Unattended AI outreach costs trust, and the market spent 2025 learning it. CommsOperator is built the other way round: the agent proposes, a person decides, and the log shows who decided what, for every message.
CommsOperator's AI agents triage conversations, draft replies, plan and write outreach, and classify prospect replies, at one of three autonomy levels: observe, draft for review, or auto-send behind a feature flag an admin controls. Policy packs set, per pipeline, the risk tier, who must approve, when two approvers are needed, banned topics, and mandatory disclaimers. Every step is logged with input, output, decision, and token use. One switch turns AI off for a workspace.
Choose how much rope the agent gets: watch only, draft for review, or send on its own. Draft for review is the default, so nothing goes out unread. Letting the agent send on its own is off until an admin turns it on, and even then every message passes the same daily cap, opt-out check, and EU review rule as a message a person sends.
Decide, per pipeline, what the AI may say, what it must never bring up, which disclaimer it has to include, and who signs off before anything risky leaves. Sponsor outreach can run looser than investor updates. Each pack sets a risk tier, a sending mode, approver roles, dual approval for high-risk content, and channel limits.
Reviewers decide in one screen instead of digging. Each draft arrives with the agent's reasoning, the compliance flags, and the policy that fired. Edit, approve, reject, or skip a step. Every human override is logged with who and when.
A record of every step the AI took and why, that you can hand to an auditor. Each step records phase, input, output, decision, tokens, duration, and outcome. An opt-in compliance engine adds trace IDs, content hashes that prove nobody edited the record afterwards, and a JSON report your compliance team can read without us in the room.
One toggle turns all AI off for a brand, instantly. AI endpoints are rate-limited per tenant and per IP so a runaway script cannot burn through your allowance, and AI usage is metered per plan so you can always see what was used.
Drafts use your own help articles, training notes, and the contact's CRM history, not the open internet. They run on OpenAI's GPT-4.1 and GPT-4o family models through the API, chosen per feature. Under OpenAI's API policy your content is not used to train its models, and CommsOperator trains none. On a dedicated instance, local Ollama inference keeps AI on your own server.
Numbers from the production instance and behaviour checked against the code. No projections, no testimonials.
What this module does not do today, and what to do about it. If one of these is a blocker, you know before you commit, not after.
Watch only and draft for review never send anything. Auto-send exists, behind a feature flag an admin turns on for a workspace, and even then every message passes the daily cap, the suppression check, and the EU downgrade. Every action, including an auto-sent one, is in the log with the decision that allowed it.
No. CommsOperator does not train models. AI calls go to OpenAI's API, whose data usage policy excludes API content from training OpenAI's models. Prompts include only the context needed for the task, not your whole mailbox.
Yes. A policy pack can require a second approver from a named role for a pipeline, alongside disallowed topics and mandatory disclaimers. Investor updates can run on dual approval while sponsor outreach runs on one.
CommsOperator is not classified as a high-risk system. For Article 50 transparency, auto-sent AI drafts to likely EU/EEA recipients are held for human approval by default, and where a workspace opts out, an AI-origin disclosure line is appended. Human-approved sends are accountable to the human sender.
Tell us what you run and we will be straight about the fit. Private beta. A person reads every request.