AI Transparency

Leveld is an AI-powered project-management assistant. This page is the customer-facing explanation of how AI fits into the Service: how model providers are selected, where in the product AI shows up, how we mark AI-generated content, and what controls you have. It is the companion to our Terms of Service and Privacy Notice.

1. What AI assistance means at Leveld

Leveld uses large language models to help you triage work from the Gmail connector, draft replies, surface the project items that need your attention, and summarise. The Service is an assistant: it proposes, you decide.

Leveld sends AI requests through OpenRouter only. Each workload is routed to an allowlisted model and provider endpoint selected under Leveld’s privacy, capability, and cost policy. Silent substitution outside that approved portfolio is not allowed. The model identifier and provider used for a request are recorded in the product audit trail.

2. How we use AI in each surface

AI shows up in the following surfaces:

• Ingest and extraction. When you connect a Gmail mailbox, Leveld reads the authorised content and extracts what looks like work — decisions, actions, risks, questions. Each extracted item links back to the verbatim source so you can verify it.
• Draft generation. Leveld drafts replies, summaries, and stakeholder updates. Drafts never send themselves (see §3).
• Register item creation. When Leveld creates a register item on your behalf, it labels itself as the author (visually distinct from human-authored items).
• Agent runs. Background processing (morning brief, periodic re-extraction, calibration) produces an activity-log entry so you can see what happened and when.

3. Human-in-the-loop guarantee

Every outbound action requires human approval before it leaves Leveld. The Service drafts; you review; you send. Leveld never auto-sends email without your explicit approval.

AI output can contain inaccuracies, omissions, or hallucinations inherent to current LLM technology. The Terms of Service oblige you to review every AI-generated draft, register item, or other output before relying on, sending, signing, or otherwise externally acting on it. Leveld is a tool that assists your work; you remain the decision-maker.

4. How we mark AI-generated content

Every AI-generated draft, register item, and message Leveld produces carries an in-product label so you can see at a glance what came from AI and what came from a human.

For outbound communications you send through Leveld, we append an attribution footer at the bottom of the message:

“Drafted with Leveld AI and reviewed before sending by the human reviewer.”

The footer serves two purposes: it gives recipients a plain-English disclosure that AI helped draft the content (satisfying EU AI Act Art. 50(1) interaction transparency); and it identifies the named human who reviewed before sending, which is the “ editorial control” carve-out under Art. 50(4) for any content that touches public-interest matters.

You must not strip, hide, or minimise the footer. Doing so breaks your obligations under our Terms of Service and removes the disclosure that protects recipients.

5. No training on your data

Customer content is not used to train Leveld models or provider models. The beta does not train, fine-tune, or serve Leveld-owned model weights from customer content.

Every AI request sets provider data collection to deny and setszdr: false. This means Leveld does not request or guarantee zero-data-retention routing. Prompt and response content may therefore be retained under the configured provider’s contract and published terms, even though it is not collected for model training. The detailed retention disclosure sits in our Sub-processor Register alongside the rest of the vendor disclosures.

6. EU AI Act limited-risk posture

Leveld is classified as a limited-risk AI system under Regulation (EU) 2024/1689 (the EU AI Act). The product is a general-purpose project-management assistant, not a high-risk system within Annex III, and not a prohibited system under Art. 5.

The Art. 50 transparency obligations that apply to limited- risk systems become enforceable on 2 August 2026. We have engineered to meet them in advance:

• Art. 50(1) interaction transparency. The first time you use the Service, an onboarding banner explains you are using an AI assistant. Every direct interaction (chat, draft generation, agent runs) carries a label.
• Art. 50(2) synthetic-content marking. AI-generated outbound carries the attribution footer (§4), and we additionally record the model identifier, prompt version, and rendered footer in our audit log as an out-of-band machine-readable marker.
• Art. 50(4) public-interest text. Outbound drafts are operational communications, not text published to inform the public. Where a draft might touch public-interest content, the named human reviewer’s editorial control plus the attribution footer carry the disclosure.
• Art. 50(5) timing and accessibility. Disclosures appear at first interaction, are written in clear English, and follow WCAG 2.2 AA accessibility guidance.

7. Your controls

You can:

• Disconnect Gmail at any time from account settings. Leveld stops ingesting from that source immediately.
• Erase what Leveld has learned about your workspace by emailing get@leveld.ai. We delete within 30 days of an erasure request, subject to legal retention.
• Delete your account from Settings → Account. Leveld revokes its Google access and erases your account data; section 5.6 of the privacy notice explains what happens and when.
• Export your data in a portable format before closing your account.

8. Contact

For AI-specific questions, contact get@leveld.ai. For privacy or data-protection questions, contact our Data Protection Officer at get@leveld.ai.

Last updated 25 September 2026. Contact get@leveld.ai for questions about AI transparency.