Terms of Use: platform rules and responsibilities

Purpose

These Terms of Use set the rules for using the Platform. See the terms of sale for commercial terms.

Publisher

DUALE AI SAS, a French Simplified Joint Stock Company (SAS) with capital of €10,000

• Paris Trade Register: 994 521 128
• Registered office: 60 rue François 1er, 75008 Paris, France
• Publication Director: Clémence Lesné
• Email: contact+info@mail.duale.ai

Hosting:

• Website: Cloudflare, Inc., 101 Townsend Street, San Francisco, CA 94107, United States
• Application data: Hetzner Online GmbH, Industriestraße 25, 91710 Gunzenhausen, Germany

Definitions

Parties and service

• Platform: the AI Agent orchestration services Duale AI provides through its SDK and API, together with the web interface for administration and usage analytics
• Customer: legal entity or individual acting in a professional capacity, holding a contract with Duale AI
• User: individual authorized by a Customer to access the Platform
• Services: the Platform’s functionalities, including submitting, orchestrating, and managing Tasks

Agent work

• AI Agent: an algorithmic software system that pursues objectives on the Customer’s behalf by selecting or carrying out steps through the Platform; it is not the same object as an Agent Identity
• Agent Identity: a tenant-scoped non-human principal used for credentials, access, Tool availability, Task attribution, and history access; it does not define a versioned AI Agent behavior
• Task: one request accepted through the public SDK or API, with an objective, deadline, response contract, and one terminal outcome; each accepted continuation is another Task
• Executor: a durable runtime scope for a bounded sequence of work within one Task
• Tool: a named, typed capability through which an execution can request an operation. Tool Definition, model exposure, dispatch, receipt, execution, Tool Result, and Effect are separate events
• Customer Tool: a Tool that the Customer supplies and hosts
• MCP Connection: a named Customer configuration through which the Platform lists and calls Tools exposed by one remote MCP server, using the credentials and remote permissions configured for that Connection
• Remote MCP Tool: a Tool exposed by an MCP server the Customer selects and made available through an MCP Connection
• Built-in Tool: a named, typed capability that Duale AI supplies and hosts for eligible Executor roles
• Effect: a state transition that actually occurs in an environment
• External Effect: an Effect that occurs outside the Platform, including in a Customer or third-party system or communication channel

Model routing

• Model pool: the set of enabled Model Routes the Customer authorizes
• Provider key: the access the Customer supplies to its own model providers

Billing, evidence, and data

• Wallet: the balance in euros the Customer loads in advance
• Audit log: the record of configuration, access, routing dispatch, and Task lifecycle events listed in Scope and limits. It records Tool dispatch decisions by name, but not whether a Tool ran, succeeded, or what it returned
• Usage analytics: the execution statistics the Platform presents in its web interface
• Content: data, text, files submitted by the User
• Outputs: results generated by AI Agents

Acceptance

You accept these Terms by ticking a box when you create the account. Duale AI does not pre-tick that box. Mere browsing does not constitute acceptance.

Duale AI retains proof of your acceptance: timestamp, version of Terms accepted, IP address, and user action.

Platform Access

Access to the Platform is limited to Users whom a Customer with an active contract has authorized. Users must be at least 18 years old.

Your credentials are personal and confidential. You are responsible for all activity on your account.

Artificial Intelligence Services

Service Description

Through the Platform, Users submit, orchestrate, and manage AI Agent tasks that run on language models.

Data Processing and Enrichment

The Platform can:

• Store and index your documents to make them accessible to AI Agents
• Perform internet searches (public data only) and retrieve the pages they return, to enrich responses
• Run a program an AI Agent writes in fresh, bounded writable state separate from your systems, and return the reported execution output. The AI Agent chooses the network setting for each Tool Call. The Platform has no Customer-level network setting. Code Execution states the current request and result limits
• Connect to a Customer-selected remote MCP server, list the Tools it exposes, and call eligible Tools for assigned Agent Identities with the remote access configured on that MCP Connection

Internal models: Duale AI uses open-source models hosted on its European infrastructure to perform analysis and add intelligence to the Platform. This processing stays on Duale AI’s infrastructure.

Web-search features are intended for public data. Do not include personal data in web-search queries unless the Customer has a lawful basis and the applicable order covers the processing.

Limitations

AI Agents can:

• Make mistakes or “hallucinate” (invent information)
• Reproduce biases
• Vary in quality depending on complexity

How the Platform selects a model

The Customer declares the Model pool and supplies its own Provider keys. For each model turn, the Platform selects an eligible Model Route from that pool. One model-directed Runtime Route can make several Provider Attempts, and a later turn can select another eligible route. Duale AI makes Provider Attempts only through enabled routes in the Model pool.

Routing is not deterministic. Two identical tasks can run on different models, and a single Task can call several providers. Duale AI records the selected model in the Usage analytics.

The Customer sets these routing criteria: target accuracy, cost sensitivity, speed preference, priority, and required and preferred skills. The criteria are weightings. They steer the choice; they never forbid one.

Model skill labels describe capabilities measured on public benchmarks. They do not describe uses Duale AI recommends. Required and preferred skills are routing weights rather than hard filters. If the Model pool cannot satisfy the full skill request, routing relaxes those preferences and selects the best eligible route; if no route is eligible, the Task fails instead of calling an unauthorized fallback.

What the Platform does not do

This section explains current Platform behavior and limits. Contractual duties and prohibitions are grouped under Responsibilities and use rules below.

Customer Tool handling is not exactly once. A dispatched Customer Tool Call can fail or time out before customer code receives it. After a host consumes it, the same (task_id, tool_call_id) can reach your code twice or more. You deduplicate on your side. If a duplicated call moves money, sends a message, or writes to a third-party system, the correction is yours.

Duale AI does not stop a Task for you. A caller whose access policy permits it to stop the Task (Task actions) can request asynchronous Stop processing. Acceptance does not confirm that the Task stopped. A Stop that takes effect does not recall work already dispatched; it can still finish and produce an Effect. Scope and limits defines the downstream Stop set, continuation, deadline, and root-wide new-route behavior. Revoking a Provider key prevents new Tasks from starting but does not stop work already running.

Customer Tool delivery and settlement are asynchronous. You set a timeout. If it expires without a Tool Host response, the Platform supplies a timeout Tool Result and the AI Agent can continue on a best-effort basis. That result does not prove whether the Customer Tool code ran or produced an Effect.

Remote MCP Tool execution is not exactly once. Duale AI does not intentionally retry a logical call when it cannot rule out that the MCP server received the request, but a duplicate can still occur. If Duale AI cannot rule out receipt and cannot prove a response, the Tool outcome remains unknown. The Customer must verify the destination before any retry or compensation.

MCP Connections provide no per-call approval step. They allow or block calls for the Connection. Assigning a Connection to an Agent Identity makes every compatible Tool that the server exposes now or later eligible for that identity. The Customer must put any required human decision in its remote system, an SDK Tool, or a separate acting process.

Deadlines are absolute dates, never relative durations.

An interrupted Task can be continued. A continuation is a new Task that inherits the task history of the Task it continues, whatever that Task’s outcome, within the limits stated in Scope and limits. It carries its own deadline.

Responsibilities and use rules

Who is responsible for what

These Terms are written on the basis that, for the Platform as supplied and used within its documented intended purpose, Duale AI acts as the provider and the Customer acts as the deployer under Regulation (EU) 2024/1689. This contractual description does not determine the statutory role for every Customer integration. Article 3 determines roles from the actual development, supply, use, and branding facts. Once Article 25 applies, any circumstance listed in Article 25(1) can reallocate provider status to a distributor, importer, deployer, or other third party. Article 25(2) states that the initial provider is then no longer considered the provider of that specific system. Each party remains responsible for the obligations that the Regulation assigns to it.

Duale AI does not accept that change. Scope and limits lists the excluded uses and sets out the only open path: write to Duale AI at contact+highrisk@mail.duale.ai before you start, and obtain a separate signed agreement. Duale AI answers within 15 working days. Without a signed agreement, the use stays excluded.

The following table summarizes operational responsibilities:

The table does not determine or reallocate statutory roles. No clause shifts an obligation that the Regulation assigns to a party. Each party answers for its own obligations before the market surveillance authority.

Output oversight

You decide where to place human oversight. The Platform imposes no review step and builds none for you. For a Customer Tool, you can put the review policy in your code. For a Remote MCP Tool, you must enforce it in the remote system, through a Customer Tool that records a proposal, or in a separate acting process. A caller whose access policy permits it to stop the Task can request asynchronous Stop processing; acceptance is not proof that the Task stopped.

Duale AI recommends human oversight for external communications, code shipped to production, and any use with contractual or financial effect.

Your Obligations

• Use the Platform in accordance with these Terms and with Scope and limits, which states its intended purpose and limits
• Make state-changing Customer Tool Calls safe to repeat. When it is uncertain whether an External Effect occurred, reconcile the external system, and define a compensating action where the destination supports one
• Select, authorize, and contract for every remote MCP server you configure; give each MCP Connection only the remote access its assigned Agent Identities need; and reconcile an uncertain Remote MCP Tool Effect before retrying
• Keep your credentials secret and do not share them
• Do not hack or bypass the Platform, or test its security without authorization
• Respect intellectual property rights of Duale AI and third parties
• Report unauthorized access to the Platform, your Customer account, Platform credentials, or an integration that can use the Platform within 24 hours after you become aware of it
• Do not resell access to the Platform as such

However, you have the right to embed the Platform in your own product and expose it to your users, under your own name and responsibility. You then remain Duale AI’s sole contractual counterparty.

Uses prohibited by the Regulation

Article 5 of Regulation (EU) 2024/1689 has prohibited eight categories of practice since February 2, 2025. Added points concerning systems that generate or manipulate non-consensual intimate material or child sexual abuse material apply from December 2, 2026. The other categories concern manipulative techniques, exploitation of vulnerabilities, social scoring, solely profiling-based criminal-risk prediction, indiscriminate facial-image scraping, emotion inference at work and in education, sensitive biometric categorization, and law-enforcement real-time remote biometric identification.

These prohibitions bind you directly, without passing through this contract. Their exact scope, conditions, exceptions, and application dates are in the consolidated Regulation (opens in a new tab).

The fines in Article 99 target operators, providers as well as deployers. Duale AI answers for its provider obligations, the Customer for its deployer obligations.

Uses prohibited by Duale AI

These contractual prohibitions apply to use of the Platform, including submitted Content and Outputs. They are separate from Article 5 of the Regulation:

• Making an automated decision about a person without having arranged human oversight
• Producing illegal, hateful, or discriminatory content; spam or disinformation; child sexual abuse material; or content that infringes third-party rights
• Hacking a system, bypassing it, or running an offensive operation against a computer system
• Impersonating someone
• Using the Platform as a safety component of a system critical to people’s safety

Transparency: Who informs whom

Article 50 of Regulation (EU) 2024/1689 applies from August 2, 2026. Article 111(4) provides a transition limited to Article 50(2): providers of systems generating synthetic content that were placed on the market before August 2, 2026 must comply with Article 50(2) by December 2, 2026. That transition does not postpone the other Article 50 duties. Whether it covers a particular Duale AI system depends on that system’s placement-on-market date and the paragraph’s scope. These Terms make no claim that the transition applies.

Article 50 assigns different duties by role and use. Under Article 50(1), a provider of an AI system intended to interact directly with natural persons must ensure that those persons are informed that they are interacting with AI, unless this is obvious in the circumstances. Under Article 50(2), providers of systems generating synthetic content must meet the applicable machine-readable marking and detectability duties, subject to that paragraph’s scope and exceptions. Articles 50(3) and 50(4) impose specific deployer disclosures for emotion-recognition and biometric-categorisation systems, specified synthetic or manipulated image, audio, or video content, and qualifying text published to inform the public on matters of public interest.

Every notice or disclosure required by Article 50(1)–(4) must be clear and distinguishable, meet applicable accessibility requirements, and reach the person no later than the first interaction or exposure.

Duale AI remains responsible for the Article 50 obligations that apply to it. No clause shifts those obligations to the Customer.

The Customer does not alter markings. It does not remove, hide, or deliberately alter any marking attached to Outputs, and offers no tool that does so.

Duale AI does not deliberately remove or hide markings already present in submitted Content. Processing can change or omit content and can fail to preserve an incoming mark. Preserve and verify the original bytes when provenance matters.

The Platform API does not itself expose a conversational interface to natural persons. Re-exposing an Output does not by itself determine whether the Customer is a provider or whether Article 50(1) applies. The Customer must assess the system and use that people actually encounter, including whether the AI system engages natural persons in a genuine, direct, two-way exchange and whether an Article 50(3) or Article 50(4) disclosure applies. An Agent Identity’s identifier and description can support attribution; neither is, by itself, an Article 50 notice or a determination of legal role.

See the consolidated Regulation (opens in a new tab) and the European Commission’s Article 50 guidance (opens in a new tab) for the statutory text and current interpretation.

Health data

The standard Duale AI offering does not include the HDS certification required by Article L1111-8 of the French Public Health Code. Health data processing requires a separate written review and agreement before use: contact+hds@mail.duale.ai

Content and Outputs

Your content remains yours. You grant Duale AI a non-exclusive, worldwide license for the duration of the contract, solely for the purpose of providing the Services.

AI outputs belong to you, subject to two reservations: third-party rights in training data or source content, and the originality requirements of copyright law.

EU law protects a work only if it is the author’s own intellectual creation. Purely AI-generated content, without significant human creative input, does not necessarily qualify for copyright protection.

Content restrictions: the contractual list under Uses prohibited by Duale AI includes the restrictions on submitted Content.

Moderation: Duale AI has the right to remove illegal content or suspend an account. Duale AI gives reasons for each moderation decision and notifies the person concerned.

• Report content: contact+abuse@mail.duale.ai (response within 48h)
• Appeal a decision: contact+appeals@mail.duale.ai (response within 15 days)

Intellectual Property

The Platform belongs to Duale AI. You have the right to use it while the contract is active, not to copy or resell it.

Personal Data

Controller for account, billing, and security data: DUALE AI SAS; DPO: contact+dpo@mail.duale.ai

For the Content you submit, Duale AI acts as processor. See the “Data processing agreement (Article 28 GDPR)” section of the terms of sale.

Data collected: name, email, connection logs, platform usage.

Legal bases and roles under GDPR:

• Contract performance (Art. 6(1)(b)) applies to account and Service administration when the data subject is the contracting Customer.
• Legitimate interests (Art. 6(1)(f)) apply to administration of a corporate Customer’s authorized Users and to Service security, fraud prevention, and improvement, subject to the data subject’s interests and fundamental rights.
• Legal obligation (Art. 6(1)(c)) applies to required invoice and accounting records.
• For Customer Content and Outputs, the Customer determines the Article 6 basis as controller. Duale AI processes that data on the Customer’s documented instructions as processor under Article 28.

Duale AI operates necessary browser error, request, and performance diagnostics on its own infrastructure to maintain the Service. These diagnostics remain active when optional tracking is refused. Customer support chat and Behavior measurement have separate optional privacy choices; creating an account or accepting these terms does not enable them. Optional browser recognition is not part of the necessary diagnostics.

Retention periods: the full table is in the Privacy notice.

Your rights (GDPR Art. 15-22): access, rectification, erasure, restriction, portability, objection → contact+privacy@mail.duale.ai (response within 1 month)

International transfers: see Subprocessors. Transfers outside the European Economic Area rely on the Data Privacy Framework where it applies, or on the 2021 standard contractual clauses with supplementary measures where required.

Subprocessors: the subprocessor list sits on the Subprocessors page. The data processing agreement required by Article 28 GDPR forms the “Data processing agreement (Article 28 GDPR)” section of the terms of sale. Questions: contact+dpa@mail.duale.ai

The model providers you call with your own Provider keys are not Duale AI’s subprocessors. They are your recipients, and putting that transfer on a lawful footing is your responsibility.

Full details: Privacy notice

Data Portability and Export

Regulation (EU) 2023/2854 (Data Act) applies from September 12, 2025.

What you can retrieve, and how:

• Audit log: a tenant administrator exports it from the web interface, in a standard format
• Usage analytics: you view it in the web interface. You cannot export it through the API
• Prompts, outputs, and configurations: on request to contact+support@mail.duale.ai

Timelines and fees:

• Change notice: 2 months maximum
• Transition period: 30 days maximum
• After the transition period ends, you keep at least 30 days to retrieve your data
• Export fees: none (switching charges eliminated from January 12, 2027)

After the Customer terminates the contract, it has 30 days to export its data. Duale AI deletes the Customer’s Content after that export period. The retention periods in the Privacy notice continue to apply to account, billing, security, and execution metadata.

Security

Duale AI uses TLS for customer-facing Platform paths. It applies application-level encryption at rest to Library documents and derived search data, database backups, and reporting copies. The public-page Web cache follows the separate rule stated in the Privacy notice. Duale AI also requires multi-factor authentication for sensitive account changes, separates data per Customer, controls access by role, and records configuration, access, and task lifecycle events. The binding list is in the data processing agreement.

When an incident requires notification under applicable law, Duale AI:

• warns you within 24 hours of becoming aware of the incident;
• notifies you of the incident within 72 hours: initial assessment, severity, indicators of compromise;
• hands you a final report within one month: detailed description, root cause analysis, corrective measures.

Personal data breach notification follows a separate rule, depending on whether Duale AI acts as controller or as processor. See Security in the Privacy notice.

Report a vulnerability: contact+security@mail.duale.ai (responsible disclosure program)

Commercial terms

The terms of sale contain the detailed rules for pricing, the Wallet, payment, availability, warranties and liability, suspension and termination, balance refunds, and force majeure. The short sections below preserve clauses stated only in these Terms of Use and otherwise point to those detailed rules. If the documents contradict each other, the contract order in the terms of sale applies.

Availability

The Availability terms state the uptime position, maintenance notice, extended-outage termination right, and refund.

Liability

Exclusions: indirect damages, including lost revenue, reputational harm, and data loss.

The Warranties and liability terms and third-party claim terms state the warranties, exclusions, liability cap, exceptions, and third-party-claim rules.

Suspension and Termination

Suspension by Duale AI: in case of serious Terms violation, fraud, security risk, or authority request. Duale AI notifies the User by email, with reasons.

The Duration and termination terms and balance terms state the suspension and termination grounds, notice periods, service effect, contest path, and refund rules. The Data Portability and Export section above states the Customer’s export period and the Content-deletion consequence.

Force Majeure

The Force majeure terms state the conditions, suspensive effect, termination right, notice period, and refund consequence.

Modifications

Duale AI notifies any modification by email 30 days before it takes effect. In case of disagreement, the Customer has the right to terminate without penalty.

Disputes

Governing Law: These Terms are governed by French law.

Jurisdiction: Courts of Paris, France.

Statute of limitations: 5 years from the date the claimant knew or was reasonably expected to know of the facts giving rise to the claim (Article 2224 French Civil Code).

Claims: contact+support@mail.duale.ai

Miscellaneous

The terms of sale set out the documents that form your contract and the order in which they apply.

Invalidity of one clause does not affect the others. Not exercising a right is not a waiver of it.

Language. These Terms exist in French and English. If the two diverge, the French version prevails.

Assignment

The Customer must not assign the contract without prior written consent from Duale AI, except to a company it controls or that controls it within the meaning of Article L233-3 of the French Commercial Code, subject to prior written notice.

Duale AI has the right to assign the contract to any successor or acquirer of all or part of its business, subject to 30 days’ notice to the Customer.

Accessibility

Duale AI works to make the Platform more accessible. When a formal accessibility compliance statement is available, Duale AI publishes it separately.

• Report an accessibility issue: contact+accessibilite@mail.duale.ai

Contact

• General inquiries: contact+info@mail.duale.ai
• Technical support: contact+support@mail.duale.ai
• Legal questions: contact+legal@mail.duale.ai
• Data protection (DPO): contact+dpo@mail.duale.ai
• GDPR rights exercise: contact+privacy@mail.duale.ai
• Data processing agreement request: contact+dpa@mail.duale.ai
• Health data: contact+hds@mail.duale.ai
• Security: contact+security@mail.duale.ai
• Accessibility: contact+accessibilite@mail.duale.ai
• Content reporting: contact+abuse@mail.duale.ai
• Appeal moderation: contact+appeals@mail.duale.ai