A finance team considering Microsoft Copilot should define the tasks it wants to support and check the environment those tasks depend on. An early question is whether the environment can support the intended experience — and for finance teams in particular, there are documented constraints that decide the answer before anyone writes a prompt.
This article sets out those constraints, what Microsoft commits to on data handling, what remains the organisation's own obligation, and how SCSB would scope a pilot that produces a decision rather than an impression.
Readiness is an environment question first
Microsoft's app and network requirements set out the prerequisites plainly. Licensed Microsoft Copilot capabilities require a Microsoft Copilot licence and a qualifying prerequisite licence; eligible Copilot Chat access is a different experience. Users need Microsoft Entra ID accounts. Deploy the Microsoft 365 apps and services required for the specific experiences in scope. Some features, including file restore and OneDrive management, require the user to have a Microsoft OneDrive account.
Two exclusions are easy to miss and both are absolute rather than configurable. Microsoft Copilot is not available where device-based licensing is used for Microsoft 365 Apps for enterprise. And for the web versions of the core applications, third-party cookies must be enabled — a setting many organisations have deliberately restricted for unrelated reasons.
Feature-specific conditions apply on top. To let Microsoft Copilot reference meeting content after a meeting ends, transcription or recording must be enabled — which is itself a decision with consent and retention consequences, not a switch to flip on the way to a demonstration. Microsoft Loop and Microsoft Whiteboard each have to be enabled for the tenant separately. And the Office Feature Updates task must be allowed to run on its normal schedule and reach the network resources it needs, because core experiences depend on it.
Check the mailboxes and experiences the team will use
Microsoft updated its requirements on 30 September 2026. The service is now named Microsoft Copilot, formerly Microsoft 365 Copilot. Its current requirements support primary Exchange Online mailboxes and content a user is permitted to access in archive, shared and delegate mailboxes. Group mailboxes remain unsupported. Availability depends on the app, licence, tenant configuration and the particular experience.
For a finance team using payables or receivables shared mailboxes, test the intended tasks with authorised users and representative messages. Record which information can be retrieved, which permissions apply and where human review is required. Shared-mailbox use alone is not a reason to rule out a pilot.
Permissions become the deployment decision
Microsoft's data, privacy and security documentation states that Microsoft Copilot only surfaces organisational data to which individual users have at least view permissions, and that the grounding process honours the user identity-based access boundary. Microsoft is direct about what follows: it is important that organisations use the permission models available in the underlying services so that the right users have the right access to the right content.
The practical consequence is that Copilot can make existing oversharing easier to discover. Also assess any new agent, connector or consent introduced by the rollout. Historic oversharing that nobody noticed while it required someone to know a file existed becomes discoverable when a colleague can simply ask a question in natural language.
SCSB's recommended sequence is therefore to review before enabling: identify the locations in scope, establish who currently has access and whether that remains appropriate, confirm how sensitive material is labelled, and resolve what the review turns up. Where content is encrypted through sensitivity labels or information rights management, Microsoft records that Microsoft Copilot honours the usage rights granted to the user, so labelling done properly continues to do its work. Enabling a new tool is not a substitute for fixing a permissions problem the review reveals.
Where data goes, and what Microsoft commits
Three commitments are documented and worth stating accurately, because they are frequently overstated in both directions.
- Prompts, responses, and data accessed through Microsoft Graph are not used to train foundation large language models, including those used by Microsoft Copilot.
- Interactions are stored — the prompt, the response, and citations to the information used — as the user's activity history. That data is encrypted at rest, and administrators can view and manage it using Content search or Microsoft Purview, including setting retention policies for it. Users can delete their own activity history from the My Account portal.
- Microsoft documents EU Data Boundary safeguards for Microsoft Copilot, but these do not cover every route. Web-search queries are outside that boundary; Anthropic models are also currently excluded from it and, where applicable, from in-country processing commitments. Customers outside the EU may have queries processed in the US, EU or other regions. Check the selected experience, enabled model providers, web-search setting and applicable terms before making a processing-location commitment.
That last point is the one Malaysian organisations should read carefully rather than skip. A Malaysian tenant is outside the EU Data Boundary. Microsoft's Advanced Data Residency and Multi-Geo offerings carry data residency commitments for Microsoft Copilot customers, so residency is a subject an organisation can address — but it is addressed by examining the commitments that apply to its own tenant and agreements, not by assuming a default.
PDPA 2010 obligations stay with the organisation
Microsoft documents product behaviour, controls and contractual commitments. It does not determine an organisation's obligations under the Personal Data Protection Act 2010, and no configuration of Microsoft Copilot discharges them.
Where the Act applies, identify the organisation’s responsibilities as a data controller and any applicable processor duties: the basis for processing, the notice given to data subjects, security measures, retention, and the treatment of any transfer outside Malaysia. Those questions are for the organisation's own legal or compliance advisers, and where a determination is needed on the current requirements, the Personal Data Protection Commissioner is the authority to consult rather than a vendor's documentation or a consultancy's article. SCSB's role here is to describe accurately what the technology does with data so that the assessment can be made on correct facts — not to reach the conclusion.
Network conditions that break it quietly
Microsoft records that several Microsoft Copilot integrations rely on WebSockets, and that networks must support full WSS connectivity to the relevant Microsoft domains. Three common configurations interfere: a perimeter that blocks the WebSocket protocol, devices attempting TLS inspection of those connections, and proxy servers enforcing aggressive connection timeouts.
All three are ordinary in a well-run corporate network, and the failure mode is unhelpful — features behave inconsistently rather than failing cleanly. Include these network checks before the pilot, because a pilot that fails on network configuration will be reported as the service not working.
Separately, privacy controls for connected experiences can remove Microsoft Copilot features entirely from the core applications. An organisation that has turned off connected experiences that analyse content will find the features unavailable, and should record why those settings are in place before deciding whether a change is appropriate.
Designing a pilot that produces a decision
A useful pilot has one defined audience, a small number of permitted tasks, a named owner, and a way to record what was learned. SCSB would fix the success criteria in advance, because otherwise the outcome is a set of impressions.
Set expectations with users explicitly: which data sources are in scope, when to verify an answer, how to handle an error, and where to escalate a suspected information-access concern. Microsoft states plainly that responses generative AI produces are not guaranteed to be entirely factual and that users should apply judgement before relying on output. For finance work, that means generated text is a draft for review, never a substitute for the source record, professional review, or an approval process.
If agents are in scope, treat them as a separate decision. Administrators can review the permissions and data access an agent requires, along with its terms and privacy statement, and control which agents are permitted.
Keep a pilot scorecard for each approved task
For an illustrative finance pilot, ask an authorised user to draft a summary of an approved management-report pack. The reviewer should be able to trace each factual statement to that pack. The example tests a drafting task; it does not delegate accounting judgement or approval to the tool.
- Task and boundary: the permitted question, intended audience and actions that still need a person.
- Source: the approved files, period, version and access rights, with unrelated material excluded.
- Expected result: the points and source references a reviewer expects to see, plus known gaps the answer should acknowledge.
- Observed result: errors, missing evidence, inappropriate disclosure, and the corrections the reviewer made.
- Effort: time to prepare, prompt, check and correct the output compared with the same task done without the tool.
- Decision: named reviewer, acceptance or rejection, remaining conditions and the next review trigger.
Stop or narrow the pilot if a user can retrieve information they should not have, a material statement cannot be supported, or review effort removes the intended benefit. Record those outcomes alongside successful tests. A useful pilot can conclude that readiness work should come first.
When a simpler answer is the right one
Not every organisation should adopt Microsoft Copilot, and SCSB will say so where the facts point that way. Where content is poorly organised, permissions need remediation or the immediate problem is a manual reconciliation, compare readiness work with the actual business need before purchasing licences. A shared-mailbox workflow should be tested against current supported experiences, rather than treated as an automatic exclusion. A permissions and information-architecture exercise is often the higher-value project, and it is the one that has to happen anyway.
Where an organisation concludes the readiness work is worth doing, an agreed Microsoft 365 implementation scope can be discussed under SCSB's Microsoft 365 implementation service. Any rollout remains subject to the organisation's own licensing, permissions, data governance and approval decisions.
For a separately scoped custom business agent, explore Microsoft Copilot Studio Implementation. Its knowledge sources, tools, licensing and acceptance tests should be assessed independently from Microsoft Copilot features in Microsoft 365.
General-information limitation
This article is general technology information about Microsoft Copilot requirements and documented behaviour. It is not legal, privacy, security or compliance advice, and it does not determine any organisation's obligations under the PDPA 2010 or any other requirement, nor confirm that a particular configuration meets them. Product requirements, controls and commitments are set by Microsoft and change over time. Confirm the current position against Microsoft's published material, and obtain appropriate legal or privacy advice for any compliance conclusion.
Exchange Online, Microsoft, Microsoft 365, Microsoft Copilot, Microsoft Entra ID, Microsoft Graph, Microsoft Loop, Microsoft OneDrive, Microsoft Purview and Microsoft Whiteboard are trademarks of the Microsoft group of companies.