Ask a small business owner what eats their day and email is almost always the first answer. Not because any single message is hard — but because there are forty of them, scattered between the urgent, the important, the ignorable, and the outright spam. Every one of them demands a judgment call, and judgment calls are exactly what your AI agent is now good enough to make.

Email triage is the highest-ROI first delegation for most small businesses. It's high volume, mostly routine, and the cost of getting it wrong is recoverable — a mislabeled message costs you ten minutes, not a customer. Here's how to build an inbox agent that sorts, summarizes, drafts, and flags with enough reliability that you trust it by week three.

Why Email Is the Right First Delegation

Before you hand your agent anything, it's worth understanding why email beats every other candidate for your first serious automation:

  • It's asynchronous. Email doesn't require your agent to interact with a live human. Mistakes can be reviewed before they matter.
  • It's high volume. Even a modest inbox produces 30–80 messages a day. The time savings compound immediately.
  • It has obvious structure. Sender, subject, thread context, and timing give the agent concrete signals to route on — no fuzzy judgment required for the bulk of messages.
  • It's already text. You don't need integrations with complex systems to start. Your agent can read and draft email on day one.

The catch: email also contains your most sensitive information, and one badly-worded reply to the wrong person can cause real damage. That's why the design of your triage system matters more than the AI itself.

The Five-Bucket Inbox Model

Stop thinking of your inbox as a list of messages. Think of it as five buckets. Every incoming message should land in exactly one of them, and each bucket has a defined handling policy your agent can execute without improvisation:

  1. Critical — needs you now. Angry customers, legal or compliance notices, urgent vendor issues, anything with money or deadlines attached. The agent flags these immediately and surfaces a one-paragraph summary.
  2. Action — needs you, not urgent. Things only you can decide, but that can wait a few hours. The agent files them into your "Action" queue with a suggested next step.
  3. Delegate — the agent can handle it. Scheduling requests, status inquiries, order confirmations, FAQ-style questions. The agent drafts a reply and queues it for your one-tap approval, or sends it autonomously once trust is established.
  4. Reference — read and archive. Newsletters, receipts, notifications. The agent summarizes anything genuinely relevant and archives the rest without bothering you.
  5. Noise — delete or unsubscribe. Spam, malware bait, and messages you never opted into. The agent removes them and reports the count in your daily digest.

That's it. Five buckets. If your agent can sort every message into one of these five, you've already reclaimed the majority of the time email used to take. Drafting comes second; sorting is the win.

Routing Rules: The Three Filters

Agents are good at fuzzy judgment, but you don't want fuzzy judgment on every message. Build a layered routing system where the most reliable signals fire first:

Filter 1: Sender authority

The sender is the strongest signal. Build an explicit list of high-priority senders (key clients, your accountant, your landlord, regulators) and an explicit list of low-priority senders (bulk mailers, automated notifications). Everything else falls through to the next filter. These lists are your guardrails — they never rely on the model's guesswork.

Filter 2: Subject and intent

For the remaining messages, the agent classifies intent from the subject line and thread context. Words like "invoice," "overdue," "refund," and "complaint" route to Critical. "Question," "availability," "quote" route to Delegate. "Newsletter," "receipt," "your order" route to Reference. Teach your agent your business's vocabulary — the terms that actually appear in your industry.

Filter 3: Body urgency

The last filter reads the body for urgency signals: deadlines, escalations, repeated follow-ups, dollar amounts, legal language. Anything ambiguous at this level defaults to Action rather than Delegate. When in doubt, the agent should surface the message to you, never silently bury it.

Here's what the routing rules look like in practice — plain-language conditions your agent can follow consistently:

# Inbox routing rules (v3)
if sender in HIGH_PRIORITY:        → Critical + summary
if sender in LOW_PRIORITY:         → Reference/Noise per sender type
if subject matches URGENT_TERMS:   → Critical + summary
if intent == "delegable_question": → Delegate + draft reply
if body mentions deadline/money:   → Critical or Action
else:                              → Action + suggested next step

Notice the hierarchy: explicit lists beat keyword matching, keyword matching beats body analysis, and the default is always to escalate rather than guess. That ordering is what keeps error rates low.

What the Agent Should Never Do

Guardrails matter more in email than in almost any other automation, because the blast radius includes your reputation. Write these rules explicitly and test them before letting anything run unsupervised:

  • Never send to a new recipient without approval. Replying to an existing thread is one thing; adding someone to the conversation is a decision only you make. For the first month, all outbound mail requires your approval.
  • Never discuss money, contracts, or confidential numbers. Not in drafts, not in summaries shared to third-party tools, not ever until you've reviewed the flow. The agent can flag these messages and stop.
  • Never apologize on your behalf or make commitments. "We'll get back to you by Friday" is a commitment. The agent drafts acknowledgments and questions, not promises.
  • Never delete anything you haven't seen. Even spam should live in a quarantine folder for a day. The cost of one wrongly-deleted client email is not worth the convenience.
  • Never fabricate context. If the agent doesn't have the facts (pricing, timelines, project history), the draft should say "I'll confirm the details with our team" — not invent them.

Write these as explicit, testable rules — the same way you'd brief a new hire. Vague instructions like "be careful with sensitive information" are how agents drift; specific conditions are how they stay locked in.

From Sorting to Sending: A Three-Week Rollout

Every inbox automation should start read-only and expand in stages. Trust is built with data, not vibes. Here's the rollout cadence we recommend:

Week 1 — Observe and sort: The agent reads your inbox, assigns every message to a bucket, and produces a daily summary. Nothing is archived, nothing is drafted, nothing is sent. At the end of each day, spot-check ten buckets and correct the agent's mistakes. This week builds your routing rules with real data.

Week 2 — Draft: The agent starts drafting replies for the Delegate bucket and suggested next steps for Action items. All drafts queue for your review. Track your edit rate: if you're changing more than 20% of drafts, the agent needs more examples of your voice. Feed it your best past replies as style references.

Week 3 — Send with approval: For the narrowest, most routine categories — scheduling confirmations, order acknowledgments, FAQ answers — the agent sends after a one-tap approval. Everything else stays in draft. By the end of the week you should know which categories can progress.

Week 4+ — Autonomous with exception review: The agent sends autonomously in the categories that hit 95%+ approval, and only surfaces exceptions. You review a daily digest: everything sent, everything escalated, everything unusual. Keep the high-risk categories on approval forever — it costs you five seconds and prevents the one email you can't unsend.

Metrics That Tell You It's Working

An inbox agent isn't a set-it-and-forget-it tool. Track these three numbers weekly and you'll know exactly when to expand or pull back:

  • Miss rate: How many messages should have been flagged Critical but weren't? This should be zero, permanently. Anything above zero means your routing rules have a hole.
  • Draft edit rate: What percentage of agent drafts do you change before sending? Under 20% means the voice is right; over 40% means the agent needs better style examples.
  • Time to inbox zero: How long does it take you to clear your Action queue each day? This is the number that pays for the whole system — and it should drop by a factor of three within a month.

Review these numbers in a weekly check-in, the same way you'd review any new hire's performance. Email volume changes, your business vocabulary changes, and your agent's routing rules need to change with them.

Start With Sorting, Not Sending

The biggest mistake we see is skipping straight to "let the agent send emails." That's the last thing you should automate, not the first. Sorting is low-risk and immediately useful. Drafting is medium-risk and takes minutes per day. Sending is where the risk lives — and where the trust you've built in weeks one and two pays off.

Set up the five buckets, write the routing rules, and let the agent spend a week just organizing your inbox. You'll be surprised how much of the stress of email disappears the moment you stop looking at a flat list of forty messages and start looking at three that actually need you.

This post is part of our practical agent series. For more on defining what your agent can do autonomously, see The Agent Autonomy Spectrum. For keeping the system honest over time, see The Weekly Agent Audit. And for the guardrails that keep outbound messages safe, see The Cost of a Hallucination.