Artificial intelligence policy

It covers every feature in THEMIS that uses a language model to read, answer, draft, review, summarise or suggest. Today these are the Agent, Tabular Review, Contract Intelligence, legal research with cited answers, document reading, drafting help, and the optional interpretation of a described scenario in Simulations. Parts of THEMIS that follow fixed rules, such as the scoring and the checks in Simulations and the transfer duty engine, do not use a model to decide anything.

• A person decides. Every output is a proposal. A person with the right role must accept it before it affects a matter, a document, money or anything sent outside the practice. The assistant cannot approve, sign, lodge, release money or confirm that an identity is verified, and those acts are blocked in the software.
• Answers can be checked. Answers are built from your files and the legal sources in your library, show the passage they rely on, and should say when the files do not answer the question.
• It is clear when AI was used. Output from an assistant feature is labelled as proposed or generated where it appears, and recorded in the AI activity log.
• Your information is not used to train models. We do not use the content of your matters, or the outputs, to train any model of ours, and we do not authorise our model providers to use them to train theirs.
• We keep it to its purpose. The assistant is a productivity tool for licensed professionals. It is not built for the uses the Artificial intelligence terms rule out, such as decisions about a person's access to credit, employment or essential services, or assisting a court.

Practice administrators have an AI control centre. From it they can switch assistant features and agents off, set limits, pause an agent, and read the AI activity log, which records who asked, about which matter, and what was produced. We can suspend an assistant feature at any time for safety or security reasons.

To answer a request, THEMIS sends the request and the relevant extract of the matter to a model provider, and only what is needed. Model providers are listed on the Sub-processors page with where they run. We give customers at least 30 days' notice before a new sub-processor starts to handle their content, so they can object.

We test the parts we can test exactly. The rules, scoring, date reading, clause detection and extraction in THEMIS are covered by automated tests and by labelled evaluation cases, including cases that check that the software leaves a fact unstated instead of inventing it. A defect found this way is fixed and the case is kept. A language model cannot be tested that way, because the same question can give different wording. For that reason its output is always a proposal that a person checks, and we do not claim a measured accuracy for it.

• Language models can write fluent text that is wrong, incomplete or out of date, can invent a citation, and can sound more certain than they should.
• A scan or photo may be misread, and a document may be missed.
• The legal library holds only the sources your practice has added, so the assistant cannot know law that is not there, and it does not replace the official text.
• Automatic checks on uploaded documents look for inconsistencies. They cannot prove that a document is genuine.

These limits are why a professional must read and check what the assistant produces before relying on it.

You must not use the assistant for biometric categorisation or social scoring, for decisions that have legal or similarly significant effects on a person without a person's meaningful review, for anything the Acceptable Use Policy forbids, or to produce content intended to deceive a court, a regulator or a counterparty. We may suspend a practice that does.

We may change the models and providers behind a feature to improve quality, cost or security. A change that alters what is sent to a third party, or who receives it, goes through the sub-processor notice in clause 4. We aim to say when a change is likely to alter the style or reliability of answers.

If an assistant produces something harmful, discloses information it should not, or behaves in a way you did not expect, tell us at support@olympuspay.co, or at security@olympuspay.co if it concerns security. We investigate, correct the cause, and tell the affected practice what we found. A personal data breach follows the process in the Data Processing Agreement, including notice without undue delay.

We want you to know where this policy stops.

• No independent body has audited or certified our AI practices.
• We do not publish measured accuracy for the language-model features.
• Our view of how the EU AI Act classifies the assistant is our own assessment and has not been confirmed by a regulator or an outside assessor.

Our leadership is accountable for this policy. We review it at least once a year and whenever an AI feature, a model provider or a law it relies on changes materially. The effective date at the top of this page shows when the current version took effect.

Write to legal@olympuspay.co about this policy, or privacy@olympuspay.co about personal data.