build.allanninal.dev

Part 4 of 7 · Email assistant series ~5 min read

How the assistant decides what to do

The brain reads the cleaned email and picks one of four things: answer directly from the knowledge file, draft a reply for your review, escalate to a human, or archive without replying. It’s allowed to be confident or to defer — never to invent.

The brain: four tools, one decision per email A vertical flow with four branches at the bottom. At the top, the cleaned email from the reader — a short paragraph stripped of quotes, signatures, and footers. An arrow leads down to the Brain, an AI with exactly four tools available. From the Brain, four labeled arrows fan out to four outcome boxes: Answer directly (left) for high-confidence questions answered from the knowledge file, Draft for review (centre-left) for borderline cases the human approves in seconds, Escalate (centre-right) for anything outside the assistant’s remit, and Archive (right) for cold outreach, sales pitches, and other no-reply mail. Each tool box describes its purpose in one short line. A bottom note reads: the AI picks one tool per email. It is allowed to be confident or to defer — never to invent. Cleaned email just what the sender wrote, plus envelope Brain AI with exactly four tools, picks one Answer reply now from the knowledge file Draft queue a reply for your one-tap approval Escalate forward to a human, briefing them on what came in Archive log it, file it, no reply required The brain picks one tool per email. It can be confident or defer — never invent.
Fig 4. Four tools, one pick per email. The brain can’t answer outside this menu.

Four tools, one decision

The most important rule for the brain is what it’s allowed to do. Most “AI for email” tools fail because they’re given too much freedom — they make up prices, promise discounts you’d never approve, agree to refunds without checking. Here, the brain has exactly four tools per email, and nothing else. If it can’t fit the sender’s message into one of these four shapes, it picks tool 3.

Tool 1 — Answer directly

The fast lane. The brain looks up the sender’s question in the knowledge file (your hours, your services, your prices, your common FAQs), writes a short reply in your tone, and SES sends it. The reply threads correctly under the original message, and the sender sees a polite, accurate answer in their inbox within seconds.

The brain is only allowed to use this tool when two things are both true: the answer is supported by a passage in the knowledge file, and the model is confident in the match. If either is missing, it picks tool 2 instead. Confidence isn’t a feeling here — it’s a number the brain has to send back with its tool choice. Below the threshold, this tool is locked.

Tool 2 — Draft for review

The middle lane, where most real mail lands. The brain writes the reply, but instead of sending it, it puts the draft in a small queue you can clear from your phone in seconds.

The draft includes everything you need to make a one-tap decision: the cleaned email, the suggested reply, the knowledge passage the brain used, the confidence score, and any flagged concerns (“sender mentioned a refund — might want to check the policy doc”). You read it. Tap approve and the reply goes out. Tap edit, fix the bit that’s off, send. Tap escalate and the email moves to your inbox.

This is the tool that turns the assistant from a risky auto-replier into a real assistant. Almost everything that isn’t a textbook FAQ goes through here. You’re still in the loop — you’re just not typing.

Tool 3 — Escalate to a human

The escape hatch. The brain picks this when:

  • The question isn’t covered by the knowledge file at all.
  • The sender is upset, asking for a refund, or asking about a sensitive topic (legal, medical, contractual).
  • The sender explicitly asks to talk to a person.
  • The email is a real-world request that needs action, not just an answer (“please change my appointment”, “can you ship to this address instead”).
  • The brain isn’t confident in any other tool.

Escalation forwards the original email to your team’s inbox, with a short auto-generated brief at the top: what the sender wanted, what relevant knowledge passages exist (so the human doesn’t have to dig), and what the brain’s best guess at a draft would have been. The human walks into a brief, not a cold thread.

Tool 4 — Archive (no reply)

For mail that doesn’t want or need a reply at all. Examples that often land here:

  • Cold outreach pitches from sales tools.
  • “Just checking in” emails from vendors.
  • Notifications that snuck through the noreply filter.
  • One-line “thanks!” replies that close out a thread.

The brain logs the message and the reason in the audit table, and the lane ends. No reply, no draft, no notification. The audit log is searchable, so if you ever need to confirm whether the email arrived, you can — but it’s not stealing your attention in the meantime.

What it never does

A short list of things the brain refuses to do, by design:

  • Quote a price, hour, or policy that isn’t in the knowledge file.
  • Promise anything — a refund, a callback, a delivery date — without explicit support in the file.
  • Accept calendar invites or commit to meetings on your behalf.
  • Click any link in the email, including unsubscribe links.
  • Continue a thread when it’s clearly out of its depth.

If the sender pushes for any of these, the brain picks tool 2 or tool 3. The shape of the four tools makes this the only natural outcome — there’s no “wing it” option in the menu.

How the four tools split your inbox in practice

A rough volume mix at small-business scale, after the rules in part 2 have already filtered out noreply and VIP traffic:

  • Tool 1 — Answer: 30–50% of remaining mail at a mature business with a good knowledge file. The same five questions, asked five hundred ways.
  • Tool 2 — Draft: 30–40%. In-between cases, longer requests, anything where confidence is good but not great.
  • Tool 3 — Escalate: 10–20%. The work that really needs you.
  • Tool 4 — Archive: 5–15%. The stuff that didn’t deserve to reach a human at all.

The exact mix shifts as your knowledge file grows. The first weeks lean heavy on tools 2 and 3 (the brain hasn’t seen enough yet); by week six, tool 1 is the largest group. Every approved draft is a small note to the future about what kind of question can be auto-answered.

In plain words

The brain has a small, fixed menu. Every email is a pick from four tools. The AI is fast and flexible inside that menu — it can phrase the same answer ten different ways, depending on the sender’s tone — but it never tries to do something that’s not on the menu. Boring, safe, correct.

All posts