Your employee may be allowed to read a document without being allowed to turn it into a customer promise. Copilot's new source choices make that distinction worth testing.
What Microsoft has actually changed
Choosing where an assistant looks sounds like a small interface improvement. For a business with customer records, internal discussions and approved guidance spread across several applications, it changes a much bigger question: which evidence should influence this particular answer? Microsoft's latest connector announcement makes that question worth revisiting before teams simply connect more systems.
The October 2026 Copilot Connectors update, published on 11 November describes Source Filter, which lets people scope a question to one or more sources. It also describes a sources list in Settings and contextual suggestions that bring connector results into an answer when those results were initially omitted. These are vendor-described capabilities, not independent evidence of better answers in your organisation. Check availability in your own tenant, licence and relevant Copilot experience before planning training around them.
Two concrete changes show why the announcement is more than a new menu. Microsoft says the ServiceNow knowledge-base attachment limit has increased from 20 MB to 64 MB. It also says the Salesforce connector now covers custom fields and additional content including Tasks, Events, Campaigns, Comments and Attachments. These are product specifications, not productivity figures. Their business significance is that the assistant may encounter a wider range of evidence than your original pilot covered.
A bigger information pool can help staff answer questions that previously needed several applications. It can also include exploratory notes, disputed comments and material intended for a different purpose. The useful management response is not to reject connectors. It is to decide what each newly available source is allowed to contribute to each type of work. An announcement about richer grounding should trigger a short review of your task boundaries, rather than an assumption that every additional source improves every answer.
A wider evidence pool is a reason to review the task, not proof that the answer has improved.
Source selection is not an access-control policy
A source filter and an access permission solve different problems. A filter narrows the information considered for a question. Access controls determine what an identity is entitled to read or do. Keeping those concepts separate matters because a useful interface choice can otherwise become an imaginary security guarantee in a manager's head.
Consider an illustrative sales manager who legitimately has access to both customer records and an internal discussion space. When preparing a customer-facing renewal explanation, they need the customer's agreed terms and approved product information. They may not need every internal commercial discussion they are permitted to read. Selecting a source can help focus the task, but the fact that the manager can access a discussion does not make that discussion appropriate for inclusion in a customer message.
Microsoft's connector announcement separately discusses Salesforce field-level and record-level controls and improvements to Confluence identity matching. That separation is important. Permission enforcement is a connector and identity configuration concern, not something to replace with a training slide saying that staff should remember to click the right filter. Nor should a prompt asking the assistant to ignore a sensitive system be treated as equivalent to restricting access to it.
What this means in practice is simple: write down the intended evidence sources for one task before turning it into a repeatable workflow. Then ask your administrator how the relevant access is actually enforced and test with representative authorised and unauthorised accounts. Separately check that an authorised employee's output stays appropriate for its intended audience. These are different acceptance checks. A pass on one cannot stand in for the other, and the article is not claiming that Source Filter itself offers a hard isolation boundary or changes the underlying permissions.
Choosing where to look is not the same as deciding who may see what.
Richer connector content needs clearer evidence roles
A comment, a customer record and an approved policy can all contain useful information without carrying the same authority. The practical challenge is to help staff and their assistant distinguish a source that establishes a fact from one that merely explains why somebody is asking a question. The new connector coverage makes that distinction harder to leave implicit.
Microsoft says Confluence Cloud results can now draw on page comments, attachments and their metadata. A comment may contain an excellent correction, but it might equally be a suggestion that nobody accepted. Similarly, a Salesforce activity can record a conversation without changing the customer's contract. The product update creates an opportunity to find more context. It does not appoint every retrieved item as an approved business instruction.
Define evidence roles in ordinary language. For a renewal response, the signed agreement establishes the existing commitment, the approved pricing record establishes the current offer, and a conversation note explains the customer's concern. If those sources conflict, the assistant should surface the conflict rather than silently choose whichever sentence sounds most helpful. This is a proposed operating approach, not a claim about an automatic evidence hierarchy built into Copilot.
The wider trend is visible in Microsoft's October Copilot product roundup, which describes a SharePoint knowledge-base experience bringing together emails, messages, meetings and SharePoint content, with an October Frontier rollout. Availability is qualified, and aggregation is not the same as approval. Before adopting any shared knowledge experience, nominate the person who decides what counts as authoritative for the task. Give them a short conflict-handling rule that staff can follow without needing to understand retrieval architecture. That ownership is usually more useful than asking everyone to write a longer prompt.
More context should not give an unapproved comment the authority of a signed agreement.
Test the new source choices against a real business decision
A convincing demonstration normally shows a question that the system can answer. A useful business test asks whether the new source choices change an answer in the way you intended. Start with a narrow task, such as preparing an internal account briefing, and a small set of representative cases for which a knowledgeable employee can establish the correct evidence.
Run each case with the existing source configuration, then with the proposed filter or connector coverage. Keep the question and review criteria consistent. Check the actual references, not just whether the prose sounds better. Record whether the response used the intended source, missed an essential fact, introduced an unsupported claim or exposed material unsuitable for the audience. Include a case with no adequate evidence and a case with conflicting sources. The expected result may be a request for clarification rather than a longer answer.
This approach has a technical rationale beyond Microsoft's announcement. In its 23 October guide to context engineering and retrieval, LangChain identifies inefficient searching, excessive context and irrelevant or stale context as distinct problems. That is framework-vendor guidance, not a performance study. Its useful implication for a business buyer is that adding information and improving retrieval are not interchangeable activities.
What this means in practice is that you should test source selection separately from model capability. A stronger model cannot establish an agreement that the selected sources do not contain. Equally, a well-chosen source does not guarantee the model will describe it correctly. Keep both checks visible. For customer-facing work, measure the employee's time to an acceptable, evidence-backed response, including correction and review. Do not report a faster first draft as a faster completed task if the reviewer must spend longer untangling where its claims came from.
Test whether the chosen evidence supports the decision, not whether the answer sounds more confident.
Use the built-in controls first, with a named owner
The strongest counterargument is fair: if Microsoft already offers permission-aware connectors and source selection, why add another process? Because the software can enforce a configured entitlement without knowing the commercial purpose of your next message. That is not an argument for building a competing platform. It is a reason to use the existing controls deliberately and keep the operating rules small enough that staff actually use them.
Begin with the capabilities already available in your approved Microsoft environment. Ask your administrator to demonstrate source selection, connector permissions, connection status and relevant diagnostics for the exact experience your team uses. Microsoft describes editing published connections and applying permission changes through an ad hoc crawl. Do not assume a configuration edit has already propagated everywhere. Establish how to verify the resulting state before relying on it for sensitive work.
For UK organisations, keep personal information obligations separate from product convenience. The ICO guidance on accuracy and statistical accuracy distinguishes accurate personal data from the statistical correctness of an AI system. A source filter is not a certificate of UK GDPR compliance. The ICO says this guidance is under review following the Data (Use and Access) Act, so check the current position when making a legal or compliance decision.
Give one business owner responsibility for the task and one technical owner responsibility for the connector configuration. Agree when to pause use: an unexplained source change, a permission failure, or a response that attributes an unapproved commitment to the business. Keep a straightforward manual route available. The next step is a bounded trial with real review, not a custom gateway or a company-wide connector rollout. Microsoft's update is useful precisely because it offers more existing controls to test. The value comes from making those controls serve a defined business task, not from treating a larger catalogue as an adoption target.
Use the existing product first, but make one person responsible for what counts as acceptable evidence.
Key Takeaways
- Check Source Filter availability in your exact tenant and Copilot experience before changing staff workflows.
- Test access enforcement separately from whether an authorised employee's response suits its audience.
- Define which sources establish commitments and which only provide background for one task.
- Compare references, accepted-response time and correction work before and after connector changes.
- Assign business and technical owners and verify permission changes have propagated before relying on them.
