Privacy Comparison

Local AI vs Copilot

Microsoft operates two "Copilot" products with very different privacy profiles. Both send your data to Microsoft's cloud. Here is exactly what that means — and how local AI eliminates that exposure entirely.

🔒Privacy Local wins — zero egress

💰Cost Local wins at scale

🔧IDE Integration Copilot ahead

🏢M365 Integration Copilot wins

Two Products, One Name

Microsoft uses "Copilot" for two distinct products. This page covers both.

GitHub Copilot

AI pair programmer integrated into VS Code, JetBrains, Neovim, and other editors. Powered by OpenAI Codex and GPT-4o models hosted on Azure.

Sends surrounding code context with every suggestion
Telemetry includes file paths, language, keystroke patterns
Business plan: 60-day prompt retention by default
Individual plan: prompts may be used for model training
$10–$19/user/month depending on tier

Microsoft 365 Copilot

AI assistant integrated into Word, Excel, Teams, Outlook, and SharePoint. Powered by GPT-4 via Azure OpenAI Service.

Accesses your entire M365 tenant — emails, docs, meetings
Processes content through Microsoft Graph on Azure
Enterprise tier: data not used for training
$30/user/month (requires M365 E3/E5)
Operates within Microsoft's EU Data Boundary (opt-in)

Important Distinction

M365 Copilot for enterprise has stronger data commitments than GitHub Copilot Individual. However, both products process your data on Microsoft's Azure infrastructure, which means CLOUD Act exposure, cross-border transfer obligations under GDPR, and third-party disclosure risk for privileged content all apply.

Head-to-Head Comparison

Local AI stack vs GitHub Copilot Business and M365 Copilot Enterprise.

Category	Local AI (Ollama + Open Models)	Microsoft Copilot (GH Business / M365)
Data residency	On your hardware only. Zero network egress. WIN	Azure datacentres. EU Data Boundary available but opt-in and incomplete.
Prompt retention	Not stored anywhere. WIN	GitHub Copilot Business: 28-day retention. M365 Copilot: subject to M365 retention policies.
Training use	Not applicable. Weights are static and local. WIN	Business/Enterprise plans: opted out. Individual GitHub Copilot: opted in by default.
GDPR compliance	No cross-border transfer. No DPA required. WIN	Microsoft offers SCCs and EU Data Boundary. More mature than most AI vendors but still a US company under CLOUD Act.
CLOUD Act exposure	Not applicable. No US company in the chain. WIN	Microsoft is a US company. All Azure data — including EU Data Boundary — subject to CLOUD Act compelled disclosure.
Source code privacy	Code never leaves your machine or network. WIN	Every code suggestion requires sending surrounding file context to Azure. GitHub Copilot transmits code snippets continuously during use.
IP / trade secret risk	No third-party exposure of proprietary algorithms. WIN	Proprietary code transits Microsoft infrastructure. May violate NDAs with clients. GitHub has been sued over Copilot training data.
Attorney-client privilege	No disclosure to third party. Privilege intact. WIN	M365 Copilot processes legal documents via Microsoft Graph. Third-party disclosure argument applies.
Air-gap deployment	Works fully offline post-download. WIN	Requires active internet. No offline mode for either product.
Code completion quality	Strong with Llama 3.3 70B and Codestral. Slightly behind Copilot on boilerplate.	GitHub Copilot remains the benchmark for IDE code completion. GPT-4o quality. WIN
M365 / document integration	Possible via local APIs but requires custom integration work.	M365 Copilot natively integrated into Word, Excel, Teams, Outlook. WIN
Fine-tuning / customisation	Full weight access. LoRA, QLoRA fine-tuning on your codebase. WIN	No fine-tuning on custom code. Context is injected at inference only.
Cost — 10 developers	Hardware amortised. Near $0 marginal. WIN	GitHub Copilot Business: $190/mo. M365 Copilot: $300/mo. Combined: $490/mo.
Cost — 100 developers	Same hardware, 10× the users. WIN	GitHub Copilot: $1,900/mo. M365 Copilot: $3,000/mo. Combined: $4,900/mo.
Vendor lock-in	Open weights. No subscription dependency. WIN	Deep M365 integration creates switching costs. Proprietary model. GitHub dependency.
Rate limits	None. Your hardware sets the ceiling. WIN	GitHub Copilot: completions-per-minute limits enforced per seat. M365: throttling during peak usage.

GitHub Copilot: What Gets Sent

GitHub Copilot's telemetry goes well beyond your current file. Understanding the full data surface is critical for IP and compliance teams.

Code Context Transmission — Every Keystroke

GitHub Copilot sends a "prompt" to Azure on every suggestion request. This prompt includes:

The current file's content (up to the cursor position)
Related files open in adjacent tabs (neighbour files feature)
File path and repository name
Programming language identifier
Editor version and plugin version telemetry

For a developer working on a proprietary algorithm, this means the algorithm's implementation transmits to Microsoft's Azure infrastructure dozens of times per hour.

The Neighbour Files Feature

Copilot's "neighbour files" feature — enabled by default — includes content from other open editor tabs in the context sent to Azure. A developer with a config file containing API keys, a schema file with PII structure, or an internal README with architecture details open in a background tab will transmit that content even if they are editing an unrelated file.

28-Day Business Tier Retention

GitHub Copilot Business retains prompts (including transmitted code) for 28 days for trust and safety purposes. This is documented in GitHub's privacy statement but is not prominently disclosed during onboarding. Enterprise tier reduces this window but does not eliminate it entirely — a 28-day window is still 28 days of proprietary code on Microsoft's servers.

M365 Copilot: The Broader Surface Area

M365 Copilot's deep tenant integration creates a larger data exposure surface than most AI tools.

Microsoft Graph Access

M365 Copilot operates through Microsoft Graph, which means it has access to — and can include in its prompts to GPT-4 — your organisation's emails, calendar entries, Teams messages, SharePoint documents, and OneDrive files. A single Copilot query like "summarise the key issues in this deal" can pull in privileged legal correspondence, financial projections, and HR communications simultaneously and send them as a combined prompt to Azure OpenAI.

EU Data Boundary — What It Does and Doesn't Cover

Microsoft's EU Data Boundary stores EU customer data within the EU for most M365 services. However:

It is an opt-in programme, not the default
Certain support and diagnostic data still flows outside the EU
It does not override CLOUD Act obligations — EU Data Boundary data can still be compelled under US law
The boundary applies at rest and in transit, but the AI processing (GPT-4 inference) occurs on Azure infrastructure that may not be fully boundary-constrained depending on load

What M365 Copilot Does Right

M365 Copilot Enterprise does not use customer data to train foundation models. Microsoft's data processing agreement for enterprise customers is among the most mature in the AI industry. For organisations already operating within Microsoft's ecosystem with appropriate DPAs in place, the incremental risk of M365 Copilot is lower than, say, using ChatGPT or Claude without enterprise agreements. The issue is not recklessness — it is structural cloud exposure that no enterprise agreement can fully eliminate.

Sector Risk Analysis

Where Copilot's cloud architecture creates specific compliance and legal exposure.

💻

Software Engineering

Proprietary algorithms, internal APIs, and unreleased features transmit to Azure continuously during GitHub Copilot use.

IP risk: Developer employment agreements typically include IP assignment clauses. Transmitting proprietary code to a third-party service may create ownership ambiguity under some jurisdictions. Client NDAs almost universally prohibit disclosure of their systems to third parties.

Local alternative: Ollama + Codestral or Llama 3.3 70B via Continue.dev VS Code extension.

⚖️

Legal

M365 Copilot processing legal documents through Microsoft Graph creates third-party disclosure of privileged communications.

Privilege risk: When M365 Copilot reads a privileged email or document as context for a query, that content is transmitted to Azure OpenAI. The voluntary disclosure doctrine may apply in some jurisdictions even when the disclosure is to an AI system operated by a third party.

Local alternative: Mistral Large 2 with local document indexing via LlamaIndex.

🏥

Healthcare

M365 Copilot accessing SharePoint or Teams where PHI is stored can inadvertently include patient data in prompts sent to Azure.

HIPAA risk: Microsoft offers a BAA for M365 that covers Copilot in enterprise configurations. However, the BAA requires careful scoping — Copilot must be configured to exclude PHI-containing data sources. Default configurations may not enforce this boundary.

Local alternative: Llama 3.3 70B on-premises with PHI-isolated document stores.

📈

Finance

Financial analysts using M365 Copilot on Excel models or Teams calls containing MNPI transmit material non-public information to Microsoft's infrastructure.

MNPI risk: Microsoft is not a broker-dealer or regulated financial entity. The chain of custody for MNPI processed through M365 Copilot is not clearly established. Compliance teams at regulated firms should treat any Azure-processed financial content as potentially problematic under SEC guidance.

Local alternative: Mistral Large 2 on an isolated finance VLAN with no internet egress.

Local Alternatives to GitHub Copilot

These models run locally via Ollama and integrate with VS Code, JetBrains, and Neovim through the Continue.dev extension.

Code Specialist

Codestral 22B

Mistral's dedicated code model. Trained on 80+ programming languages. Fast inference, strong completion quality.

ollama pull codestral

Best Overall

Llama 3.3 70B

Excellent general coding plus reasoning. Strong on architecture decisions and code review, not just completion.

ollama pull llama3.3

16GB RAM

Llama 3.1 8B

Lightweight, fast. Runs on developer laptops. Good for boilerplate, refactoring, and inline completion.

ollama pull llama3.1

8GB RAM

Mistral 7B

Smallest capable model. Suitable for code completion and explanation on any modern laptop.

ollama pull mistral

Continue.dev — Local Copilot in VS Code

Install the Continue extension in VS Code, point it at your local Ollama endpoint (http://localhost:11434), and you have a GitHub Copilot-equivalent that never contacts Microsoft. Supports tab completion, inline chat, and codebase-aware Q&A — all running on your machine.

Cost Comparison

50-person engineering team. GitHub Copilot Business + M365 Copilot combined vs local AI stack.

Local AI StackOllama + Llama 3.3 70B, 2× A100 80GB, shared inference server
Hardware (amortised 3yr)$833/mo
Power$80/mo
Ops / maintenance$200/mo
Per-seat cost (50 devs)$0
Total / month~$1,113

Microsoft Copilot Stack

50 users, GitHub Copilot Business + M365 Copilot

GitHub Copilot Business (50×$19)$950/mo

M365 Copilot (50×$30)$1,500/mo

M365 E3 base (if not owned)$1,800/mo

Hardware / ops$0

Total / month~$4,250+

3-Year Total Cost

Local AI stack: ~$40,000 (hardware + ops). Microsoft Copilot stack: ~$153,000 (subscriptions only, excluding M365 E3 if already licensed). That is a ~$113,000 difference over three years for a 50-person team — before accounting for compliance overhead, DPA management, and legal review costs that the cloud approach requires.

When to Use Each

The right answer depends on your data sensitivity and existing Microsoft investment.

Use Local AI When

Working with proprietary algorithms or unreleased product code
Operating under NDA with clients who prohibit third-party disclosure
GDPR, HIPAA, ITAR, or financial regulation applies to your data
You need fine-tuned models on your specific codebase
Team size makes per-seat licensing prohibitive
You need offline or air-gapped operation
You want zero CLOUD Act exposure

Copilot Remains Useful For

Teams deeply invested in the M365 ecosystem
Non-sensitive code on public or open-source projects
Developers without local GPU hardware
M365 document workflows where AI integration is seamless
Organisations with mature Microsoft DPAs already in place
Teams that need Copilot's polished IDE UX without setup friction

Replace Copilot in 15 Minutes

Install Ollama, pull Codestral, connect Continue.dev. Private code AI running before your next commit.

Quick Start Guide → vs ChatGPT