What agents actually do
with a real inbox.

Agent reads the inbound email and decides a follow-up sequence is warranted for this contact.

It creates an ad-hoc enrollment with custom prompts — referencing the specific ask, the timeline mentioned, the tone of the conversation. Not a template. Written for this email.

Steps fire on schedule: day 0, day 3, day 7. Each message is generated at send time using the original context snapshot.

Contact replies — sequence exits automatically. No stale follow-up goes out after a conversation has already started.

Email arrives. Phishing detection runs immediately — suspicious signals are flagged before anything else happens.

Workflow rules evaluate the message: who sent it, what it's about, how it matches configured conditions. Labels applied, thread routed to the right inbox.

An AI condition evaluates whether the email needs an immediate reply. If yes, a draft is prepared and held for review — not sent.

Priority score computed. High-priority messages surface in the agenda. Everything else is handled quietly.

Agent queries the relationship graph: last contact date, interaction frequency, topics discussed, connection strength.

It reads the contact's enrichment data — AI-generated digest of what this person cares about, based on your actual history with them.

Message composed with that context embedded — the prompt instructs the LLM to reference specific shared history. Not "Hi {{first_name}}, hope you're well."

If the contact doesn't respond, a sequence picks up — each step still aware of what was said before.

Agent token is issued with requires_approval on email:send. Any send the agent attempts is automatically held.

Agent reads inbound email, runs an AI workflow, and generates a reply. The reply is queued as a held action — not sent.

You receive a task in your agenda: "Action needed — reply ready for review." Open the approval queue, read the draft, approve or reject.

Approved — message sends immediately. Rejected — discarded, agent moves on. Everything logged either way.

Inbound email to a shared address fans out to every team inbox via a distribution list. All agents see it.

Thread assigned to the most relevant agent. Internal notes visible to the team, invisible to the sender.

Contact shared across inboxes — overlap detection shows which team members already have a relationship with this person and how strong it is.

New team member added — the onboarding flow surfaces their top contacts automatically, seeded from existing team relationships.

A new message arrives with a question, request, attachment, or exception. The agent identifies what context it needs before it responds.

It retrieves relevant thread history, contact details, prior files, extracted fields, labels, and recent activity from the inbox record.

The draft, answer, or escalation is generated with citations to the operational context the inbox already holds.

If confidence is low or the action is sensitive, the agent routes it to a human with the supporting context attached.

When you sign up, your tenant is assigned a region for stored workspace data. AI processing follows the tenant's enabled configuration and consent state.

AI features run only when enabled. We don't train general-purpose models on customer content, and retention behavior should be configured before launch for the tenant's selected provider path.

Every action your agent takes is written to an immutable audit log — who did what, when, to which resource. Searchable and exportable.

LLM consent is managed per-tenant and reversible. Revoking it stops all AI features immediately and removes stored AI-generated content. Data deletion requests are processed end-to-end via the API.