checks, limitations and next decisions

If the assistant answers inconsistently, inspect the exact sources available for the question. Check whether the needed material exists and whether the user is authorized to access it. Record source gaps as content work rather than trying to hide them with prompt changes. Define how additions, revisions and deletions will reach the pilot before relying on it for changing information.

A practical starting point

Work through these checks in order. Use the results to decide what needs to change in your own workflow.

1. Identify the sources needed for the task and confirm authorized access.
2. Remove obsolete duplicates and label authoritative versions.
3. Create a small representative test set before building the interface.

A mistake to avoid

Do not upload sensitive material to an unapproved service because it is convenient for testing.

How to check progress

Measure source coverage and the number of unresolved conflicts before evaluating answers.