Velorith gives every member of a company their own AI agents. This page explains how these agents work, which AI models they use and who provides them, how we mark AI-generated content and where people stay in control. It supplements our terms of service and our privacy policy and reflects our transparency obligations under Regulation (EU) 2024/1689 (AI Act).
1. Agents are AI systems
Velorith agents are AI systems within the meaning of the AI Act. They are built on large language models that generate text based on probabilities. Agents answer questions, draft texts, search the company brain, search the web through Parallel, visit websites in a hosted browser at Browser Use, work with files and commands in a member's workspace and, with approval, carry out actions on a member's computer. Everything an agent writes or does is generated by AI on the instruction of a member, even where the reply reads as if a person had written it.
Velorith is a general assistant for work tasks. It is not designed for the prohibited practices or high-risk uses listed in the AI Act. Our acceptable use policy excludes these uses.
2. Models and providers
All model calls are sent from our infrastructure to OpenRouter, which forwards each request to the provider that runs the model. We currently use the following models.
| Model | Developed by | Run by | Used for |
|---|---|---|---|
| GLM 5.3 Flash | Z.ai | Fireworks, Together, Parasail or Baseten for agents, and other providers selected by OpenRouter for company brain processing (see section 3) | Default model for agents, and for extracting and classifying company brain content and recognising text in uploaded images |
| Claude Haiku 4.5 | Anthropic | Anthropic | Agents, if a member selects this model |
| Claude Sonnet 5 | Anthropic | Anthropic | Agents, if a member selects this model |
| text-embedding-3-small | OpenAI | OpenAI | Search index of the company brain (embeddings) |
Members can choose the model for a conversation in the chat. The models we offer and the providers that run them can change as better or more suitable models become available. We keep this page and our list of subprocessors up to date and inform customers about new subprocessors in advance as described there.
The software also contains a function for connecting a member's own ChatGPT subscription. This function is not available, and no requests are sent to OpenAI via this route.
3. Your data and model training
We do not use customer content to train AI models. Agent requests go through our model gateway, which sets data collection to deny, fixes the providers listed above and disables automatic fallback to other providers, so an agent request is never rerouted to a provider we have not chosen. With data collection set to deny, OpenRouter forwards a request only to providers that, according to OpenRouter's classification, do not use request data for training. Requests for extracting and classifying company brain content and for recognising text in images also set data collection to deny, but do not fix the providers, so OpenRouter may route them to other providers that meet this condition. Embeddings for the search index are created by the company's brain server through OpenRouter, which forwards them to OpenAI. These embedding requests currently do not carry the data collection setting. How long the model providers keep request data is governed by their own terms with OpenRouter. Details on what data is transmitted are set out in our privacy policy and our data processing agreement.
4. When you are interacting with AI
Agents in Velorith are always shown as AI agents with their own name, and in the web app also with their own avatar. Members know from the product that they are talking to an AI system and not to a person.
At launch, members will also be able to reach their agents via iMessage and WhatsApp. Replies in these channels carry the name of the agent, and the line is presented as an AI agent, so that it is clear that the conversation partner is not a person.
5. Marking of AI-generated content
At launch, messages that agents send to people outside the company through the channels Velorith provides are marked as AI-generated, in line with Art. 50 of the AI Act. Customers and members may not remove or conceal this marking.
Content that leaves Velorith in other ways is not marked automatically, for example text that a member copies out of a conversation, files that an agent creates in a workspace or entries that an agent makes in a connected third-party service. If a customer uses such content to communicate with third parties, it is responsible for any disclosure required by law, as set out in section 5.4 of our terms of service.
6. Limits of AI output
AI output can be wrong, incomplete, outdated or inappropriate, even when it sounds convincing. Agents can misunderstand instructions, invent facts or sources, or draw wrong conclusions from documents. AI output is not legal, tax, medical or other professional advice. Please check important output before you rely on it or pass it on, in particular figures, quotes, legal statements and anything that will be sent to others.
7. Human oversight and approvals
People decide what agents may do. Actions with external effect, such as executing commands, sending messages, changing data in connected services or acting on a member's computer, are classified by their potential impact. Depending on the approval rules that the company sets, an action is either submitted to a member for approval before it is executed or executed without individual approval. Approval requests expire after 15 minutes if nobody decides them. Every approval decision is recorded in the company's audit log.
The service provides an emergency stop with which authorised members can halt agent activity in the company.
Automatic screening with Jev (planned)
At launch, we plan to offer an auto-accept mode for actions that agents carry out on a member's computer. In this mode, a planned action is assessed by Jev, a classifier provided by TypeSafe, before it is executed. Jev receives text only, namely the planned steps, the accessibility tree of the active window, the name of the app and the task the member approved. It does not receive screenshots.
- Fixed rules come first. Apps that a member has set to never, the company's lock list and kinds of action that always require approval, such as sending, paying, deleting or changing permissions, are never decided by Jev
- Confidence decides. If Jev is highly confident that the action matches the approved task, it runs. If confidence is medium, the member is asked. If confidence is low, the action is blocked
- Fails closed. If Jev does not answer in time, returns an error or is unavailable, the member is asked, as without auto-accept
- Traceable. Every automatically accepted action is logged with the label and confidence assigned by Jev, and approval requests show both to the member
Jev is itself an AI system and can misjudge an action. It supports the company's approval rules but does not replace the company's own review of which actions it allows without approval.
8. Reporting problems with AI content
If an agent produces content that is harmful, unlawful, misleading or clearly wrong, please tell us at [SUPPORT_EMAIL]. Please include the time, the agent and, if possible, the conversation concerned. If you become aware of misuse of Velorith or of a security issue, for example an agent bypassing approvals or accessing data it should not see, please report it to [SECURITY_EMAIL] as described in our vulnerability disclosure policy.