Data Sovereignty Isn’t Data Residency: Audit Microsoft Fabric Copilot and SQL Server Copies

Yet as database veteran Johan Lamberts argues in a pointed analysis for ITWeb, confusing a physical street address with actual data sovereignty is a fast track to a breach notification. While South Africa spent September debating national digital destiny at Durban's GovTech conference and weighing ministerial sovereignty agendas, domestic organizations were busy losing control of their most sensitive records—not to foreign intelligence agencies across an ocean, but one unmanaged copy at a time to third-party suppliers right down the road, and through single-click Microsoft Fabric tenant switches toggled on a Tuesday afternoon.

For enterprise IT administrators, database architects, and Microsoft 365 governance leads everywhere, the lesson goes far beyond Southern Africa. Data residency only answers where your bytes sit at rest. Data sovereignty answers who—who holds the encryption keys, who has valid credentials to query the tables, who patches the aging SQL Server or Windows Server engine underneath, and which administrator just authorized Azure OpenAI to ship your prompts across continents because an executive wanted a prettier Power BI dashboard.

The Illusion of the Local Address​

To understand why the "where" question is largely a solved problem in South Africa, look at the physical infrastructure footprint. Microsoft has operated Azure cloud regions in Johannesburg and Cape Town since 2019 and committed an additional R5.4 billion toward South African cloud and AI infrastructure through the end of 2027. Amazon Web Services and Google Cloud also maintain local regions, contributing to a wider national build-out that, as noted in President Cyril Ramaphosa's 2026 State of the Nation Address, encompasses 55 data centers and more than R50 billion in projected three-year digital infrastructure investment.

National policy reflects a similar pragmatism. South Africa's National Policy on Data and Cloud, introduced by the government in mid-2024, mandates domestic storage strictly for government data touching national security, while leaving cross-border commercial transfers to the existing guardrails of the Protection of Personal Information Act (POPIA). Even so, Lamberts notes that a recent PwC Africa cloud survey found 45% of organizations cite geopolitics as a driver of their infrastructure choices—a concern that too often starts and ends with asking a cloud vendor which city hosts the server rack.

Why isn't a local rack enough? First, there is legal jurisdiction. Lamberts points to testimony from July 2025, when Microsoft's legal director in France told a French Senate inquiry under oath that the company could not guarantee data held for French public bodies would never be handed over to United States authorities without French consent—while noting it had never happened and that Microsoft is contractually bound to challenge unfounded demands. That reality is not a Microsoft policy quirk; it is the result of U.S. statutory reach (such as the CLOUD Act), which applies to any U.S.-headquartered hyperscaler regardless of where the concrete is poured.

Second, and far more urgent for daily IT operations, is that hyperscalers have responded to jurisdictional anxiety with genuine engineering controls: customer-managed encryption keys, strict data boundaries, and localized processing options. The tools to lock down cloud estates exist. The problem is that enterprise teams rarely configure them properly, and even fewer govern what happens when data is copied out of primary systems.

Section Summary:

• Residency vs. Sovereignty: Local cloud regions (like Azure's Johannesburg and Cape Town facilities) solve physical residency ("where"), but sovereignty requires operational control over access, keys, patching, and jurisdiction ("who").
• Jurisdictional Reality: Physical location inside a country does not automatically immunize data held by foreign-headquartered hyperscalers from foreign legal statutes, making customer-held encryption keys and strict technical boundaries essential.

How Control Actually Leaks: Nightly Extracts and Shadow Copies​

If you ask a CISO where customer data lives, they will show you an architectural diagram of their primary ERP, their core SQL databases, or their Azure tenant. If you ask a veteran database administrator, they will tell you where the data actually lives.

Lamberts describes the kind of sprawl visible in almost any mature Windows and SQL Server estate without even looking hard:

• The Forgotten Reporting Server: A nightly finance extract set up years ago because reporting queries were slowing down the production ERP. The target server was never added to the backup schedule, its database engine sits two versions behind production, and its share permissions are wide open to the Everyone group.
• The Seven-Copy Customer Table: A core identity table replicated into the CRM, the marketing automation tool, the loyalty platform, an external KYC/identity-verification vendor, and finally a data warehouse meant to consolidate them all—which merely becomes copy number seven.
• The Immortal Spreadsheet: An Excel export containing 40,000 national identity numbers dumped onto a departmental file share for a 2022 project and emailed around twice since.

None of these secondary stashes show up on a formal IT asset register, because asset registers track systems, not copies. Citing a June 2026 Redgate survey of 2,150 database professionals, Lamberts notes that 77% of organizations operate without a formal data governance process. Every unmanaged extract carries its own patch state, its own Active Directory or Entra ID permission scope, and—when handed to a vendor—its own legal jurisdiction.

September's Domestic Supply-Chain Reckoning​

The danger of unmanaged copies stopped being theoretical for South African enterprises in September 2026. A string of high-profile third-party breaches proved that an organization's own perimeter can remain completely intact while its customers' identities are auctioned on the dark web.

On September 9, 2026, the Dire Wolf ransomware group claimed on its leak site to have exfiltrated 200 gigabytes of data—allegedly 3.57 billion rows across production databases and Amazon S3 storage—from RelyComply, a South African anti-money-laundering and identity-verification provider serving regulated financial firms. According to an independent public safety notice published by Ubuntu Guard Cyber, RelyComply reported unauthorized access to its environment on September 2 and identified the breach on September 8. While Ubuntu Guard Cyber noted that the ransomware group's specific volume claims were not independently verified, the downstream fallout was immediate: major financial platforms including EasyEquities, SA Home Loans, Satrix, and Peregrine Capital had to notify customers in mid-September that their identity documents, ID numbers, or property addresses may have been exposed—even though the financial institutions' own core systems, logins, and account balances were untouched.

Lamberts highlights two other September supply-chain shocks reported in the local market: a vehicle-tracking provider breach exposing data tied to more than two million subscribers by September 8, and a life-insurance policy-administration supplier whose decommissioned support-ticket system was compromised via a reused employee password, exposing roughly 400,000 customer records across 45 insurers via file attachments rather than core policy databases.

In the wake of the insurance supplier incident, both South Africa's Information Regulator (enforcing POPIA) and the Prudential Authority (enforcing Joint Standard 2) delivered a cold shower to executives: legal accountability and the 24-hour notification clock belong to the responsible party (the bank or insurer), not the outsourced vendor that lost the files. Furthermore, as Lamberts observes, even when you sign a domestic contract with a local supplier, their privacy terms frequently permit cloud processing across the UK, the European Economic Area, the U.S., or elsewhere. If you do not know where your supplier's copies sit, your domestic cloud strategy is theater.

Section Summary:

• Copy Sprawl: Sovereignty is routinely lost inside national borders through untracked ERP extracts, unpatched staging databases, and exported spreadsheets.
• Third-Party Exposure: The September 2026 RelyComply incident and parallel supplier breaches forced unbreached banks and insurers to send breach notices because customer identity copies were compromised in vendor environments.
• Regulatory Liability: Regulators enforce compliance against the primary data owner, regardless of which downstream supplier dropped the ball.

The Tuesday Afternoon Sovereignty Decision: Microsoft Fabric and Copilot​

You don't even need a third-party supplier breach to inadvertently export your data across borders. Sometimes, all it takes is an impatient business analyst and an obliging IT administrator clicking a toggle in the Microsoft Fabric admin portal.

Consider an enterprise running Microsoft Fabric on a capacity in the South Africa North region. All standard Fabric workloads run locally, and OneLake storage keeps the data resting squarely inside South African borders. However, what happens when users want to turn on Copilot in Fabric?

Because the Azure OpenAI Service models powering Fabric Copilot do not reside in every regional data center, Microsoft takes a conservative, privacy-first engineering stance out of the box. As verified in Microsoft Learn's documentation for Copilot and Agent tenant settings, while the base setting allowing users to access Azure OpenAI-powered features is enabled, cross-border data movement outside the US and the EU Data Boundary is blocked by default.

Specifically, Microsoft provides two distinct tenant-level switches under the Copilot and AI settings group that are disabled by default for capacities outside the EU data boundary and the United States:

As Microsoft's architectural documentation for Fabric Copilot explains, when a user submits a prompt, Copilot performs a preprocessing step called grounding. Grounding collects contextual data—such as lakehouse schemas, semantic model metadata, session chat history, or specific data points from workspace items the user has permission to view—and bundles that augmented prompt to send to the Azure OpenAI Service. Microsoft explicitly notes that Azure OpenAI does not train models on customer data, does not cache data at the GPU location, and respects the user's existing workspace roles and item permissions. Furthermore, Microsoft warns in its admin documentation that if users connect Fabric data agents to non-Fabric services—such as Microsoft Foundry, Microsoft Copilot Studio, Microsoft 365 Copilot, or an MCP server—responses returned by those agents may travel outside Fabric's compliance boundary and geographic region under the receiving service's terms.

Microsoft's defaults and documentation here are exemplary: the cross-border switches ship turned off, and the labels state in plain English exactly what happens when you flip them.

The vulnerability, as Lamberts points out, is human governance. On a random Tuesday afternoon, a project team complains that Copilot isn't working in their South Africa North workspace. A Fabric administrator logs into the Admin Portal, sees the toggle blocking external processing, flips it to Enabled for the entire organization so the dashboard team can meet their deadline, and closes the ticket. No board resolution is passed. No Data Protection Officer minutes the cross-border transfer risk. In thirty seconds, an admin has made an enterprise-wide sovereignty decision.

What About Microsoft 365 Copilot In-Country Processing?​

The tension between AI adoption and local processing extends to Microsoft 365 Copilot as well. On November 4, 2025, Microsoft announced plans to bring in-country data processing for Microsoft 365 Copilot interactions to 15 countries, placing South Africa in a wave of eleven nations scheduled to receive local processing during 2026.

However, in an April 3, 2026 update to that announcement, Microsoft refined its timeline and broadened the scope of local data inferencing to span Microsoft 365 Copilot, Microsoft Copilot Studio, Power Platform, and Dynamics 365. Under the revised schedule, Microsoft stated that local inferencing is expected by the end of 2026 for Australia, India, the United Arab Emirates, the United Kingdom, and the United States—with Canada following in 2027, Japan in 2028, and EU/EFTA nations served regionally via the EU Data Boundary. South Africa was omitted from the revised April 2026 timeline without a newly assigned target date. For South African enterprises, that means waiting on vendor hardware rollouts is not a substitute for governing cross-border AI traffic today.

Section Summary:

• Granular Fabric Controls: Microsoft ships cross-geographic Azure OpenAI processing and storage settings disabled by default in Fabric for regions outside the US and EU.
• What Actually Travels: Enabling cross-region processing doesn't export an entire OneLake database, but it does allow prompts, schemas, metadata, and user-accessible grounding data points to be processed abroad—and enabling cross-region storage keeps notebook and data agent chat history abroad for up to 28 days.
• Shifting AI Roadmaps: With South Africa left without a target date in Microsoft's April 2026 update for local M365 Copilot inferencing, admins must actively audit tenant toggles rather than assuming AI traffic stays domestic.

The Five Pillars of a Sovereign Windows & Data Estate​

If you want your organization to actually own its data—rather than merely renting a local zip code—Lamberts proposes five operational questions that every IT director and board must be able to answer. Let's translate those into a concrete checklist for Windows Server, SQL Server, and Microsoft Cloud administrators:

1. Can you count and govern every copy?
   Stop treating your CMDB strictly as a hardware and VM inventory. Audit nightly SQL Server Integration Services (SSIS) packages, scheduled PowerShell exports, staging databases, and departmental file shares. Know who owns each extract, lock down Everyone and broad Entra security group permissions, and kill legacy reporting copies that have outlived their usefulness.
2. Do you hold your own encryption keys?
   All three major hyperscalers support Customer-Managed Keys (CMK) and Bring Your Own Key (BYOK) architectures (such as Azure Key Vault Managed HSM). Yet most enterprises still leave key management to the vendor's default platform keys. For truly sensitive workloads, take custody of the cryptographic keys.
3. Have you audited the credentials that can reach personal data?
   According to Lamberts, Sophos's 2026 ransomware study—examining 135 attacked South African organizations over the prior year—found that in 85% of local incidents, the ransomware deployment and the organization's most serious identity compromise were the exact same event (compared to 67% globally). Attackers aren't cinematic hackers breaking ciphers; they are logging in with valid credentials. Enforce phishing-resistant MFA, strip standing privileges with Entra ID Privileged Identity Management (PIM), and audit service accounts querying production databases.
4. Is the underlying database and OS engine still supported?
   A database sitting in a Johannesburg or London data center on an unpatched or end-of-life operating system is not sovereign—it is just a compromised server with a local IP address. As of October 2026, Lamberts notes the ticking clock on legacy Microsoft platforms: Windows Server 2016 is down to its final three months before Extended Support ends in January 2027, and SQL Server 2017 has roughly twelve months remaining before its October 2027 end-of-life date. If your data copies sit on expiring engines, migrate or apply Extended Security Updates (ESUs) before attackers do the inventory for you.
5. Do you know what your suppliers hold—and can you take it back?
   Under corporate governance frameworks like South Africa's King V (Principle 10) and data protection laws worldwide, the board owns third-party risk across jurisdictions. Demand architecture diagrams from KYC vendors, payroll processors, and SaaS providers. Know which cloud regions they use, how they segregate your tables, and verify that they have a tested mechanism to cryptographically wipe and return your data when a contract ends.

Step-by-Step: Auditing Your Microsoft Fabric Copilot Geo-Processing Settings​

Before your next compliance meeting, verify that nobody has quietly authorized cross-border AI data flows in your Microsoft Fabric tenant:

1. Sign in to the Microsoft Fabric Admin Portal using an account with the Fabric Administrator role.
2. Navigate to Tenant settings and scroll down to the Copilot and AI section.
3. Locate Data sent to Azure OpenAI can be processed outside your capacity's geographic region, compliance boundary, or national cloud instance.
4. Locate Data sent to Azure OpenAI can be stored outside your capacity's geographic region, compliance boundary, or national cloud instance (as well as the parallel toggles for OpenAI as a subprocessor).
5. Verify whether these settings are Disabled. If business requirements necessitate enabling them for specific workloads, do not enable them for The entire organization. Instead, scope access strictly to dedicated Specific security groups and consider delegating the setting to the capacity level so only non-sensitive capacities allow cross-border processing.
6. Document the configuration decision, the specific security groups allowed, and the formal sign-off from your Data Protection Officer or risk committee.

The Bottom Line for IT Leaders​

Cloud residency is a commodity you can purchase with a credit card, and hyperscalers have made it easier than ever to keep primary storage inside your national borders. Operational control, however, cannot be bought off the shelf. It is won or lost in the trenches of database administration: tracking down forgotten CSV exports, retiring aging Windows Server 2016 boxes, locking down supplier schemas, and putting formal change-control around the AI tenant settings that bridge your local data to global compute.