Why AI Needs More Than Access to Your Files

A folder can contain a current supplier agreement, three drafts, a scanned invoice and a market report, all with names that mean little outside the team that saved them. Giving an assistant access makes those files available. It does not settle which one answers a particular question.

Several problems can get in the way:

• Unclear categories: two teams use different names for the same document type or business topic.
• Old versions and duplicates: an answer can draw on material that has been superseded if the retrieval system does not account for version status.
• Unreadable content: scans need text preparation before a text-based workflow can use them.
• Missing business context: a document may describe warehouse robots without using your investment team’s category name, “Warehouse automation.”
• Overbroad access: an assistant may make already-accessible information easier to discover, exposing weaknesses in existing sharing settings.

When an assistant makes oversharing easier to discover

Imagine a salary spreadsheet or an acquisition plan saved in a SharePoint location that everyone in the company can read. Few people may know it exists. An assistant such as Microsoft 365 Copilot can make that information easier to find or summarize for employees who already have access—even if the file owner never intended such a broad audience.

The risk is that overly broad permissions become much easier to use. Copilot does not need to bypass access controls for sensitive information to reach an unintended audience. A prompt telling an assistant to be careful is not a substitute for restricting access at the source.

Before connecting an assistant, check who can read sensitive locations, review broad sharing links, and apply the appropriate permissions and information-protection controls. Continue reviewing those settings as files and teams change.

StayCharted can help organize prepared records by business category. Those predicted categories do not secure a file or grant or revoke access. Your storage and governance systems remain responsible for enforcing who can see it.

Microsoft’s Copilot governance guidance specifically addresses oversharing and data hygiene. Category labels help organize content; access controls still need to be enforced by the systems that store and retrieve it.

Think of a table beside your files, with one row per file or record. It might live in SharePoint columns, a database or your existing document platform. You do not have to convert every document into a spreadsheet.

Start with the fields needed for a specific task. Dates, amounts, ownership and permissions should come from authoritative records or a suitable verified extraction process. A category prediction does not establish these facts.

Imagine a team asks: “Find company overviews about warehouse automation.”

Before consistent categories, relevant material sits among market reports, diligence notes and unrelated company presentations. Filenames and wording vary. Search may still find useful results, but the team has to check what each result actually is.

After classification and review, the catalog could look like this:

The search system can use Document type = Company overview and Business topic = Warehouse automation to narrow the candidate records, then retrieve their contents subject to the user’s permissions.

These are two separate classification tasks. Train a document-type model and a topic model if you need both columns. Neither requires StayCharted to host or search the original files.

Metadata filtering is an established retrieval technique: Azure AI Search supports filters on category and other nonvector fields. Your search integration must actually use the fields; adding a tag alone does not change an assistant’s behavior. Incorrect or overly narrow tags can also exclude relevant material, so check retrieval results as well as classification accuracy.

StayCharted AI Model Trainer learns your categories from text your team has already labeled. The useful input is a training table containing prepared text and the approved category—not a storage location on its own.

1. Choose one category field. Start with document type, business topic or diligence topic. Write down what each category includes and where its boundaries lie.
2. Collect reviewed examples. Supply representative text and labels. If production inputs will be summaries, use comparable summaries for training. Keep related versions or near-duplicates from leaking across your training and evaluation data.
3. Train and inspect the results. Look at held-out accuracy, individual mistakes and performance by category. A strong overall score can hide a weak category.
4. Categorize new records. Fill a spreadsheet in StayCharted, or send prepared text through its API. API access is available on Essentials and Business.
5. Review and save the answers. Use confidence to help choose records for human review. Your integration writes the category and review status back to the catalog.
6. Improve the next version. Add approved corrections to your examples, retrain, evaluate and publish when the new version is ready. Corrections do not silently alter the model already in use.

Confidence helps prioritize review; it does not identify every mistake. Choose a cutoff using your evaluation and the consequences of a wrong category. Explore model options and current plan details.

StayCharted provides the classification model. Your existing systems prepare text, keep files and versions, enforce permissions, maintain the search index and retrieve the original content.

The model can suggest that a record is a Supplier agreement or that a note concerns Competition. It should not be used to decide that an agreement is legally signed, that a file is the authoritative current version, or that a person should receive access.

The category catalog and the training dataset serve different purposes: the catalog describes records for everyday use; the training dataset teaches the model from approved examples. New catalog labels should not automatically become training truth without review.

Pick a recurring task, such as separating supplier agreements, purchase orders and insurance certificates. Test one model on representative records before expanding to more categories or departments.

For an initial batch, export prepared text with stable record IDs, fill the category column in StayCharted, review the results and import the answers back. For new records arriving over time, connect your system to the API and respect its rate limits.

If your team uses Microsoft, the Power Automate and SharePoint guide describes a flow that sends prepared text to StayCharted and writes back category and confidence. It is a documented integration pattern awaiting an end-to-end Microsoft tenant test.

Better categories give people and search systems another way to find relevant information. They complement content preparation and governance; they do not guarantee the correctness of an assistant’s final answer.