Indie Hackers: Building in Public Intro

Building an AI Email Assistant That Doesn't Make Stuff Up

Hey IH,

I'm building Inbox SuperPilot — a Chrome extension that lives in Gmail and drafts replies from your docs and CRM, citing its sources for every claim.

The short version: I got tired of AI being helpful right up until the moment it becomes a liability.

TL;DR

• Generic AI drafts fast but gets business-specific details wrong — wrong pricing, deprecated features, invented policies
• Knowledge-grounded AI retrieves from your actual docs before writing, then cites the source so you can verify
• Inbox SuperPilot lives inside Gmail, connects to Drive, Notion, Confluence, HubSpot, and 20+ sources, and never auto-sends — every draft is for your review

The moment that started this

The day ChatGPT quoted the wrong price to a $50K prospect, I stopped trusting generic AI for email.

Like a lot of founders, I spend an embarrassing amount of time in email. Enterprise SLA inquiries. Support tickets. Investor updates. Endless onboarding threads. On a normal day, easily 2–3 hours.

So of course I tried using ChatGPT to speed things up. It was genuinely fast — until it told a customer our Pro plan was $29. It's $25. Same draft also referenced a feature we'd already deprecated. I caught it before sending, which makes this a less dramatic story than I'd like, but it scared me enough to rethink the whole approach.

The problem with AI email isn't writing quality. It's factual accuracy. If the model doesn't know your business, it shouldn't be speaking for it.

People treat this like a prompting issue. I don't think it is. You can write better prompts, paste more context, babysit every draft harder. But if you're manually feeding the model your pricing, policies, and conversation history every time, that's not an email workflow. That's part-time model supervision.

I started asking around founder networks and heard the same stories. Almost everyone had at least one near-miss where an AI copilot confidently hallucinated a refund policy or promised a feature that wasn't on the roadmap.

What we built

Inbox SuperPilot works inside Gmail as a Chrome extension. No separate inbox, no new app.

You connect your knowledge sources — Google Drive, Notion, Confluence, help center content, CRM data from HubSpot or Salesforce — and when you open an email, it drafts a reply pulled from that material. The things I care about most:

• It cites its sources so you can verify exactly which document a claim came from
• It creates drafts only — you review, edit, and send yourself
• It adapts your writing voice per recipient, because a factually correct email that sounds robotic still damages trust
• It works inside Gmail, because nobody wants another inbox product

That citation layer is the whole point. If AI answers a pricing question or explains a policy, I want to see which document it pulled that from. Otherwise you're outsourcing risk to an ungrounded model and hoping for the best.

The two camps of AI email

I think this market is splitting into two camps.

There's AI that helps you write better — cleaning up wording, adjusting tone, getting past a blank page. That's the ChatGPT/Gemini bucket, and it's useful. We wrote a detailed breakdown of where each of these tools falls short for business email in ChatGPT vs Gemini vs Claude for Email: Where Generic AI Falls Short.

Then there's AI that can answer accurately on behalf of your business. Did it use the right pricing? Reflect the current policy? Answer every question in the thread? That's a much harder problem, and it's the one I'm working on.

A polished wrong answer is still a wrong answer. And a $50K prospect who gets quoted the wrong price doesn't care how smooth your tone was.

Anyone technical is reading this and thinking: "That's just RAG." Fair. The concept isn't new. What matters is the execution: hybrid semantic and keyword retrieval so you find the right doc and not just a similar one, a quality guard that checks whether every question got answered, and source citations on every claim. The goal was never to build a safer ChatGPT. It was to build the first email workflow where you stop babysitting the model entirely.

If you want to see how this plays out in a real workflow, we wrote up the customer support use case in detail: Why Generic AI Fails Customer Support Teams.

Privacy in one line

Zero-retention APIs, no training on your data — your inbox stays yours. Full details on our security page.

Where things stand

We're in early launch. Pricing:

• Free forever — 50 drafts/month, no card required, no countdown timer
• Pro — $25/month for 500 drafts + expanded KB access
• Founder's rate — $19/month locked forever for the first 100 users

I wanted people to actually use it before deciding, not get pressured into a trial while they're still connecting their first knowledge source.

What I most want feedback on

Does the positioning land?

Our tagline is "AI email that actually knows your stuff." Subhead: "Draft replies grounded in your docs and CRM, with sources cited for every claim — right inside Gmail."

The behavior change question underneath is real: most people's default is still "paste context into ChatGPT." I believe people switch when they've been burned by wrong info, when they're tired of the copy-paste ritual, or when they want something embedded directly in Gmail.

But I'd love to know what the actual trigger would be for you.

Why I'm posting now

Because this is the stage where outside feedback is most useful. Not "would you use AI?" — that's too broad.

The real question: do people care enough about hallucinations in email to want a purpose-built tool instead of a general-purpose model?

I think yes. Especially once money, customer trust, or policy accuracy is on the line. But that's exactly the kind of belief that gets stronger — or gets corrected — by talking to smart people in public.

If you've built in this space, sell to founders, or have strong opinions on positioning, I'd genuinely appreciate the pushback.

Further reading

• Why Generic AI Fails Customer Support Teams (And What to Use Instead)
• ChatGPT vs Gemini vs Claude for Email: Where Generic AI Falls Short
• Superhuman vs Inbox SuperPilot: Productivity Suite vs Email Accuracy

References

1. Customer-specific AI beats generic models — SAP
2. Knowledge-centered GenAI for support teams — Elastic
3. Retrieval-augmented generation for knowledge-intensive NLP tasks — Lewis et al., Meta AI Research