build.allanninal.dev

Part 1 of 7 · Email assistant series ~5 min read

An email assistant on AWS for a few dollars a month

Your inbox is doing too much work. Most of what lands there is repeats — the same five questions, asked in different words. Here’s how to build a small email assistant that reads each new message, replies from your own knowledge when it can, drafts the rest for your approval, and quietly escalates anything important.

The whole system on one page

Before any code, here’s the shape of what we’re building.

System architecture: three outside surfaces, three inside AWS At the top, three external surfaces in a row. On the left, “The sender” — the person emailing your business. In the middle, “Your knowledge” — FAQs, hours, tone, sender allowlist. On the right, “Your team” — the inbox where escalations land. Each connects via an arrow to the AWS account container below. The sender has a two-way arrow labeled “email” representing both inbound mail and outbound replies. Knowledge feeds into the brain. The brain forwards complex emails to your team. Inside the AWS account are three components in a row, mirroring the layout above. On the left, the Reader — receives the email, parses it, strips quoted threads and signatures. In the middle, the Brain — decides one of four moves: answer, draft, escalate, archive. On the right, the Sender — sends the auto-reply, queues drafts for review, or forwards to a human. Internal arrows flow left to right. A note at the bottom reads: every email gets one of four outcomes. The AI never invents an answer that isn’t in your knowledge file. The sender emails in Your knowledge FAQs, tone, allowlist Your team when a human is needed email in & out guides escalate when needed AWS account Reader receives the email, strips the noise Brain picks one of four: answer, draft, escalate, archive Sender does what the brain decided clean text decision Every email gets one of four outcomes — and never an invented one.
Fig 1. Three outside surfaces, three pieces inside AWS. Mail in, reply or escalate back, with a four-tool brain in the middle.

What you set up once (the outside)

  • A receiving address — hello@yourbusiness.com or similar. Either pointed straight at the assistant, or forwarded from your existing inbox so nothing about your address changes for senders.
  • A short knowledge file — your FAQs, your hours, your prices and policies, the tone you want, and a small allowlist of senders who should always go straight to a human. Lives in a Google Doc you can edit anytime.
  • An escalation inbox — the address the assistant forwards to when an email needs a human. Your normal inbox, a shared team inbox, a help-desk queue — whichever you already use.

What runs on every email (the inside)

  • The reader — takes the raw inbound email and turns it into a clean, focused message. Strips quoted threads, signature blocks, footers, and trackers, so the brain reads what the sender actually meant to say.
  • The brain — reads the cleaned email and picks one of four moves: answer directly from the knowledge file, draft a reply for your review, escalate to a human, or archive without replying.
  • The sender — carries out the decision. Sends the auto-reply with proper threading. Or queues a draft you approve in seconds. Or forwards the email plus a one-line summary to your team. Or files it quietly.

In plain words

An email lands. The cloud reads it. A small AI decides whether it’s a question your knowledge file can answer, a question that needs a human, or noise. If it can answer, it does — in your tone, citing your own words. If it can’t, it puts a draft on your desk or hands the whole thing to your team. The sender hears back fast, and you stop spending mornings retyping the same five replies.

Total cost runs in coffee-money territory at typical small-business volume — a few cents per email, going up smoothly with how often the inbox rings.

Design rules that shaped every decision

  • Stay inside the AWS always-free quotas where possible. SES has a per-email cost, but it’s in fractions of a cent.
  • The assistant answers from your knowledge file only — never invents prices, hours, or promises.
  • Auto-send only when confidence is high. Anything borderline becomes a draft you approve, not an apology you send later.
  • Threading must be preserved. The reply must show up under the original message in the sender’s mail client, not as a fresh thread.
  • Configuration lives in a Drive doc you can edit. Updating tone, hours, or the allowlist never needs a deploy.

Why this shape

Most “AI for email” tools fall over for one of three reasons. They charge a per-seat fee that costs more than the work they’re doing. They make up prices and policies in their replies. Or they’re a black box where your customer history sits inside someone else’s product.

The setup above is the smallest one I could find that handles all three. One way in (your address), one way out (your team or the sender), three small pieces in the middle that read, decide, and act. Everything stays in your AWS account, so the data is yours and you pay per use, not per seat.

The next five posts walk through each piece in turn — how an email enters and gets routed, how the reader strips it down, how the brain picks one of four moves, how a reply stays accurate, and what the whole thing actually costs. One diagram per post. A final engineering reference at the end gives engineers the dense version with precise service names and model IDs.

All posts