What does an employee asking through approved Claude do with a new supplier request?
In this fictional approved Claude view, an operations employee asks about supplier intake through Brain. The handbook says to use the intake form and identify the purchasing owner. The cited answer supports that internal task without opening a nearby restricted investigation. This illustration connects no real Claude session; the source owner remains responsible for maintaining the procedure.
Supplier intake handbook
Source excerpt · Supplier intake handbook
New supplier requests use the intake form. Include the purchasing owner and the purpose of the proposed service.
Shadow AI grows where useful work and approved tools diverge
Shadow AI is AI use that falls outside an organization’s approved tools, policies or oversight. Understanding which information people need can help a team offer an approved way to consult it.
Shadow AI is the use of AI tools or workflows outside an organization’s approved rules and oversight. It can begin with an ordinary task: someone needs to explain a process quickly, cannot find the approved reference and uses a convenient assistant. The risk depends on what information they share, which account processes it and what rules apply. Do not assume every employee or every outside tool is behaving the same way.
An approved alternative needs to help with the task that led to the workaround. Brain makes an intentional company knowledge collection available to compatible AI clients, with sources behind the answer. For a process question, a teammate can consult the approved handbook instead of assembling a new archive of copied files. Familiar clients can fit that workflow when their use and account configuration are approved by the organization.
Useful answers and access boundaries need to arrive together. Give the collection an owner, decide which groups and agents should reach it and test a restricted question as well as an ordinary one. A source citation helps the reader inspect the basis of an answer; it does not grant access to a private folder. The demonstration below makes that distinction visible using fictional material and a limited-access state.
Keep the wider policy practical. Explain permitted purposes, approved accounts, sensitive-data restrictions and how to request a missing tool. Review the activity your connected systems actually record, and use other authorized evidence where needed. Brain’s workflow and records do not discover every personal account or every external upload. Combine a useful approved path with proportionate controls and clear ownership, then investigate the reasons people still choose another route.
Connect. Ask. Govern.
From scattered documents to a shared answer
01
Connect the approved supplier references
Choose the supplier intake handbook and acceptable-use guidance the team is allowed to consult. Keep restricted investigation notes outside the ordinary employee collection and name the owner of the maintained process.
02
Ask where approved work happens
Let an approved compatible client consult the intake procedure for a specific supplier question. Inspect the cited instruction so the permitted workflow is useful without a new bundle of copied company documents.
03
Review the connected activity
Test a restricted request and inspect the identities and activity the system actually records. Use the integration request route for missing tools, and investigate wider AI use through other authorized evidence where needed.
See the idea in action
shadow ai: questions with evidence
Fictional examples. These interactions do not query your Brain or test real permissions.
In this fictional approved Claude view, an operations employee asks about supplier intake through Brain. The handbook says to use the intake form and identify the purchasing owner. The cited answer supports that internal task without opening a nearby restricted investigation. This illustration connects no real Claude session; the source owner remains responsible for maintaining the procedure.
Supplier intake handbook
New supplier requests use the intake form. Include the purchasing owner and the purpose of the proposed service.
The fictional investigation folder is restricted to its review group. Switch to limited access and the answer and excerpt are withheld. An approved AI workflow should help with permitted work without turning nearby sensitive documents into context. This illustration does not inspect a real employee account or establish live access for your organization.
Restricted vendor investigation
Private review: investigation notes and unresolved findings for a proposed vendor.
For the fictional week of 21–27 September 2026, the summary records a procurement assistant’s allowed handbook read on Monday, an operations assistant’s allowed read on Wednesday and a denied vendor-investigation request on Friday. Review the dated records against approved tasks. This scoped illustration covers those connected identities only; it cannot report unrelated personal accounts or unrecorded uploads.
Fictional weekly connected activity: 21–27 September 2026
Review window: 2026-09-21 through 2026-09-27 UTC. 2026-09-21T09:15Z — procurement-assistant — supplier-handbook — read allowed. 2026-09-23T14:20Z — operations-assistant — supplier-handbook — read allowed. 2026-09-25T10:05Z — operations-assistant — vendor-investigation — read denied. Scope: these connected identities and objects only; personal accounts and external uploads are not covered.
The fictional acceptable-use note asks staff to explain the tool and task through the integration request path before sending confidential material to another service. A useful request describes the work without attaching the private documents. The organization decides approval; submitting an integration request does not make a tool available or authorize its use.
Approved tools request note
Describe the tool, business task and source type. Do not include confidential documents, credentials or private employee records.
Open a question, then inspect its source excerpts.
Brain is a knowledge layer reached through MCP, the Model Context Protocol. Claude, Cursor and Codex are examples of compatible clients in the existing product. Client support and setup differ, so check the current connection instructions rather than assuming every assistant has the same capabilities.
The current public setup describes Google Drive and Notion as connected sources, alongside documents you choose to add. Start with a focused collection and inspect the actual connection screen for availability. A tool appearing in a roadmap or illustration does not mean its connector is ready for your account.
Google Drive
Connect the documents your workspace needs.
Notion
Bring approved pages into a shared knowledge layer.
Your documents
Add a focused set of knowledge you own.
Keep access intentional
Reduce opportunities for AI data leakage in the approved path
Set the source collection, employee groups and agent scope deliberately. A maintained intake reference should be easier to consult than a private investigation folder. Review recorded connected activity in that context, while recognizing that personal accounts and unrelated external uploads need other authorized evidence. Clear acceptable-use rules and a practical tool-request route complement the product boundary.
Fictional content-blind access record
Actor
Example teammate
Action
Read approved source
Outcome
Allowed within the example workspace
Fictional metadata only. No document content is shown, and this is not a record from your account or proof of live enforcement.
Avoid copying the supplier archive for one process question
The approved intake question needs the process and its relevant exceptions, not every historical vendor file. Focused context can reduce repeated input when it is sufficient for the task. Compare real usage and source-backed answer quality in the pilot, including retrieval and output charges, instead of treating adoption or token savings as a guaranteed result.
Combine a useful approved path with proportionate controls
Choosing an approach for this workflow
Approach
What to consider
Restrict particular AI uses
Can be appropriate for specific data or tasks. Make the restriction clear and provide a permitted way to complete necessary work. Whether a broad ban is effective depends on the environment; do not assume every organization needs the same rule.
Block domains or endpoints
Can enforce part of a technical boundary in managed environments. Coverage depends on devices, network paths and configuration. Assess it alongside approved alternatives, rather than assuming a block identifies every account or answers every legitimate knowledge question.
Publish an acceptable-use policy
Sets expectations and accountability. Concrete examples, accessible approved tools and a request process make it easier to follow. A policy document alone does not reveal which sources a particular assistant retrieved.
Offer a governed Brain workflow
Helps approved compatible clients consult maintained company knowledge with intentional access and records. Evaluate adoption and the actual connected activity. It can contribute to reducing workarounds, while other controls address activity beyond the connected workflow.
shadow ai: frequently asked questions
What is shadow AI?
Shadow AI is AI use outside an organization’s approved rules and oversight. It may involve an unapproved tool, account, source connection or purpose. The label does not mean every employee is misusing AI. Establish your approved workflow and data restrictions clearly so people can distinguish a permitted task from a risky workaround.
Why is shadow AI a risk?
Unapproved use can move company information into services, accounts or retention arrangements the organization has not assessed. It can also make ownership and activity harder to review. The severity depends on the information and use. Identify the actual data path and applicable rules before treating every external AI interaction as the same incident.
How do you find out which AI tools staff use?
Use authorized evidence from your managed systems, approved-tool inventory and discussions with the teams doing the work. Brain’s connected activity can describe the identities and requests it records, but it is not a universal detector of personal accounts or external uploads. Respect the organization’s privacy and monitoring requirements when investigating broader usage.
Should companies ban ChatGPT?
That depends on the proposed use, data sensitivity, account terms and available controls. Some restrictions may be necessary; other tasks may fit an approved configured account. State the rule precisely and provide a practical permitted workflow. Brain does not decide your acceptable-use policy or authorize a client your organization has prohibited.
How do you write a company AI policy that works?
Describe approved purposes, tools, accounts and data categories in terms people can apply to daily work. Name the source owner and approval route, explain restricted material and provide examples. Pair the policy with usable tools and proportionate controls. Review questions and incidents so confusing rules or missing workflows can be corrected deliberately.
How does Brain reduce shadow AI?
Brain offers a source-backed knowledge workflow through compatible clients so approved work can be easier to complete without repeated document copying. Intentional access and records help make that workflow reviewable. Measure whether teams use it and why workarounds persist. It supports an approved alternative; it does not guarantee elimination of all outside AI use.
Does an approved workflow prevent every ChatGPT data leak?
No tool selection alone can guarantee that company information never reaches an unintended account or service. Assess the actual configuration, granted actions, account terms and employee workflow. Brain governs its connected knowledge path; other clients and external uploads have separate boundaries. Use approved synthetic tests and authorized incident evidence rather than assuming universal coverage.
Give one team a practical approved path
Choose a recurring internal question, connect its approved references and make the source-backed answer easy to inspect. Use that pilot to improve the policy and the workflow together.
