Enterprise AI coding assistants with permission and privacy controls

Enterprise AI coding assistants address data privacy, no-training terms , and repository-level permission controls by pairing strict data-handling policies with code intelligence layers.sourcegraph.com+1

When processing developer prompts and indexing documentation or codebases, the following platforms enforce explicit zero-training and permission inheritance controls for enterprise tiers:

1. Sourcegraph Cody (Enterprise)
   • Repository Permissions: Relies on Sourcegraph's native code intelligence platform, ensuring that contextual code retrieval strictly respects the user's underlying read permissions and fine-grained access controls. It also uses Context Filters to limit which repositories Cody can read or index.
   • No-Training & Retention: Explicitly guarantees that customer code, prompts, and completions are never logged for model training by Sourcegraph or third-party LLM partners (with zero-retention policies applied at generation time).sourcegraph.com+2
2. GitHub Copilot (Business & Enterprise)
   • Repository Permissions: Inherits organization and repository-level access rights managed through GitHub's core platform boundaries. Enterprise tiers add deeper policy management and compliance controls.
   • No-Training & Retention: Enforces a strict no-training policy where proprietary code, snippets, and prompts are processed in real time and never used to train foundation models. Prompt/code retention for IDE interactions is wiped or not permanently retained under enterprise privacy modes.briangershon.com+2
3. Cursor (Business & Enterprise)
   • Repository Permissions: Integrates codebase indexing with workspace-level administrative controls, audit logs, and tracking APIs.
   • No-Training & Retention: Requires Privacy Mode (which is enabled by default for teams and enterprise plans). Under Privacy Mode, Cursor enforces zero data retention (ZDR) agreements with underlying model providers (OpenAI, Anthropic, etc.), ensuring data is not stored or used for model training.cursor.com+2
4. GitLab Duo (Enterprise / Dedicated)
   • Repository Permissions: Operates within GitLab’s native project- and group-level permission hierarchy, ensuring code suggestions and chat features only process code accessible to the authenticated user's authorization level.
   • No-Training & Retention: Implements strict zero data retention (ZDR) agreements with its AI sub-processors (such as Anthropic and Google), completely barring model training on customer inputs/outputs.docs.gitlab.com+1

If you are evaluating these tools for a specific environment, let me know:

• Your primary git host/version control platform (GitHub, GitLab, self-hosted, etc.)
• Whether you prefer a cloud-hosted or self-hosted/air-gapped LLM deployment

I can help narrow down the best fit for your compliance and security requirements.

Enterprise AI coding assistants: privacy, retention, and repository access

As of October 2026, the strongest candidates to evaluate are Sourcegraph Cody Enterprise, Cursor Enterprise, GitHub Copilot Enterprise, and Amazon Q Developer Pro. They offer different combinations of repository-level authorization, no-training commitments, data-retention controls, and enterprise administration.

The key distinction is that “we don't train on your code” is not the same as “we never retain your code or prompts.” If your requirement covers source code, documentation, retrieved repository context, and developer prompts, assess all four data types across every client, agent, and model provider.

Shortlist

• 

1. Sourcegraph Cody Enterprise

(Strongest fit for permission-aware code retrieval)

Sourcegraph documents that Cody enforces repository permissions during code retrieval: it retrieves only snippets the user has permission to read. It also states that Enterprise customer data is not used for training and describes zero-retention arrangements with its model partners.sourcegraph.com+6

• Repository permissions: Explicitly enforced during retrieval.
• No training: Enterprise company data is not used to train models.
• Retention: Model-provider inputs and outputs are covered by zero-retention commitments under the documented Model Provider terms.
• Validate: Retention of service-side logs, indexing data, documentation, and prompts under your specific deployment and contract.

Sources: Cody privacy FAQ ·Model Provider terms

---

• 

2. Cursor Enterprise

(Strong configurable privacy controls)

Cursor enables Privacy Mode by default for Enterprise teams and supports organization-wide enforcement. It documents zero-data-retention agreements with most inference providers, plus controls for model access, agent permissions, and repository blocklists.prod.cursor.com+6

• Repository controls: Enterprise repository blocklists and agent access controls.
• No training: Privacy Mode prevents Cursor and its covered model providers from training on customer code and prompts.
• Retention: Most models use zero-retention arrangements, but some models have separate retention requirements and require admin approval.
• Important caveat: Cloud Agents temporarily store encrypted repository copies while working. This is not a universal no-storage architecture.

Sources: Privacy and data governance ·Enterprise controls

---

• 

3. GitHub Copilot Enterprise

Strong fit for organizations already using GitHub.

GitHub provides enterprise policy controls for agents, approved tools, and repository access. Its published retention information distinguishes IDE chat and completions from other access methods: IDE prompts and suggestions are not retained by default, whereas other uses can retain them for 28 days.github.com+6

• Repository permissions: GitHub repository authorization and enterprise-managed agent controls; verify the behavior of each agent and integration.
• No training: Review the current Copilot Business/Enterprise contractual terms for the specific features enabled.
• Retention: Not uniform across surfaces; CLI, web, mobile, and agent workflows need separate review.
• Important caveat: Do not interpret the IDE retention policy as a blanket zero-retention guarantee for every Copilot feature.

Sources: Enterprise agent management ·Copilot privacy and retention

---

• 

4. Amazon Q Developer Pro

Good fit for AWS-centric enterprises.

AWS explicitly states that content from Q Developer Pro is not used to train underlying foundation models. AWS also documents encryption and access controls for stored content.aws.amazon.com+6

• Repository permissions: Evaluate access behavior for your connected repositories and IDE/agent workflows.
• No training: Explicit Pro-tier commitment.
• Retention: AWS stores some content securely to operate the service; no-training does not establish zero retention.
• Important caveat: Confirm the precise retention period and whether all code, documentation, and prompt paths are covered by your intended contract.

Source:Amazon Q Developer FAQs

How they compare against your requirements

This matrix distinguishes documented capabilities from items that still need contractual verification. A checkmark does not mean every product surface or integration has identical behavior.

Sources:sourcegraph.com+6

My recommendation

• For strict repository permission enforcement: Start with Sourcegraph Cody Enterprise. Its documentation directly describes permission checks during code retrieval.
• For enforceable privacy settings and granular agent restrictions: Evaluate Cursor Enterprise, paying particular attention to cloud-agent storage and model exceptions.
• For a GitHub-native governance strategy: Evaluate Copilot Enterprise, but map retention by product surface rather than relying on one general privacy statement.
• For AWS-centered development: Evaluate Amazon Q Developer Pro, with a separate review of retention and repository authorization.

What to require in the contract and security review

Before selecting a vendor, ask for written answers to these five questions:

1. Scope of no-training terms: Do they cover source code, repository documentation, retrieved snippets, system prompts, developer prompts, chat history, generated output, and agent tool results?
2. Retention: What is retained by the vendor, its subprocessors, and model providers? Specify duration, logs, caches, backups, abuse monitoring, and deletion timelines.
3. Permission enforcement: Are repository and file-level permissions rechecked when context is retrieved, including for indexing, embeddings, agents, and cross-repository search?
4. Admin enforcement: Can administrators prevent users from disabling privacy protections, selecting noncompliant models, using personal API keys, or enabling unapproved agents?
5. Exceptions and evidence: Are all commitments contractually binding, and can the vendor provide a DPA, subprocessor list, retention schedule, and security documentation?

Bottom line: Sourcegraph Cody Enterprise is the clearest starting point when permission-aware repository retrieval is the primary requirement. Cursor Enterprise is a strong alternative when configurable, enforceable privacy and agent controls are central. Neither no-training claims nor zero-retention commitments alone establish that every form of documentation, prompt, log, and agent artifact is never stored.

For an enterprise AI coding assistant, I’d evaluate vendors against two separate requirements: zero use of your private data for model training and enforcement of your existing repository access controls. These are distinct security guarantees, and a vendor satisfying one does not automatically satisfy the other.

What your enterprise solution should guarantee

1. No training on your private data

Your source code, prompts, chat history, code completions, and documentation must not be used to train or improve shared models. Look for an explicit contractual commitment covering all data types, including data processed by subprocessors.

---

2. Repository permissions are enforced

The assistant must respect your GitHub, GitLab, Bitbucket, or other source-control permissions. A developer who cannot access a private repository or file must not be able to retrieve its contents through AI chat, code search, indexing, or agent tools.

---

3. Enterprise-grade data governance

Require SSO, SCIM provisioning, audit logs, configurable retention, encryption, contractual data-processing terms, and clear policies for model-provider subprocessors.

Shortlist: the strongest candidates

These three vendors are worth evaluating first, based on their published enterprise privacy and access-control documentation.

• 

1. Sourcegraph Cody Enterprise

(Strongest documented permission-aware retrieval)

Sourcegraph explicitly documents that Enterprise customer data is not used for training and that its code retrieval process enforces users' read permissions before sending relevant code to the language model. This makes it a particularly strong candidate when fine-grained repository access is a primary concern.sourcegraph.com+1

• Training: No training on company data, according to its Enterprise FAQ.
• Permissions: Explicitly permission-aware code retrieval.
• Evaluate: Whether permissions remain consistent across indexing, chat, IDE integrations, and agent workflows.

Official documentation: Cody privacy and permissions FAQ ·Enterprise features

---

• 

2. GitHub Copilot Enterprise

(Strong choice for GitHub-centric organizations)

GitHub states that Copilot Business and Enterprise customer data is not used to train its models, with contractual protection under its Data Protection Agreement. It is a natural candidate when your organization already manages repositories and developer identities through GitHub.docs.github.com+1

• Training: No model training on Business or Enterprise customer data.
• Permissions: Evaluate repository access, organization policies, and the behavior of coding agents against your existing GitHub permissions.
• Evaluate: Whether access restrictions cover every connected tool and agent action, not just ordinary IDE suggestions.

Official documentation: Copilot features and data protection ·GitHub Copilot documentation

---

• 

3. Cursor Enterprise

(Strong choice for AI-first development workflows)

Cursor documents an organization-enforceable Privacy Mode, zero-data-retention arrangements with model providers, and enterprise controls for restricting repositories, models, and agent access. Its Protected Git Scopes feature adds organization-level restrictions for cloud agents and related services.cursor.com+2

• Training: No training on your data when Privacy Mode is enabled.
• Permissions: Repository blocklists, per-user repository checks, and Protected Git Scopes.
• Evaluate: Cloud agent permissions, approved model retention policies, and any use of personal API keys.

Official documentation: Privacy and data governance ·Repository and model access controls

How I would compare them

These are based on vendor documentation, not an independent security audit. The repository-control mechanisms are not identical, so treat this as a shortlist rather than a definitive certification of compliance.

The security questions I would require vendors to answer

Before signing a contract, ask each vendor to demonstrate these scenarios in a proof of concept.

Proof-of-concept checklist

of

My recommendation

Start with Sourcegraph Cody Enterprise if enforcing fine-grained access to code across repositories is your highest-priority requirement. Its published documentation directly describes permission checks during code retrieval. Include GitHub Copilot Enterprise and Cursor Enterprise in the evaluation if your existing development platform or preferred agent workflow makes them a better fit.

One important distinction: not training on your data does not necessarily mean never storing your data. If your policy also requires zero retention, private-network deployment, or data residency, assess those requirements separately.

One question to narrow this down

Which source-control environment do you use?

1. Source control

GitHub Enterprise

GitLab

Bitbucket

Multiple platforms

---

2. Deployment requirements

SaaS with enterprise controls

Private cloud or VPC

Self-hosted or on-premises

---

3. Permission enforcement

Enforce permissions at every retrieval and agent action

Repository-level allowlists and blocklists are sufficient

[Refine vendor recommendation]

When evaluating enterprise AI coding assistants, your two requirements—strict zero-training data policies and enforcement of existing repository-level permission structures —are the core gating items.

Leading enterprise-tier tools meet these requirements, though they do so by anchoring into your existing git host or cloud IAM architecture.

Top Enterprise AI Coding Assistants

1. GitHub Copilot Enterprise / Business
   • Training Policy: Strict No-Training. Prompts, code snippets, and telemetry data for GitHub Copilot Business and Enterprise are processed in real-time and are never logged, retained, or used to train or improve GitHub's or third-party foundational models.
   • Repository Permissions: Inherits directly from your GitHub organization and repository access controls. Copilot Chat and codebase indexing only search or reference repositories that the authenticated user explicitly has read access to within that GitHub enterprise boundary.
   • Best For: Teams already living inside the GitHub ecosystem.github.com
2. GitLab Duo (Enterprise / Ultimate Tiers)
   • Training Policy: Strict No-Training. GitLab explicitly contractually guarantees that customer data, prompts, code, and chat outputs are never used to train foundational or multi-customer models (leveraging enterprise instances of models from partners like Anthropic and Google).
   • Repository Permissions: Deeply honors GitLab’s native project and group-level permission inheritance (Guest, Reporter, Developer, Maintainer, Owner). Features like codebase context and chat interactions respect the precise visibility and branch-protection settings of the user.
   • Best For: Organizations standardizing their SDLC on GitLab.forum.gitlab.com+1
3. Amazon Q Developer (Pro / Enterprise)
   • Training Policy: Strict No-Training. AWS guarantees that customer content (code, logs, prompts, generated outputs) sent to Amazon Q Developer is not used to train or fine-tune AWS foundational models.
   • Repository Permissions: Integrated tightly with AWS IAM (Identity and Access Management) . Access to specific repositories, codebase indexing, and specialized actions can be explicitly bounded and constrained via IAM policies and resource-based permissions.
   • Best For: Enterprise environments deeply integrated into AWS infrastructure and IDEs (like VS Code or JetBrains via the AWS Toolkit).docs.aws.amazon.com+1
4. Sourcegraph Cody (Enterprise)
   • Training Policy: Strict No-Training. Cody offers explicit enterprise privacy guarantees, ensuring your codebase context and chat logs are isolated and never fed back into public model training data. It can also be deployed with local/private LLMs (via Ollama, vLLM, or enterprise-hosted instances).
   • Repository Permissions: Cody syncs repository permissions directly from your code host (GitHub, GitLab, Bitbucket, or Perforce). It builds its internal vector search index strictly respecting who is allowed to view what repository; if a developer doesn't have read access to a repo in the source host, Cody's context engine won't index or retrieve snippets from it for them.
   • Best For: Multi-host or large polyrepo enterprises wanting flexibility over LLM backend choices while strictly enforcing host-synced permissions.

Key Evaluation Checklist for Your Security Team

• Data Residency & Zero Retention: Verify whether prompts are stateless (zero retention after the API response is served) or if they feature temporary caching for performance optimization. Ensure zero retention is configured.briangershon.com
• Context Indexing Boundaries: Clarify how the assistant indexes your codebase. Does it index locally on the developer machine, or does it build a remote enterprise vector database? Ensure the remote vector index queries user identity tokens at runtime to filter out unauthorized repos.
• Audit Logging: Check if the platform logs AI usage events for compliance review without capturing raw proprietary block text that could violate internal policies.

To help narrow down the best fit, could you tell me:

• What git host/version control platform (GitHub, GitLab, Bitbucket, etc.) do you use?
• What IDEs does your engineering organization primarily use?
• Do you have a preference for cloud-hosted AI vs. self-hosted/private model routing?