How DeskGrade uses AI

1.About this document

This document states DeskGrade’s terms and commitments. It is not legal advice.

2.Voice, messages and marketing

Our recording policy requires recording to be off by default. Employee meetings, incoming calls, outgoing calls and shared rooms each require a defined purpose, lawful basis, audience notice and audio/transcript retention before recording is used. AI disclosure does not replace a recording notice. One general messaging and interaction layer uses channel adapters. Service messages must be distinguished from marketing; marketing requires applicable permission and working opt-out and suppression controls that agents may not override.

3.Purpose and customer deployment records

DeskGrade is intended for productivity and customer support. Recruitment ranking, employee scoring or evaluation, employment decisions and behaviour- or personality-based work allocation are outside its intended purpose. We commit to a platform system record with distinct product and agent profiles. The customer must record purpose, necessity, lawful basis, data sources, notices and any required consultation. We commit to providing easy, versioned management of that deployment record in DeskGrade. The customer remains responsible for its own basis; the effect of customisation, redistribution and substantial modifications on responsibilities is assessed case by case.

4.AI identity and assistance

Agents in DeskGrade are AI systems, not people. In DeskGrade chat, messages written by an agent carry an agent icon and an “AI” label, and appear under the agent’s own name, not a person’s.

When an agent participates in a conversation, we commit to a first-interaction AI disclosure and a visible AI identity, including for external recipients and room guests. AI assistance is disclosed by default even when a person reviews or approves the message. Human review or approval will be described only when it actually occurred. A suitable first notice is: “You are interacting with a DeskGrade AI assistant.” A human-approved message remains AI-assisted; approval does not make the agent a human author.

5.Content made by agents

We commit to preserving provider provenance marks and using standards-aligned machine-readable marking in supported export and channel formats. For supported media, this includes IPTC/C2PA without a visible stamp. Mark preservation through copying and transport must be verified for the format; a visible AI label or internal log alone does not establish machine detectability. When sharing output, disclose AI involvement as required.

6.People stay in control

• A personal agent acts with the permissions of the person it works for, never more.
• Company agents act with a role the company gives them, and cannot reach another company’s data.
• An agent operates within its delegated authority and applicable policy. Approval is required for dangerous or irreversible actions; ordinary reversible actions may run directly. Review proposed actions and output before relying on them.
• Every agent action is written to the audit log together with the change itself.

7.Memory

Agents can remember preferences, decisions and facts about the work. People can review what is remembered, correct it and have it forgotten. Work and personal material stay apart.

8.AI models we use

DeskGrade connects to OpenAI directly, Vercel AI Gateway, OpenRouter as a backup, and Anthropic using our own key through Vercel AI Gateway. We prefer providers with zero retention and no training on customer inputs; this is a routing preference, not a guarantee for every request. Provider terms and the configured route determine processing. Hosting in Stockholm, Sweden does not mean that AI inference, support or all onward processing stays in the EU. EU-only model routing is not currently guaranteed. See the current directly connected providers and each gateway’s published provider list. Pilotify does not train models on customer content.

9.Limits of agent output

AI output can be incomplete, out of date or wrong, and can sound confident when it is. Agents cite their sources where they answer from company knowledge; check important answers against the source before you rely on them.

10.Uses we do not support

DeskGrade is not designed for, and may not be used for, decisions that have legal or similarly significant effects on people based solely on agent output, or for uses the EU AI Act classifies as high-risk, such as screening job applicants or assessing creditworthiness, unless we have agreed to a lawful use in writing; prohibited AI uses are never permitted. See our terms of use.

11.Our role under the EU AI Act

Pilotify is the provider of DeskGrade as an AI system: we design how agents identify themselves, how their output is attributed and which safeguards apply. A customer that uses DeskGrade in its own business is a deployer, and is responsible for using it as instructed, for informing the people it deals with where the law requires, and for the AI literacy of its own staff. The providers of the underlying general-purpose AI models carry their own obligations for those models.

12.Questions and concerns

If you think an agent has acted wrongly or you were not told you were dealing with an AI, contact jan@pilotify.app or use the contact form.