Files and spreadsheets often contain the real operating layer of a business: pipeline trackers, portfolio updates, budgets, action registers, research, and meeting packs. Connecting them can remove manual searching and reconciliation, but an unrestricted connector can also amplify inconsistent data and broad permissions.

Separate storage, index, reasoning, and action

Approved source → Permission-aware retrieval → Task context → AI processing → Validation → Controlled write-back

Each layer has a different responsibility. Combining them into one opaque automation makes errors and leakage difficult to detect.

1. Define the approved source

Select folders, sites, workspaces, or tables—not an entire cloud account. Assign business owners, define authoritative versions, remove duplicates, and document refresh frequency.

2. Preserve source permissions

The retrieval layer should respect user identity and source-level access. A relevant document is not automatically an authorised document. Use service identities with the minimum scope and separate environments and credentials.

3. Build usable context

For documents, retain file name, owner, version, timestamp, section, and link with each indexed passage. For spreadsheets, define the table, field meaning, data type, unique key, permitted filters, and quality rules. Do not ask a model to infer an undocumented schema.

4. Validate before write-back

Check required fields, data types, duplicates, totals, permissions, and conflicts. High-impact updates should become proposed changes for human approval. Preserve the previous value and action log.

Example: portfolio operating review

  1. A scheduled workflow reads approved portfolio KPI tables and current board packs.
  2. Validation flags missing periods, broken totals, and stale submissions.
  3. The AI layer drafts variance explanations and questions linked to sources.
  4. The portfolio owner reviews and approves the briefing.
  5. Approved actions are written to the action register with owner and due date.

The model accelerates preparation; it does not silently redefine the KPI, approve the explanation, or overwrite the underlying record.

Reliability and cost controls

How skills use the connection

The connection is shared infrastructure; a skill is a bounded use of it. A meeting-preparation skill may read approved calendars, documents, and CRM records. A portfolio-review skill may read a specific KPI table and produce a controlled briefing. Skills selected from the existing Structishen Skills Library inherit the client’s data boundaries and do not expand them by default.

Connect the minimum source to a measurable task.

Structishen designs the permission model, retrieval path, validation, approval, and evidence trail around the business outcome.

Discuss a connected workflow →