Skip to main content

Legal

AI Use Policy

Codepixel is an AI-augmented software development agency. We use business-grade AI tools to help us deliver client work faster, with stronger consistency, and at lower risk. We publish this policy so that clients, prospects, and partners can see exactly how we work with AI — and what we won't do.

This policy applies to AI usage across all Codepixel client engagements and internal operations. For data processed through this website, see also our Privacy Policy.

1. Our approach in one paragraph

Codepixel uses business-grade AI tools to augment human engineering work, not replace it. Every AI-assisted output passes through human review before it reaches a client. We restrict ourselves to a small set of vetted Tier A tools that contractually do not train on customer data, and whose data handling we describe below. Clients always retain the option to require human-only delivery.

2. Approved tools

We use only the following business-grade AI tools for client work:

Vendor / ToolUsed forTier / Account type
Claude Team (Anthropic), in Codepixel's company organisation, including Claude CodeCode generation, code review, technical writing, drafting and researchTeam (business plan under Anthropic's Commercial Terms and DPA)

Claude Team runs under Anthropic's Commercial Terms of Service and Data Processing Addendum (DPA):

  • Anthropic does not train its models on our or our clients' inputs and outputs, and the feedback features that would share conversations with Anthropic are switched off for the whole organisation.
  • Chats and projects are kept until we delete them and are removed from Anthropic's systems within 30 days of deletion. Claude Code session data is kept for 30 days, and Claude Code keeps session history on the engineer's encrypted laptop for up to 365 days. Anthropic keeps data longer only where the law requires it or to enforce its usage policy.
  • Anthropic stores data in the United States; transfers are covered by the Standard Contractual Clauses in its DPA. Claude Team has no EU data residency option. If your engagement requires EU-only processing, we agree in writing before work starts which tools, if any, may be used.
  • Company accounts only. No personal AI accounts for company or client work, and no consumer AI plans for client work. Every member signs in with a company Google Workspace account and is invited to our Claude organisation individually.

AI features built into the business software we already use, such as Google Workspace and ClickUp, run on our company accounts with training on our data switched off.

We update this list when our tooling changes. If a tool moves out of Tier A status (for example, if a vendor changes its data-handling defaults), we stop using it for client work and remove it from this list.

3. What we do NOT do

The following are out-of-bounds for Codepixel's AI usage:

  • No consumer-grade AI for client work. Company accounts only: we never put client code or data into a personal, free or consumer account of any AI provider.
  • No secrets in prompts, and no production data without your consent. Passwords, API keys, database credentials and other secrets are never sent to AI tools. Your end users' personal data and your production data stay out as well, unless you give written consent. We test with synthetic or masked data. Our organisation settings also block our AI coding agent's file tools from reading secret files such as .env files.
  • No special-category data without explicit approval. GDPR Article 9 data (health, biometric, religious, political, etc.) is not sent to AI tools unless the client explicitly approves and we put appropriate safeguards in place (pseudonymisation, contracted DPA, vendor certifications).
  • No autonomous decisions on client systems. AI connectors and agents touch your systems only with your written agreement. They may propose changes; humans approve and execute them. There is no auto-merge or auto-deploy gated only by AI.
  • No model fine-tuning on client code without consent. We only use general-purpose models. We do not fine-tune or distill models on a client's codebase unless explicitly contracted to do so.

4. Mandatory human review gates

Every Codepixel deliverable passes through two AI-assisted review gates, with humans as the final authority:

  • Gate 1 — Per Pull Request. Before any PR can merge, it must pass:
    • Automated secrets scanning (we use tools that flag accidentally committed credentials).
    • AI-assisted security and correctness review (with human reviewer sign- off).
  • Gate 2 — Pre-Delivery. Before a deliverable ships to a client:
    • A codebase-level AI review surfaces broader correctness, architecture, security, and maintainability issues.
    • Findings are captured in a written Findings Register, then triaged and closed (fix, accept-with-rationale, or document for handoff) by a human reviewer who signs the register.

These gates are part of every engagement. They are described in our Security Statement at a high level.

5. Data handling guarantees

For client work performed with AI assistance, we guarantee:

  • No training on your data. Anthropic does not train its models on our or our clients' inputs and outputs, and the feedback features that would share conversations with Anthropic are switched off for the whole organisation.
  • Known retention. Chats and projects are kept until we delete them and are removed from Anthropic's systems within 30 days of deletion. Claude Code session data is kept for 30 days. Claude Code also keeps session history on the engineer's encrypted laptop for up to 365 days. When a project closes, we delete that client's Claude projects, chats and local session history if the contract or the client asks us to.
  • Known data location. Anthropic stores data in the United States; transfers are covered by the Standard Contractual Clauses in its DPA.
  • Synthetic or masked test data. If we need to test code against real-shaped data, we use synthetic or masked data. Real personal or production data goes into an AI tool only with your written consent.
  • Records you can review. A register of approved AI tools, their settings and owners, plus version control and peer review that show what changed and who approved it. On request, we walk you through the parts that cover your engagement.

6. Disclosure

Before work starts we tell you, in the contract or SOW, which AI tools we use on your engagement (Section 7). We do not label individual pull requests or documents, because every output passes the human review in Section 4. Where your contract asks us to flag AI use on a deliverable, we do.

7. Client controls

Before work starts we tell you which AI tools we use. You can opt out or restrict them. We ask for your written consent before personal data, production data or your restricted information goes into an AI tool. Your choices are recorded in the contract or SOW:

  • Opt out completely from AI assistance, in which case the engagement is delivered entirely by humans. Pricing and timeline may adjust.
  • Restrict to specific tools (e.g., "we approve Claude Code in the repository but not Cowork or the browser tools").
  • Restrict data types (e.g., "you may use AI for documentation but not on source code").
  • Give or refuse consent. Your written consent is required before personal data, production data or your restricted information goes into an AI tool. If you work in a regulated sector, or your signed contract requires prior consent, we also ask for it before using AI tools on your engagement at all.
  • Approve connections to your systems. AI connectors and agents touch your systems only with your written agreement.
  • Review our records. On request, we walk you through the records described in Section 5.

8. Compliance alignment

Our AI practices are designed to align with:

  • EU General Data Protection Regulation (GDPR) — as a controller for our own data and as a processor under our client SOWs. We do not use AI to make solely automated decisions with legal or similarly significant effects on data subjects (GDPR Art. 22).
  • EU AI Act (Regulation 2024/1689) — we operate as a "deployer" of general-purpose AI systems in low- and limited-risk contexts. General-purpose model obligations have applied since August 2025, and the high-risk obligations apply later under the amended timeline. Neither set changes our classification, and we review it whenever our tooling or use cases change.
  • Canada PIPEDA and BC PIPA — same human-review-gate and no- training-on-data commitments apply to Canadian client data.
  • Sectoral regulations — for engagements in regulated sectors (health, finance, public sector), we tighten controls and document them in the SOW.

9. Incident handling

If an AI tool contributes to a defect, data exposure, or security incident on a client engagement, our incident response covers it like any other root cause. We notify the affected client per the engagement SLA and our Security Statement. We do not hide AI-related causes in post-mortems.

10. Internal use of AI

Beyond client deliverables, Codepixel uses AI internally for:

  • Engineering productivity (code completion, documentation, refactoring).
  • Sales and marketing operations (drafting, research, content review).
  • Internal automation (the same review-gate principle applies).

Internally the same rules apply: company accounts only, and client data only in the tools listed in section 2.

11. Updates to this policy

This policy reflects our practice as of the effective date above. We update it as our tooling, contractual posture, and regulatory environment evolve. Material updates are reflected in the version number and a change- log entry below.

If you are a current client and an update materially affects how we work on your engagement, we will notify you per the SOW.

12. Contact

Questions, audit requests, or feedback on this policy:

  • Email: privacy@codepixel.me (subject: "AI Use Policy")
  • For client engagement-specific AI questions, use your usual Codepixel point of contact.

Change log

  • v1.4, November 2026. Section 6: clients learn which AI tools we use up front in the contract or SOW. We no longer label individual pull requests or documents, unless the contract asks for it.
  • v1.3 — October 2026. Added that Claude Code keeps session history on the engineer's encrypted laptop for up to 365 days and that we delete it with a client's projects and chats when asked. Organisation settings now block our AI coding agent's file tools from reading secret files.
  • v1.2 — October 2026. Company and client AI work now runs in a Claude Team organisation under Anthropic's Commercial Terms and DPA. Updated the approved tools, retention, data location and client-choice sections, added the records we keep (tool register, version control and peer review) and aligned the EU AI Act dates with the amended timeline. Cursor Business retired on 2026-10-07; Claude Code is our AI coding tool.
  • v1.1 — August 2026. Updated the EU AI Act wording to reflect that the staged enforcement obligations are now in force. No change to our tooling, gates, or practices.
  • v1.0 — Initial publication (May 2026).