Im Überblick
- The customer contract sets the first limit: confidentiality, subcontractor and location clauses can each bar sending code to a cloud AI service.
- Armenian copyright protects only code with a human author; output produced by technical means without human creative activity is excluded.
- Employees’ code belongs to the employer by law unless the contract says otherwise; contractors’ code needs a written assignment.
- Personal data sent to an AI provider abroad needs a lawful basis and, outside the Personal Data Protection Agency’s adequacy route, the Agency’s prior permission (decision within 30 days).
- Procurement teams ask for documents: a tools policy, a per-customer matrix, training records and account settings.
Usually yes, within limits the customer contract sets before Armenian law comes into it: if the contract treats source code as confidential, controls subcontractors or fixes where work may be done, an AI coding assistant that sends prompts to a provider’s servers can breach it on the first request. Armenian law then adds two layers, copyright rules that decide whether AI-assisted code can be owned and assigned at all, and personal-data rules that apply the moment customer data reaches the tool.
What customer contracts say about AI coding assistants
Most development agreements signed before 2023 do not mention AI. They still govern it through four ordinary clauses.
Vertraulichkeit
A prompt containing the customer’s code goes to the AI provider, a third party. If the code is confidential information under the contract, that is a disclosure. Armenian civil law points the same way: a contractor under a work contract may not disclose technical information received from the customer without the customer’s consent, and information kept confidential for its commercial value is protected as a trade secret. Whether a particular upload breaches these rules depends on what was sent and on the provider’s terms.
Express AI clauses and supplier policies
Large buyers increasingly write AI rules into supplier programmes. Three published examples show the range:
- Intel’s Generative AI supplier policy (July 2023) prohibits entering Intel confidential information, including software code, and personal data into generative AI tools.
- Cisco’s Supplier AI Policy (April 2025) requires Cisco’s permission before its protected data is used to train, retrain or improve an AI model; a permission normally runs for one year.
- Microsoft’s Supplier Data Protection Requirements, version 12 (March 2026), require prior notice of subcontractors, approval of international transfers of Microsoft personal data, and add a separate section for suppliers that use or provide AI systems.
These are customer requirements and bind only suppliers who sign up to them. They are a reliable preview of the questions your own customers will ask.
Subunternehmer
Whether an AI provider counts as a subcontractor depends on how the contract defines the term. Where it does, flow-down clauses usually require the provider to accept terms at least as protective as yours, and the large AI vendors rarely sign bespoke terms. Read the definition before you assume the tool sits outside it.
Permitted locations
A clause permitting development in Armenia covers where your engineers sit. The AI provider processes and may store prompts in its own data centres, often in the United States or the EU. Those are two separate questions, and the second one is the one customers’ security teams now ask about.
Permission to use a tool and permission to put customer information into it are also separate. A vendor’s promise that it does not train on prompts is a term of your licence with the vendor. The confidentiality clause owed to your customer is satisfied by that customer’s consent or by a contract term permitting the tool.
Building a permissions matrix per customer
Each customer contract produces a different answer, so the working tool is one row per customer, filled in from the contract and any supplier policy it incorporates. A minimal version:
| Kolonne | Was aufgezeichnet werden soll | Woher kommt es |
|---|---|---|
| Tools allowed | Named tools and account type (enterprise only, none, any approved) | AI clause, supplier policy, written approval |
| Inputs allowed | Customer code, documentation, test data, personal data: yes or no for each | Confidentiality and data clauses |
| Training and retention | Whether training opt-out and zero or limited retention are required | AI clause, data protection addendum |
| Standort | Where the provider may process and store prompts | Location and transfer clauses |
| Approval or notice | Who at the customer must approve, and how far in advance | Subcontractor clause |
| Evidence owed | Attestations, logs, audit rights | Audit and reporting clauses |
Review a row whenever you add a tool, the provider changes its model or its own subprocessors, or the team starts using AI for a new task such as test-data generation. Each of these can fall outside an approval given for the original setup.
The approved-tools policy: what to put in it
The matrix answers what each customer allows. The policy tells engineers how to work inside it:
- Tool tiers: approved enterprise accounts, tools allowed only for internal or open-source work, and banned tools, including personal accounts on any service.
- Datenklassen: what may never go into a prompt (customer personal data, credentials, keys, code marked confidential) and what may, per the customer’s row in the matrix.
- Account Einstellungen: training opt-out, retention period and region, configured centrally so individual engineers cannot change them.
- Genehmigungsverfahren: who adds a tool and who checks the matrix before a project starts.
- Review of output: code review and licence scanning for AI suggestions, since an assistant can reproduce third-party code.
- Vorfälle: what an engineer does after pasting the wrong thing, and who decides whether the customer or the authorities must be told.
To bind staff, adopt the policy as an internal employer document and have each engineer acknowledge it in writing; our Einhaltung der Beschäftigungsvorschriften work covers how internal rules are introduced. Where engineers are engaged through an employer of record, the EOR is their legal employer, so the policy has to reach them through the EOR-Vereinbarung and its contract with you.
Who owns AI-assisted output?
Under the Armenian Copyright Law (HO-142-N) the author is the natural person who created the work, and a computer program is protected if it is the original result of the author’s creative work. The law expressly excludes results obtained with technical means without human creative activity. Code generated entirely by an AI tool therefore has no copyright to own or assign, while code an engineer writes using AI as a tool can be protected. How much prompting, selection or editing turns AI output into protected code has not been tested: no Armenian court decision or official guidance on AI-generated works has been found.
For protected code, the usual chain of title applies:
- Angestellte: economic rights in work created within employment duties belong to the employer unless the contract provides otherwise. Moral rights stay with the engineer and cannot be transferred.
- Auftragnehmer: nothing passes by law. You need a written assignment. Without clear wording, a licence under the Civil Code is non-exclusive, covers Armenia only, and runs for five years where no term is stated.
- Vergütung: a clause in which the author waives the right to remuneration is void.
A customer’s IP-assignment clause can transfer only what your company holds. If a deliverable contains machine-generated code, or code an assistant copied from a third-party project, the warranty of title in your contract may be wrong even though every engineer signed the right papers. Our software IP ownership audit sets out how to check the chain from engineer to company.
Personal data: what Armenian law adds
If prompts contain customer personal data (user records, logs with names or emails, production database extracts used as test data), the Law on Personal Data Protection (HO-49-N) applies to the upload.
Your role and the mandate
An Armenian company processing data on a customer’s instructions is normally an “authorised person”, close to a processor under the GDPR. The customer’s mandate must be in writing and state the legal basis, purpose, categories of data and data subjects, possible recipients and the protection measures. If the mandate does not comply with the law, the company must refuse and warn the customer in writing. Uploading data to an AI tool the mandate does not mention takes the processing outside it.
Lawful basis and minimisation
Processing needs the data subject’s consent, a deemed-consent case listed in the law, or an express statutory ground. Armenian law has no general legitimate-interests ground, and data may not be used for a purpose different from the one it was collected for. If the task can be done with anonymised data, processing identifiable data is prohibited; deleting names alone may not anonymise a record that still identifies someone indirectly.
Transfers to an AI provider abroad
Sending personal data to servers outside Armenia is a cross-border transfer. It needs no permission where the destination provides adequate protection, through an international treaty or by appearing on the Personal Data Protection Agency’s list of adequate countries. Otherwise the transfer may proceed only with the Agency’s prior permission, under a contract with safeguards the Agency approves. The application is made in writing before the transfer and names the country, the recipient, the data and the purpose; the Agency decides within 30 days. The data subject’s consent does not replace this permission, and an AI provider’s standard data processing addendum is not an Agency approval.
The Agency approved its adequacy list in December 2025. Which countries it covers, and in particular how it treats the United States, where most AI coding assistants process prompts, is being confirmed against the Agency’s published decision, so check the provider’s processing region against the current list before relying on it.
Security and incidents
The law requires you to destroy or block data you no longer need, protect data with encryption, restrict access to authorised users and keep data confidential during and after employment. Where data leaks from an electronic system, the processor must announce it publicly at once and report it to the police and the Agency at the same time; there is no 72-hour window and no risk threshold. Unlawful processing that is discovered must be cured within 3 working days, or the data destroyed, and the data subject informed. Whether an unauthorised paste into an AI tool counts as a leak or as unlawful processing has not been decided.
The Law on Cybersecurity (HO-442-N), in force since 4 January 2026, places duties on operators in vital sectors, including IT and digital infrastructure, and generally excludes micro and small companies unless they run critical infrastructure. It contains no general ban on sending code to AI services. If your customers’ data includes EU residents’ data, the GDPR may apply alongside Armenian law.
Evidence for procurement responses
A security team will ask for documents. The set that answers most AI questions:
- the approved-tools policy and a register of tools in use, with account type and settings;
- the customer’s row in the permissions matrix, and any written approvals it relies on;
- staff training and policy acknowledgement records;
- the AI providers’ contract terms on training, retention and processing location, with any Agency permission for transfers;
- a data-flow diagram showing what leaves your network and where it goes;
- your subcontractor list and incident log;
- an ISO 27001 certificate, where you hold one. Microsoft, for example, accepts a valid ISO 27001 certificate in place of its security section only; its data-handling and AI requirements still apply.
The wider questionnaire is covered in our security questionnaire checklist for Armenian suppliers. Build the matrix row for each customer from its contract first; switch on tools for that customer’s project only once the row is filled in and any approval it requires is on file.
Häufig gestellte Fragen
Can contractors use AI coding tools on client code?
Does a “no training” setting satisfy our confidentiality clause?
Who owns code an AI assistant wrote?
Do we need permission to send personal data to an AI provider outside Armenia?
Is an AI provider a subcontractor under our customer contract?
Does Armenia’s Cybersecurity Law ban uploading code to AI services?
What should an AI use clause in a software development agreement cover?
Letzte Aktualisierung: 13. Oktober 2026
