AI data preparation: check the information before you connect it to a workflow

Match the preparation to the task

For model training, keep final evaluation examples separate from the examples used to develop the model. Google’s dataset guidance explains training, validation and test sets and why duplicate examples can distort evaluation.

The rest of this guide focuses on preparing business information for a workflow. It does not replace the statistical and engineering work needed to train a predictive model.

Inventory the sources the job actually needs

List the records required for one question or result. For a report, that might mean the client agreement, metric definitions, approved figures and recent changes. An entire shared drive may contain a great deal of material with no bearing on that report.

Assign a source owner who can confirm meaning and approve corrections. Record the original link, client identity, access rules and when the information was collected. Keep the date a fact became valid separate from the date you imported it.

Said describes our current build as bringing context together from sources including SharePoint, email and Slack at periodic times. That establishes reported source coverage, not proof that every record is complete, current or suitable for every question.

Correct inconsistencies without losing their meaning

• Use stable identifiers for clients, people and records.
• Confirm metric definitions and units before joining values.
• Identify old versions and preserve the rule for which version applies.
• Record missing information explicitly rather than replacing it with a plausible value.
• Keep evidence of important corrections and the person who approved them.

For documents, check what is lost during extraction. A table heading, footnote or section title can change the meaning of a paragraph. If you divide a document into smaller passages for retrieval, keep enough surrounding information and a link to the original for review.

For reporting, distinguish period dates from collection dates. Check duplicate exports, revised figures and partial periods. A clean format cannot make mismatched periods comparable.

Preserve access and uncertainty through the workflow

Check permissions before retrieval and again at the points where output is shared. A summary must respect the same client boundary as its source. Test with records from different clients and a user who should see only one.

Our current build carries provenance so old, missing or contradictory information is visible to the agent and its reviewer. Said describes confidence declining when a record is not seen again by its validity date. When the evidence cannot support a reviewable answer, the system must say it does not know.

Our delivered analytics work used coded names and transformed numbers before AI. That is a specific handling choice, documented in the case study. Any proposed transformation needs its own assessment of what information remains identifiable or useful.

Test the prepared information on known questions

Write expected answers and their authoritative sources before testing. Include an older agreement, missing input, conflicting decision and wrong-client record. Check what the workflow retrieves, what it omits and what it should hold for a person.

Correct the cause of a failed test in the source or preparation rule. Preserve a test set for later changes. The knowledge-base example demonstrates several of these boundaries with synthetic records.