> For the complete documentation index, see [llms.txt](https://docs.solcis.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.solcis.ai/legal-centre/ai-transparency-statement.md). # AI transparency statement Where Solcis uses AI, what it can and cannot do, and how you stay in control of it. This statement describes where the Solcis platform uses artificial intelligence (AI), what that AI can and cannot do, and how our customers stay in control of it. ## About this statement * **Who publishes it:** Solcis Pty Ltd ("Solcis", "we"). * **What it covers:** the Solcis platform as we supply it to customers. It does not cover how Solcis as a company uses AI internally, or the AI agents our customers govern with Solcis. * **Who it is for:** customers and prospective customers, their risk, compliance and procurement teams, auditors, regulators, and members of the public. * **Version:** 1.0, published 19/10/2026. We review it at least once every 12 months, and whenever a release changes how AI is used in the platform. We prepared this statement in line with ISO/IEC 42001:2023 Annex A controls A.8.2 and A.8.5 (information for users and other interested parties). Its structure follows the Australian Government's Standard for AI transparency statements. The appendix maps it to each framework we considered. ## What Solcis is and the role AI plays Solcis governs AI agents. Before an agent carries out an action, Solcis checks it against the controls the organisation has adopted and returns a decision. After the check, it records that decision in a tamper-evident log. In security terms, Solcis is a policy decision point. The customer's agent framework is the enforcement point that acts on the decision. Solcis does not take actions on its own. Most of Solcis does not use AI. Rule matching, pattern-based content checks, decision logging, shadow mode, the kill switch and human approvals are all deterministic: the same input always gives the same result. AI is used only where a policy author chooses it, for example a rule that asks a model to judge an action. AI is also used in optional authoring and monitoring aids that help people write and maintain policies. Solcis is self-hosted. It runs in the customer's own environment, and the AI models it uses run alongside it in that environment by default. ## Where AI is used in Solcis Solcis uses AI in six places. Some are off until the customer configures a model. In each case, AI either supports a human decision or can only make an agent action more restricted. AI cannot approve an agent action on its own. | Capability | What the AI does | Human role | Default state | | ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------- | | AI judgements in policy rules | Assesses an agent action against a written criterion, or screens it for content-safety, prompt-injection or personal-information risks | The AI can deny or escalate to a person, but cannot allow. If the model fails or is unavailable, the action is held for a person. | Used only in rules a policy author writes to use it | | Policy compilation | Turns adopted controls into draft policies | A person reviews and approves every policy version before it is tested | Customer-initiated | | Policy drafting assistant | Writes a first draft or revision of a policy from selected controls | A person approves every draft | Off until the customer configures a model | | Control import | Extracts controls from an uploaded document | A person reviews the import. If extraction is incomplete, the import is refused rather than partly applied. | Customer-initiated | | Drift detection | Rates how significant a change to a control or agent is | Recommendations go to a person to approve. With no model available, changes are flagged for human review. | On, recommend-only | | Decision review | A larger model re-checks earlier AI judgements and suggests improvements | Findings go to a person. This capability never changes policy and never trains models. | Customer-initiated | The customer can replace any default model with one they chooses. If they choose an externally hosted model, for example a commercial large language model reached over the internet, the provider of that model is the customer's supplier, not ours. A conversational policy editor (powered by the customers chosen AI model) is also available. It stays off until the customer turns it on, and a person approves any policy it produces. Two further features, a console assistant and an AI severity rating, are locked off in the current release. This statement will be updated if either is enabled. ## Human oversight and customer control People at the customer organisation decide what Solcis enforces, when it enforces it, and how held actions are resolved. Solcis gives them these controls: * **Decision outcomes.** Each check returns allow, deny, modify, step-up or defer. Step-up and defer hold the action until a person resolves it. If anything goes wrong, the action is blocked. * **Severity labels.** A decision can carry a severity of low, medium, high or critical. This label records how serious a matter is. It never changes the outcome on its own. * **Human approval.** A held action is released only when an authorised person resolves it, with separation of duties. If no one can resolve a hold, the action is treated as blocked. * **Staged rollout.** Each policy moves through draft, staging, shadow and production. A person approves each version before it moves forward. * **Shadow mode.** A policy can run in shadow mode, for the whole engine or rule by rule. In shadow mode, Solcis records what it would have decided without affecting the agent. * **Kill switch (roadmap).** Authorised people can stop agent actions across the engine. Each use is recorded as a deliberate decision. * **Turning off AI judgements.** Customers can switch off AI judgements entirely. Any AI judgement that is skipped is recorded as an escalation, never as an allow. One limitation: if an AI judgement does not flag an action and no other rule objects, the action is allowed. Customers should combine AI judgements with deterministic rules for high-impact actions. See [Accuracy, limitations and known risks](#accuracy-limitations-and-known-risks). ## Our role and how Solcis is classified Solcis supplies the platform. Customers decide how to use it and remain responsible for their own AI agents and their own compliance. This statement gives customers the information they need to meet their own obligations. It does not say that using Solcis makes a customer compliant. **Roles.** * Under the EU AI Act, Solcis is the provider of the platform and the customer is the deployer. * Under data protection law (GDPR, UK GDPR, the Australian Privacy Act, ADGM and UAE law), the customer is the controller of any personal information in agent actions. * Because Solcis runs in the customer's environment, Solcis does not receive, store or process that information. This is why there is no data processing agreement for the product. **EU AI Act risk classification.** Solcis assesses AI agent actions against an organisation's controls. It does not assess, score or make decisions about individual people, so we do not consider it a high-risk system under Article 6 and Annex III. The intended purpose is to govern AI agent actions, and this classification depends on that purpose. A customer that uses Solcis to monitor or evaluate workers, or as the deciding step in an Annex III use such as credit or employment decisions, should assess its own obligations. **Transparency for people interacting with AI (EU AI Act Article 50).** Where Solcis features let users talk to an AI model, such as the drafting assistant or conversational editor, the interface makes clear that the content is generated by AI. **Third-party models.** Solcis uses open-weight models published by third parties. It does not modify, fine-tune or retrain those models. Where a customer chooses a general-purpose model from another provider, that provider's own obligations apply. ## How AI in Solcis uses data Solcis AI processes data inside the customer's environment, and customer data is not used to train AI models. | Topic | What happens | | -------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Data the AI sees | For AI judgements: the action an agent intends to take, its source and destination, and the parts of its payload the rule refers to. These may include personal information. For authoring aids: control and policy text. | | Where processing happens | In the customer's environment, by models deployed next to Solcis. Data leaves that environment only if the customer configures an external model or service. | | Model training | Solcis does not train or fine-tune models on customer data. | | Learning across customers | Solcis does not combine data from different customers. | | Samples kept for decision review | When decision review is enabled, recent decision inputs are kept in the customer's environment for 48 hours by default. The customer can set this between 1 and 168 hours. | | Decision log | Payload content is not written to the decision log. User, source and destination identifiers and the reasons for each decision are recorded. | | Data sent to Solcis | Only an hourly count of policy evaluations for billing, and only where the deployment is metered and the customer turns it on. No decision content is sent. Air-gapped deployments send nothing. | | Model sources | Default models are open-weight models, each model is signature-checked before it is deployed. A register of models and their licences is available to customers on request. Customers may deploy and integrate with their own models. | ## Accuracy, limitations and known risks AI judgements in Solcis make mistakes, and customers should not rely on them as their only safeguard. The main risks we have identified, and how they are addressed: * **Missed problems (false allows).** A model may fail to flag a harmful action. AI judgements cannot allow an action on their own, but a missed detection with no other objection results in an allow. We recommend deterministic rules and human approval for high-impact actions. * **Unnecessary blocks (false denies).** A model may flag a legitimate action. Shadow mode lets customers measure this before enforcing, and held actions go to a person. * **Evasion.** A crafted input may try to mislead a model. Deterministic checks run alongside AI judgements, and failures result in a hold, not an allow. Heavily disguised inputs, such as encoded text, should be decoded before a model sees them. * **Personal information in agent actions.** Inputs to AI judgements may contain personal information. Processing stays in the customer's environment, and payload content is kept out of the decision log. * **Unfair outcomes.** We have not yet tested the models for different outcomes across groups of people. * **False assurance.** A Solcis decision record shows that a control was checked. It is not proof that an agent's overall behaviour complies with a law or standard. We do not yet set a minimum accuracy that a model must meet before it can be used for enforcement. Results in a customer's environment will differ, and customers should validate models against their own workload. The fast inline check, drift detection, decision review and the drafting assistant have not yet been benchmarked. ## Records and audit evidence Solcis records every decision before returning it, so customers can show auditors what was decided, when and why. * **Tamper-evident log.** Each decision record is hash-chained and digitally signed (Ed25519). Anyone with the public key can verify it offline. * **What a record shows.** Each record shows the outcome, the policy and rule that produced it, any severity label, and the reason given. Where the customer enables it, the record also shows which model made the judgement. * **Independent witnessing (optional).** Records can also be published to a transparency log and co-signed by a witness outside the customer's control. This is available with Enterprise licences. * **Who holds the records.** The decision log sits in the customer's environment. Solcis does not hold copies. These records can support a customer's own record-keeping and oversight duties, for example under Articles 12 and 26 of the EU AI Act for the customer's agents. Whether they are enough for a particular obligation is for the customer to decide. Solcis accepts OSCAL control catalogues as input. It does not yet produce OSCAL assessment results as output. ## Governance and accountability Solcis is building an AI management system aligned to ISO/IEC 42001, together with an information security management system aligned to ISO/IEC 27001. Neither is certified yet. * **Accountability.** CTO is accountable for responsible AI in the Solcis platform. CISO approves AI impact assessments and changes to how AI is used. * **Impact assessment.** We assessed the impact of the Solcis platform on individuals, groups and society using ISO/IEC 42005. The assessment covers the risks set out above. It is reviewed whenever we add or change an AI capability. * **Model selection.** We evaluate each default model for accuracy, licence terms and origin before including it, and record the results in a model register. * **Release controls.** AI features we have not been able to validate, such as the AI severity rating, are locked off in released versions. * **Monitoring.** Customers can use shadow mode and decision review to monitor AI judgements in their environment. We use customer feedback and reported issues to review our choice of default models and our policy packs. * **Incidents.** If we become aware of an AI-related issue that could affect customers, for example a model fault that leads to wrong decisions, we will notify affected customers within agreed timeframes, per contract. Security vulnerabilities can be reported under our [vulnerability disclosure policy](/legal-centre/vulnerability-disclosure-policy.md). * **Staff.** People who build and support Solcis receive training on AI risks and responsible use, in line with Article 4 of the EU AI Act. ## Questions, concerns and your rights To raise a question or concern about AI in Solcis, email . We will acknowledge your message and tell you how we will handle it. * **If you are affected by a decision about you.** Solcis decides about agent actions, not about people. But an organisation may use an AI agent governed by Solcis in a process that affects you. That organisation is responsible for the decision. Under laws such as the GDPR, UK GDPR, the Australian Privacy Act and the ADGM Data Protection Regulations 2021, you may have rights to information about automated decisions, to human review and to contest the outcome. Contact that organisation first. We will help our customers respond. * **Personal information we hold.** See our [privacy policy](/legal-centre/privacy-policy.md) or contact . * **Security issues.** See our [vulnerability disclosure policy](/legal-centre/vulnerability-disclosure-policy.md). * **Customers.** Customers can ask for more detailed technical information through their account contact. This includes model details, test results and deployment guidance. ## Changes to this statement We update this statement at least every 12 months. We also update it when a release adds, removes or materially changes an AI capability or a default model. | Version | Date | Change | | ------- | ---------- | --------------- | | 1.0 | 19/10/2026 | First published | ## Appendix: regulatory alignment map This statement addresses transparency provisions in six frameworks. For most of them, the duty falls on our customers as deployers or controllers, and Solcis provides the information they need. Dates are as at 19 October 2026. | Framework | Provision | What it asks for | Who it binds | | ----------------------------------------------------------------------- | -------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- | | ISO/IEC 42001:2023 | A.8.2 System documentation and information for users | Information users need about the AI system | Solcis | | ISO/IEC 42001:2023 | A.8.3 External reporting | A channel for interested parties to report adverse impacts | Solcis | | ISO/IEC 42001:2023 | A.8.4 Communication of incidents | A plan for telling users about AI incidents | Solcis | | ISO/IEC 42001:2023 | A.8.5 Information for interested parties | Determine and document what is reported to interested parties | Solcis | | ISO/IEC 42001:2023 + ISO/IEC 42005 | A.5.2–A.5.5 AI system impact assessment | Assess impacts on individuals, groups and society | Solcis | | EU AI Act (Reg (EU) 2024/1689, as amended by Reg (EU) 2026/1744) | Art 50(1), (5); applies from 2 Aug 2026 | Tell people when they are interacting with an AI system, by the first interaction at the latest | Solcis as provider | | EU AI Act | Art 6, Annex III; high-risk duties apply from 2 Dec 2027 | Classify the system and its intended purpose | Solcis | | EU AI Act | Art 13(3) (used as a model, not a binding duty) | Instructions for use: purpose, accuracy, risks, oversight, logging | High-risk providers | | EU AI Act | Arts 12, 14, 26 | Logging, human oversight, deployer monitoring and log retention | Customers' high-risk systems | | EU AI Act | Art 4 | Support AI literacy of staff | Solcis and customers | | GDPR | Arts 13(2)(f), 14(2)(g), 15(1)(h), 22 | Tell people about automated decisions, the logic involved and their consequences | Customers as controllers | | UK GDPR (as amended by the Data (Use and Access) Act 2025) | Arts 22A–22D; in force 5 Feb 2026 | Safeguards for significant automated decisions: information, representations, human intervention, contest | Customers as controllers | | DORA (Reg (EU) 2022/2554) | Arts 28, 30(2)(a)–(b) | Describe ICT services, subcontracting, and data processing locations | Financial entity customers (by contract) | | Australia: Standard for AI transparency statements (DTA AI policy v2.0) | Statement content requirements | Purpose, usage, unreviewed public impact, monitoring, compliance, last updated date, contact, annual review | Commonwealth entities (structure adopted voluntarily) | | Australia: Privacy Act 1988 | APP 1.7–1.9; from 10 Dec 2026 | Privacy policies must describe automated decisions that significantly affect individuals | Customers as APP entities | | Australia: Guidance for AI Adoption (NAIC, Oct 2025) | Practice 4, Share essential information | Disclose AI use, capabilities, limitations and supply chain | Voluntary | | Australia: APRA CPS 230 | Para 54 | Material service provider agreements | Regulated customers (by contract) | | ADGM Data Protection Regulations 2021 | Notice and automated decision-making rights | Tell people about automated decisions and the logic; human intervention and contest | Controllers in ADGM | | UAE PDPL (Federal Decree-Law 45 of 2021) | Art 18 | Right to object to automated decisions with legal or serious effects | Onshore UAE controllers | | UAE Charter for the Development and Use of AI (2024) | Principles 5, 6, 7 | Transparency, human oversight, governance and accountability | Voluntary | --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.solcis.ai/legal-centre/ai-transparency-statement.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.