Inbo: An AI email assistant that drafts every reply in your voice

01 / Overview

About Inbo

02 / The challenge

What Inbo needed to solve

The emails that cost people the most time are the middle ones: a client question that needs three careful paragraphs, a delay to announce, a firm but polite follow-up.

1. 01

   Routine replies eat hours, and using a chat assistant means copying, prompting and pasting, which costs more time than typing.

2. 02

   Generic AI replies do not sound like the sender, and the right tone changes from one recipient to the next.

3. 03

   Emails about money, contracts or an angry customer must never be answered by a machine without a person checking first.

4. 04

   The product had to work the same way across two mail providers, Gmail and Outlook.

03 / What we built

How we built Inbo

We built Inbo as one Next.js application in TypeScript that watches the user's mailbox, sorts every message, drafts replies in their voice and puts the results in a review queue.

• 01

  Gmail and Outlook intake

  Gmail arrives through Google Pub/Sub push and Outlook through Microsoft Graph notifications, with a 60-second polling fallback. New mail is picked up without the user doing anything.

• 02

  Three-step sensitive hold

  Automated mail is discarded, a keyword check flags five hold categories, and a language model confirms. Financial, legal and other sensitive emails are held for a person, and an Explain why panel shows the reason.

• 03

  Drafts in your voice

  Routine mail gets a draft written from the thread and a style profile that records formality, sign-off and tone for each sender. Every edit the user makes feeds back into that profile, so drafts get closer over time.

• 04

  Review queue

  The queue shows held drafts first and routine replies below, with counts for waiting, needs judgment and ready to send. Each draft can be sent, edited then sent, left or discarded from one screen.

• 05

  Optional autopilot

  Autopilot can send replies on its own, but it is off by default and never overrides a hold. Users stay in control of what leaves their inbox.

• 06

  Searchable history

  A history log records what Inbo drafted, sent, held, discarded or skipped, searchable by sender. Users can check any past decision in seconds.

• 07

  Marketing site and onboarding

  We built the marketing site, pricing page and a four-step onboarding with sign-in through Google or Microsoft. New users can connect an inbox without a card.

• 08

  Plans and reliable processing

  Stripe handles Free, Pro, Teams and Enterprise plans. A background worker takes jobs from a BullMQ queue five at a time, retries each up to three times and skips duplicates, and OpenAI drafts with Anthropic as a backup.

04 / The results

What changed for Inbo

Inbo takes an email from arrival to a ready draft, so users review replies instead of writing them from scratch.

• Replies are already drafted in the user's own voice when they open the queue.
• Sensitive messages about money, contracts or complaints are held for a person instead of being answered automatically.
• Drafts improve as the style profile learns from each edit.
• Drafting keeps working if one AI provider fails, because Anthropic takes over when OpenAI does not respond.
• Message metadata is kept for 90 days by default, and users can schedule deletion of their account at any time.