AI projects succeed or fail on data more than on models. Here is a practical checklist for getting your documents and databases ready for AI assistants and agents.
Data matters more than the model
When an AI assistant gives a wrong answer, the cause is often the content it was given rather than the model itself: outdated documents, contradictory policies, or data it shouldn't have seen. Preparing your data is the least glamorous part of an AI project, and the one that most affects the result.
Step 1: Pick one use case
Start from a specific job, such as answering staff policy questions, drafting replies to customer emails, or extracting data from invoices. The use case tells you which data matters. "Prepare all our data for AI" is not a project anyone can finish.
Step 2: Inventory the sources
For that use case, list where the information lives: shared drives, help centres, wikis, PDFs, email templates, CRM records, databases. For each, note:
- Who owns it and keeps it current.
- How it's accessed: export, API or manual.
- Who is allowed to see it.
Step 3: Clean up documents
- Remove or archive outdated versions. If two documents disagree, the AI will too.
- Fix the structure. Clear headings, one topic per section, and tables that are real tables rather than images.
- Convert scans to searchable text, and check the accuracy of anything converted automatically.
- Add context that people know but documents don't state, such as which regions or products a policy applies to.
Step 4: Get structured data in shape
For agents that work with records such as orders, customers or bookings:
- Consistent identifiers across systems.
- Documented fields: what "status = 3" actually means.
- An API or a reliable export, rather than screen scraping.
- Test data that resembles real data without exposing real customers.
Step 5: Decide on access and privacy
- Apply the same permissions to AI answers as to the source documents.
- Keep personal and sensitive data out unless the use case truly needs it.
- Check how your AI provider handles data, including retention and whether it may be used for training.
- Consider hosting models in your own cloud for highly sensitive material.
Step 6: Build an evaluation set
Collect real questions or tasks with known good answers. Use them to test the system before launch and every time you change content, prompts or models.
Conclusion
Good AI starts with tidy, current, well-permissioned data, scoped to one clear use case. Do that groundwork first and the model choice becomes much less critical. Our AI development team can help you assess your data and choose a first project.