AI Transparency Statement

In short: the AI sorts your inbox, drafts replies, plans outreach, summarises meetings, and answers from your knowledge base. A person decides what gets sent, unless an administrator turns on automatic delivery, which is off by default and gated as section 3 describes. Models run on the OpenAI API and are not trained on your content; we train none. Every step is logged. EEA recipients get a human review, or a disclosure line. In short boxes are a plain-language guide. If a summary and the full text differ, the full text applies.

This statement describes how AI works in CommsOperator so a customer, a security reviewer, or a recipient can judge it. It is written for the transparency expectations of the EU AI Act and for security questionnaires that now ask about model provenance, retention, and human oversight.

1. What the AI does

In short: six jobs, and for each one the table says who has the final word. Triage runs automatically and a person can always overrule it; everything that leaves the building is a person's decision unless an administrator chose otherwise.

How to read this table: each row is one AI feature, what it produces, and who decides what happens with the result.

Feature	What it produces	Who decides
Inbox triage	A classification: Focused, Other, or Spam	Automatic classification; a person can reclassify and the inbox rules always win
Reply drafting	A proposed reply grounded in the thread, knowledge base, and CRM context	A person edits and sends
Sales agent	Outreach plans, drafted steps, compliance checks, reply intent classification	A person approves each step unless an administrator enables automatic delivery for a fixed template
Knowledge chatbot	Answers from your published content with citations	Answers automatically where the widget is set to auto-reply, with escalation to a person
Content tools	Drafted articles, changelog entries, task plans, summaries	A person reviews before publishing
Meeting summaries	Summaries, participants, next steps from transcripts	A person reviews; tasks are created only on confirmation

2. Providers and data handling

In short: drafts run on OpenAI's API. Your content is not used to train their models, and we train none. The AI sees the thread and the matching articles, not your whole mailbox. A dedicated instance can keep inference on the instance.

AI processing runs on the OpenAI API using GPT-4.1 and GPT-4o family models, selected per feature.
Under OpenAI's API data usage policy, content sent through the API is not used to train OpenAI's models.
CommsOperator trains no models on customer data and builds no cross-customer profiles.
Prompts carry only what the task needs: the conversation, retrieved knowledge, and relevant CRM fields.
Embeddings of knowledge base content are stored in the customer's own workspace for search.
Dedicated instances can run local inference with Ollama, so prompts never leave the instance.

3. Human oversight

In short: watch only, draft for review, or send on its own. Draft for review is the default. Auto-send needs a feature flag, a permitting policy pack, and every send check; one toggle turns all AI off for a workspace.

Three autonomy levels: observe, draft for review, and auto-send. Draft for review is the default.
Auto-send requires a feature flag, a policy pack that permits it, and passes suppression, daily caps, and the EU AI Act check. For fixed-template campaigns it also requires an administrator to confirm the exact rendered sequence; any change invalidates that consent.
Policy packs set risk tiers, required approver roles including dual approval, disallowed topics, required disclaimers, and channel limits.
An administrator can disable all AI in a workspace with one switch.

4. Logging and auditability

In short: every AI step and every human override is recorded, so you can show an auditor what the AI did, what a person changed, and when. An opt-in compliance engine adds tamper-evident hashes and a machine-readable report.

Every agent step records its phase, input, output, decision, token use, duration, and outcome. Human actions on drafts (unlock, edit, reject, manual send) are logged with the actor and time. An opt-in compliance engine adds trace identifiers, content hashes for tamper detection, and a machine-readable compliance report for authorised auditors. Administrative actions are recorded in the audit log with actor, resource, IP address, and user agent.

5. EU AI Act

In short: CommsOperator is not a high-risk AI system. From 2 August 2026, an AI-drafted message headed for an EEA recipient waits for a person unless the workspace opts out, in which case the message says it came from an AI.

CommsOperator is not a high-risk AI system under the EU AI Act: it does not make decisions producing legal or similarly significant effects on individuals. For Article 50 transparency duties, which apply from 2 August 2026:

Automatically sent AI-drafted messages to recipients likely to be in the EEA are downgraded to a human approval by default, so an accountable person sends them.
Where a workspace explicitly opts out of that downgrade, an AI-origin disclosure line is appended to the message, telling the recipient the message was generated and sent by an AI assistant and that they can reply to reach a person.
Recipient location is inferred from the contact's country field, then the email domain. Errors default to the safer behaviour: hold for human approval.
The chatbot identifies itself as an AI assistant and offers a route to a person.

6. Known limits

In short: the AI can be wrong, triage is a heuristic, EEA detection is a heuristic, and expired agent logs are not yet purged automatically. We would rather say so here than have you find out.

AI output can be wrong, out of date, or confidently plausible. It is a proposal, never advice.
Grounded answers depend on the quality of your knowledge base; gaps produce a refusal, not an invention, but a wrong article produces a wrong answer.
Classification is a heuristic: a legitimate message can land in Other, and a well-crafted spam message can land in Focused.
EU residency inference is a heuristic; a .com address with no country on file is not flagged as EEA.
Automated purging of expired agent logs is not yet enabled; a six-month working window is computed and reported.

7. Reporting a problem

In short: if the AI produced something harmful, misleading, or unlawful, email office@commsoperator.com with the workspace and the approximate time. We can trace the exact run.

If an AI feature produced something harmful, misleading, or unlawful, tell us at office@commsoperator.com with the workspace and approximate time. We can trace the exact agent run and its inputs.