AI Use Disclosure

Lumevo uses large language models through a Hermes agent runtime. In the workspace that shows up as Tally in Ask AI, and as the worker behind a My Work task.

• It reads the prompt, the recent thread or task, and any company-library passages or connector reads that the run is allowed to use.
• It produces a draft: a chat reply, a file, a structured question, a proposed task, a proposed automation, or a proposed external action.
• It stops. A person answers, reviews, accepts, retries, or approves.

Ask AI is chat-only. It cannot send mail, change a record, or mark work done. My Work can propose a mailbox draft; creating that draft in Outlook still needs review by the task owner and approval by an owner or manager. Company connectors are read-only. Automations only create ordinary My Work tasks when the clock fires.

• It does not make a decision that grants or refuses a significant legal or contractual right.
• It does not have spending authority, trading authority, or a right to send mail.
• It does not complete a task. Done is a human-recorded outcome.
• It does not train Lumevo-owned models on Customer Content.
• It does not run as an unsupervised employee.

The live states a person will see are the real ones:

• Needs My Input — the run cannot continue without an answer from the person who created the task
• Needs Review — a draft is ready; accept, revise, or retry
• Needs Approval — a proposed external action is waiting for an owner or manager
• Needs Attention — something broke and a person has to look
• Done or Failed — recorded, not implied

An approval is a decision record. It is not proof that mail left the building or that a third-party system changed. The current mail executor creates a draft. Unknown action types fail closed.

A turn may send Customer Content to: the Atlas application server in the operator’s environment, the database, private file storage, and the configured inference provider used by Hermes. If the task is allowed to read a connector or a mailbox, that content can be included. Library files are only used if a manager or owner has approved them for AI use.

Ask AI history is stored by Lumevo and purged 30 days after the thread goes quiet. Task drafts, files and approvals stay with the workspace until the workspace ends or the Customer asks for deletion as in the Terms.

The inference provider is a subprocessor. Its region and its own training policy must be named on the Subprocessor List before paid production use. Lumevo will not opt in to a provider training switch where the provider offers a business no-training mode.

Drafts can be wrong, incomplete, out of date, or too confident. They can miss a qualifier that was in the source file. They can invent a citation. They can reproduce bias that was in the prompt or the library. A person who knows the work has to read the output before anyone relies on it.

Lumevo is not a law firm, accountant, financial adviser, migration agent or medical device. If a Customer uses it in a regulated workflow, the Customer remains the regulated person.

Untrusted text — an email, a web page, a uploaded PDF — can try to instruct the model to ignore Lumevo’s rules. The product is built so the model does not get send-mail or connector-write credentials. Residual risk remains: the model may still put something it should not into a draft, and general network egress from an agent environment is a known remaining control gap described in the Data and Security Terms. Do not treat the sandbox as a reason to skip review.

As between Lumevo and the Customer, the Customer may use drafts produced for it in the workspace. That does not mean the draft is free of third-party rights. Models can emit text that is close to material they were trained on. The Customer is responsible for checking outputs it publishes or sends.

Australia has not enacted a text-and-data-mining copyright exception. Lumevo does not train its own models on Customer Content, which is the part of that debate this product actually controls.

There is no EU-style AI Act in force in Australia as of September 2026. Existing law still applies: the Privacy Act, the Australian Consumer Law (including misleading conduct and unfair contract terms), copyright, and sector rules. OAIC guidance on commercially available AI products expects organisations not to paste personal information into public AI tools. Lumevo is a private workspace with human gates, not a public chatbot.

From 10 December 2026, APP entities must say in their privacy policy if a computer program makes, or is substantially involved in, a decision that significantly affects an individual’s rights or interests. Lumevo’s position is that it does not make those decisions. The Privacy Policy states that. Customers who use drafts as an input to hiring, credit or similar decisions must make their own APP 1.7 disclosure.

Voluntary Australian AI safety practices still matter for tenders. Lumevo’s relevant practices are: named accountability (the Customer owner), draft-first operation, human review, logging of task states, and no claim of certified high-risk AI.