AI inbox triage and follow-ups: a playbook

The inbox problem: everything urgent, nothing prioritized

Open a business inbox on a normal Tuesday and every message renders exactly the same way: a bold subject line and a timestamp. The supplier dispute that could cost you a customer looks identical to a newsletter. So triage happens by scanning — and because new mail keeps landing, it is work you redo all day.

Follow-ups are worse, because they are invisible. The quote you sent last week, the invoice now overdue, the client who said "let me check and get back to you" — none of these appear in the inbox at all. An inbox records what arrived; it holds no state for what you are waiting on. Those threads live in someone's head until they fall out of it. And the recurring admin — the Monday summary, the weekly chase list — arrives on the calendar's schedule, competing for the same scarce attention.

The playbook below fixes all three with one agent and five design decisions. The through-line: the agent does the reading, sorting, drafting, and remembering on its own; anything that leaves under your name waits for a human.

Step 1 — Define the triage rules the agent applies

You already have triage rules; they just live in your head. Write them down in three lists. Labels: the categories your mail actually falls into — new enquiry, existing customer, booking, invoice, supplier, newsletter. Priorities: what jumps the queue — anything mentioning a cancellation, anything from a key account, anything overdue. Routing: who or what handles each category — sales threads to the person who owns the pipeline, booking requests toward the calendar, noise archived.

The agent then applies those rules to every message the moment it lands: it labels, sets a priority, and routes. These are internal actions — they change nothing outside your workspace — so they run around the clock without approval gates. Ground the rules in your own documents, too: Olano agents classify and draft against a knowledge base built from what you upload — price lists, policies, SOPs — so "is this a warranty case?" is answered from your warranty policy, not a guess. Start with one workflow: a single inbox, your ten most common message types, nothing more.

Step 2 — Reply drafts: the agent drafts, you approve

For most messages, triage should end with a draft, not a to-do. The agent reads the thread, checks the relevant documents, and writes the reply — then queues it. This is the line the whole playbook stands on: drafting is internal, so it runs on its own; sending is an external action, so anything outbound waits for approval by default. The agent spends the minutes; you spend the seconds.

How much waits in the queue is a dial, not a switch. Olano agents run at explicit trust levels 0–4, and most inbox workflows start at level 1 (Assistant), where every send needs sign-off. As your review history grows, routine categories — opening hours, order status, standard pricing — can be promoted toward level 3 (Autonomous), where in-scope replies send without per-message approval while per-category gates keep financial actions and anything unusual queued for a human. The full ladder is covered in our playbook on automating customer enquiries with human approval.

Approval requests arrive as ordinary messages in the channels your team already uses — Slack, WhatsApp, Telegram, or any of the 15+ channels Olano supports — so reviewing the queue means answering a message, not learning a new dashboard.

Step 3 — Follow-up chasing on your schedule

This is where the agent adds the state your inbox lacks. Whenever a thread ends with you waiting — a quote awaiting a decision, an invoice awaiting payment, a question awaiting an answer — the agent puts it on a watch list. Persistent per-agent memory keeps that list alive across sessions, so a thread opened three weeks ago is still being watched three weeks later, whether or not anyone remembers it.

You set the cadence: nudge after three days, escalate to a firmer note after seven, stop after three attempts, never chase certain senders. When an interval lapses, the agent drafts the follow-up in your tone and queues it for approval like any other send. Nothing chases anyone until you say so — but nothing gets forgotten either. The deals that quietly cost the most are the ones that died not from a "no" but from silence.

Step 4 — Recurring admin, on schedule

The third leg is the admin that recurs whether or not the inbox is busy, handled with scheduled tasks and proactive monitoring. Typical jobs: an 8am digest of what arrived overnight, sorted by priority, with drafts already attached; a weekly list of every thread still waiting on someone else, with days outstanding; an end-of-month sweep of unanswered enquiries. Monitoring works the other way around: the agent flags matches immediately, so a message from a key account or anything mentioning a cancellation reaches you the moment it lands, not in tomorrow's digest.

Step 5 — The audit trail

Every action in this playbook — the label applied, the draft written, the approval given or declined, the follow-up sent and when — lands in an immutable audit trail. It earns its keep three ways. When someone asks "did we ever reply to them?", you look it up instead of reconstructing it. When declines cluster in one category, the trail shows you whether the fix is a triage rule or a stale document. And when you are deciding whether a category is ready for level 3, you decide from a record of hundreds of reviewed drafts, not a feeling.

What it plugs into

Email reaches the agent as a channel: Gmail through the built-in integration, or any standard IMAP inbox. But triage rarely stops at email — a booking request needs the calendar, a sales thread should update the pipeline — so the agent works through 75+ built-in integrations, each with its own specialist subagent: Google Calendar for scheduling context, and HubSpot, Salesforce, or Google Sheets on the CRM side. Digests and approval requests, meanwhile, can meet your team in whichever of the 15+ messaging channels you already have open.

A single managed inbox workflow like this is typically live within 1–2 business days of onboarding — live in days, not months — and every build is quoted and approved before we start. If mornings in your business begin with an hour of scanning and sorting, the AI inbox assistant is a good first workflow.