A development agency's AI usage policy should be agreed before the first commit. It should answer five questions in writing:
Which AI tools engineers may use
Which client code and data may enter them
How long each tool provider retains that data
Who reviews and owns AI-generated code
How AI use is disclosed to the client
If those answers arrive after kickoff, you are negotiating them while your codebase is already passing through tools you never approved. 724SOFTWARE is a selected Anthropic partner in Vietnam, and its engineers use Claude Code as part of normal delivery work. Every engagement already includes security and NDA setup during onboarding, which is the natural point to settle these terms.
TL;DR
A client-facing AI usage policy is different from an internal AI acceptable use policy. You need both.
Name every approved tool, its account type, and the data tier it may touch.
Record each tool's retention setting, including Claude Code data retention, and the date it was checked.
Require human review on every AI-assisted change, because AI-only output weakens IP ownership.
Agree how AI use is labelled in commits before sprint one.
About the Author: Written by the 724SOFTWARE delivery team. The company has 200+ professionals (58% senior level) who run dedicated teams for clients in 10+ countries and use Claude, Cursor, Gemini and NotebookLM in daily delivery.
What is an AI usage policy for a development agency?
An AI usage policy for a development agency is the written agreement that governs how generative AI touches one client's code, data and deliverables. Most agencies already have an AI acceptable use policy for staff. That internal rulebook does not tell a client what happens to their repository.
Document | Audience | Answers
|
|---|---|---|
AI acceptable use policy | Agency employees | What staff may do with AI tools in general |
AI usage policy (client-facing) | Client and agency | Which tools touch this client's assets, under what controls |
Contract / SOW | Legal signatories | Liability, IP assignment, confidentiality |
Two frameworks are used for structuring AI governance:
ISO/IEC 42001 defines a certifiable AI management system.
The NIST AI Risk Management Framework is voluntary and is organised around four functions: Govern, Map, Measure and Manage.
Neither is required for a single client policy. Borrowing their headings still keeps the document auditable.
Which AI tools and data should the policy name?
Once the purpose is clear, the first concrete section is a tool register. AI assisted software development usually involves several tools with different jobs. A blanket "AI is allowed" clause hides exactly the differences that matter.
A practical register lists, for each tool:
Tool and account type: for example, Claude Code on a business account, Cursor, or NotebookLM.
Approved use: code generation, review, test scaffolding, or spec summarisation.
Data tier allowed: public, internal, confidential, or regulated.
Retention and training settings: as currently configured.
Owner: the named person who confirms those settings stay unchanged.
Pair the register with a default rule on data:
Source code and internal documents may enter approved business-tier tools.
Production data, credentials and personal data may enter no AI tool without the client's written sign-off.
Why does Claude Code data retention belong in the contract?
Building on the tool register, retention is the column clients most often leave blank. It is also the column that decides where copies of their code live.
A prompt behaves more like an email attachment than a keystroke. Once sent, a copy sits on the provider's systems under the provider's retention rules, not yours. Your repository's access controls stop applying at that moment.
That is why Claude Code data retention, and the equivalent setting for every other tool, should be written down rather than assumed. Terms can differ between consumer and commercial account types, and providers update them. The policy should record:
Which account type engineers use (personal accounts are not permitted)
Retention and training-use settings as configured
The date someone last checked the provider's terms
What happens to prompt history when an engineer leaves
Our opinion: a quarterly re-check is a reasonable minimum. A policy citing last year's vendor terms gives false comfort.
Who owns and reviews AI-generated code?
A related but distinct question is ownership. This is where AI use stops being a security topic and becomes a legal one.
Under US law, code generated solely by AI cannot be copyrighted because it lacks human authorship.
Commercial teams also face infringement risk from AI training data.
Confidential code fed into public models can lose trade-secret protection.
The operational answer is a review gate, not a ban:
A named human reviewer approves every AI-assisted change before merge.
Reviewers remove snippets of unclear origin.
The contract's IP assignment clause explicitly covers AI-assisted work.
If the product ships AI features, the team tests against the OWASP Top 10 for LLMs, which covers risks such as prompt injection and excessive agency.
Human review improves quality. It also strengthens the human-authorship record behind the code.
How should AI use be disclosed and logged?
With ownership settled, the remaining gap is evidence: how a client later proves what was AI-assisted. For an agency, labelling must be light enough that engineers actually do it.
Method | Effort | What it proves
|
|---|---|---|
Commit trailer or PR label | Low | Which changes involved AI |
Tool named in PR description | Low | Which retention terms apply |
Sprint report summary | Medium | Usage patterns across the team |
Already happening: the EU AI Act requires technical documentation, logging and human oversight for high-risk AI systems.
Likely: European clients will ask agencies for similar records even for lower-risk products. Building the habit now costs less than retrofitting it later.
What should an AI usage policy template include?
Pulling these sections together, a usable AI usage policy template fits on two pages and is signed before sprint one:
Scope and definitions of "AI-assisted"
Approved tool register with account types
Data tiers and prohibited inputs
Retention settings with check dates
Human review and merge rules
IP assignment for AI-assisted work
Disclosure method
Incident procedure if restricted data enters a tool
Review cadence
Do not skip item 8. Treat a prompt containing credentials like any leaked secret: rotate it, log it, and notify the client.
Frequently Asked Questions
Is an internal AI acceptable use policy enough for clients?
No. It governs staff behaviour but does not commit the agency to specific tools, settings or disclosure for your engagement.
Should AI tools be banned on sensitive projects?
Usually not. Restrict data tiers and require business accounts instead. Bans tend to push use onto unmonitored personal accounts.
Who owns the policy?
A named delivery lead on the agency side, with sign-off from the client's CTO or security owner.
Does AI-generated code belong to the client?
Only what the contract assigns, and only what carries human authorship. Make the assignment explicit.
How often should the policy be reviewed?
At least quarterly, and whenever a tool or its terms change.
About 724SOFTWARE
724SOFTWARE is a Vietnam-based technology company providing dedicated teams, embedded staff augmentation and offshore development centres for web and mobile application development. As a selected Anthropic partner in Vietnam and an official Cursor partner, it trains engineers to use Claude Code in normal delivery work. Its processes are aligned with ISO 27001:2022 and ISO 9001, it is GDPR compliant, and it scales teams from 1 to 50+ pre-vetted engineers in 2 to 4 weeks with a <10 minute, 24/7 incident response commitment.
Before your next engagement starts, ask your delivery partner to complete the tool register and retention column above. To see how 724SOFTWARE answers it, get in touch at https://724software.com.vn/.
