Prepare the information the employee's task actually needs. General drafting may need only a clear brief. Comparing files needs usable, permitted documents. Recurring company search needs a maintained source collection. Updating a business system also needs valid records and fields, plus authority to make the change. Choose the information route before commissioning a company-wide data cleanup.
For an enablement lead and departmental owner, the immediate decision is what must be prepared for a bounded trial. Some work will uncover wider problems worth fixing. That does not make every data-platform improvement a prerequisite for every employee AI task.
Choose the information route
Start with the accepted output and ask where its facts will come from. “Use AI in customer service” is too broad to answer. “Compare the dimensions of two desks using their current specification sheets” makes the preparation visible.
These routes can coexist within one team. They are different information needs, not stages every organization must climb.
| What the employee wants | Preparation to examine first | What would expand the work |
|---|---|---|
Draft general wording | Purpose, audience, tone and facts the employee supplies | The draft must include company-specific promises or customer details |
Analyse selected files | Applicable versions, permitted use and readable content | Important facts sit in diagrams, scans or linked attachments |
Find answers repeatedly across company sources | Defined source scope, access, content structure and an update route | More departments, conflicting sources or frequent changes |
Create or update a business record | Correct record identifiers, required fields, validation and action authority | Sending, retries or changes that affect other records |
Suppose Sam Taylor works in customer service at a furniture distributor. Asking for clearer wording for a generic delivery update does not require the whole order database. Adding a promised arrival date does require a reliable source for that order. Creating a replacement-parts request introduces another dependency: identifying the correct order line and part.
Keep the broader operating decision in the tools-and-workflows guide. Here, use the chosen task to decide which information needs work first.
Make the selected files usable
For Sam's desk comparison, prepare the current sheets for the relevant product variants. Keep model names, dimensions, units and qualifications together. A width without its unit, or a dimension drawing separated from the variant it describes, leaves the assistant to guess at a relationship the source should settle.
Inspect what the selected tool actually receives:
- Text and structure. Check that headings, table columns, footnotes and reading order survive processing.
- Visual information. Try a page whose answer depends on a diagram or image, rather than testing only a plain paragraph.
- Completeness. Include a referenced attachment when it is necessary and permitted. Leave unrelated material out.
- Applicability. Identify the version and product variant being compared. Resolve uncertain source status with its owner.
Do this through the interface employees will use. A file opening successfully does not establish that its tables or diagrams reached the answer.
OpenAI's PDF visual-retrieval FAQ, checked in October 2026, illustrates why the route matters. It describes visual retrieval for PDFs supplied in ChatGPT Enterprise conversations, while PDFs stored as GPT Knowledge or Project Files use text-only retrieval. That is a specific product distinction, not a rule for every account or assistant.
The need can be very concrete. In a May 2025 public discussion, a Reddit user who described themselves as a security technician wanted on-site access to manuals, wiring diagrams, parts lists and work notes. Replies proposed tools, but did not establish a successful deployment. Their question shows what a text-only test could miss: finding a manual is different from making its diagrams available for inspection.
Keep the original available to the reviewer. If Sam cannot verify which measurement belongs to which desk, a polished comparison is still unfinished work. Use the human-review guide to define the checks that make the result usable.
Prepare a maintained collection when the task needs one
Repeated questions across a catalogue create a different commitment from supplying two files once. Define the collection needed for the chosen questions, with responsibility for three things:
- Scope. Which catalogue, variants and applicable versions are included? Keep superseded material out of the ordinary answer route.
- Access. Which employees may retrieve the material? Ask the technical owner to demonstrate the restriction through an ordinary account and correct inappropriate existing permissions.
- Updates. Who maintains the source and checks that changes reach the assistant? An uploaded copy may remain unchanged after its original is revised.
Microsoft's Copilot setup guidance includes pilot testing and SharePoint access and oversharing reviews. These are substantive deployment tasks. A small trial is not permission to skip the controls required for its sources.
If the system breaks documents into passages for retrieval, check that a passage keeps the context needed to interpret it. Microsoft's Azure guidance on retrieval-augmented generation (RAG) explains the tradeoff between oversized passages and pieces too small to preserve meaning. It recommends testing approaches against the document types involved. The tool may manage this processing; employees do not automatically need to build their own retrieval pipeline.
For Sam's catalogue, a useful test asks about two similarly named desk variants and checks that each answer retains the right dimensions. Expand the collection when the work requires it and the team can maintain it. Do not use document count alone as evidence that the arrangement is ready.
Treat system actions as record preparation
A product sheet can explain a replacement part. It cannot, by itself, establish which customer's order line should receive that part or whether a request has already been created.
Before allowing an AI-assisted route to create Sam's replacement request, the business and technical owners need to settle:
- Record identity. Which order line and part identifier will be used, and how will a mismatch be caught?
- Required values. Which fields must be present, and which must come from an authoritative system rather than generated text?
- Action boundary. Is the tool drafting a request, saving one or sending it onward? Who accepts that step?
- Uncertain results. What happens if a field is missing or a save appears to time out? Check the system before repeating an uncertain action.
Power Automate's action documentation makes the distinction tangible: actions can update records or send emails, and their configuration has required fields and options. Preparing documents is only part of preparing such a workflow.
If the order line cannot be identified, keep the request unresolved for a person to check. Do not make a demonstration work by inventing an identifier, adding unnecessary access or treating a draft as an accepted record.
Test a small prepared packet before expanding
Assemble enough approved material to exercise the task and its important exceptions. There is no useful universal file count. The packet should reveal whether preparation preserves the facts and relationships that matter.
- Agree the reference result. For the desk comparison, establish the applicable models, dimensions and units with the product owner.
- Try the ordinary route. Use the actual employee account and interface, including the file-selection or search step.
- Include a revealing exception. Try a similar variant, a missing dimension or an answer carried only by a drawing. Check that uncertainty remains visible.
- Inspect the next handoff. Confirm that the employee can verify the answer and, where actions are in scope, that the correct record would be affected.
Record what needs changing before the next attempt: a missing attachment, a damaged table, an outdated copy, a source-access or retrieval issue or an ambiguous order field. Assign the repair to someone able to make it. Include preparation and checking effort when deciding whether the task remains worthwhile.
You have enough preparation for the next bounded trial when its required information is permitted, usable and checkable through the chosen route, and unresolved cases have a responsible destination. That is a trial decision, not proof that every company source is ready. Use the acceptable-use policy guide for the employee-facing boundaries around accounts, inputs and review.
Questions about preparing data for employee AI tools
Must we clean all company data before using AI?
The preparation should follow the task and its required controls. A general wording exercise may need no company records; recurring internal search can require substantial source and access work. Identify the necessary information, then test whether it is usable and permitted. The information-route comparison helps separate these commitments before expanding the scope.
Do scanned PDFs always need OCR first?
Check how the selected tool processes the PDF and how the file is supplied. Some routes can interpret visuals; others rely on extracted text. Optical character recognition can help make scanned text available, but it does not by itself preserve a diagram's meaning or a table's relationships. Test a representative page and compare the result with the original, as described in the file checks.
Is uploading company documents the same as training a model?
Supplying documents for a task or retrieval is different from training a model. In Microsoft's description of RAG, retrieved material grounds the response. Do not assume that employee use requires a training dataset or model-development project. Verify the selected service's handling of uploaded content separately with the responsible owner. Then prepare the sources for the actual task and retrieval route.
How many documents should a pilot include?
Include enough approved material to cover the chosen task, relevant variants and important exceptions. A large file count is not evidence of completeness, and a small count is not automatically sufficient. Start with questions whose answers a subject owner can verify, then build a representative test packet around them.
What extra preparation is needed when AI updates a system?
An update needs the correct target record, valid required fields, enforced action authority and a way to handle missing information or uncertain results. A readable source document does not establish those conditions. Have the workflow owner and technical owner demonstrate the record and action checks before allowing changes to ordinary business records.
