Golden Horizons
CAPABILITY · OPS & BACK-OFFICE
AI Email Assistant
Email triage plus drafted replies. Inbox security-sensitive, nothing important missed.
Audit this workflow →What it does
Reads your inbox, labels and prioritizes threads, drafts replies in your voice for review, and unsubscribes from noise. Works as a daily digest or a continuous background process. Owner approves before anything sends.
Owners who come to us about email rarely lead with volume. They lead with the nagging fear that something important is sitting in there unseen. They're right to worry. An inbox with 200 unread messages isn't only wasted time. It's a client waiting on a contract, a vendor who needed a decision three days ago, and a referral that went cold while newsletter noise piled on top.
Generic filters don't fix that. Shoving every unfamiliar sender into a "review later" folder just relocates the anxiety. The missing piece is context. Is this email from the attorney's client about an upcoming deposition, or a newsletter from that same firm? The subject line won't tell you. The sender alone won't either. You need something that understands what matters to this owner, in this business, today.
The build works in three layers. An AI reads every incoming message and assigns an intent label: URGENT-respond, CALENDAR, INFO, NEWSLETTER, or AUTO-DELETE. That labeling logic trains on the owner's past behavior, including how they handled similar senders, which thread types they forward to staff, and which ones they delete unopened. Over the first two weeks it learns what the owner treats as urgent versus what only looks urgent on the surface.
Next, for anything labeled URGENT-respond or CALENDAR, the system drafts a reply in the owner's voice, pulled from their past outbound patterns, and surfaces it in a daily digest or a team chat alert. The owner reads the draft, edits if needed, and approves. Nothing sends without that approval. This keeps the owner in the loop while killing the blank-page problem. Replying to a draft that's 70% done and already in your tone takes two minutes. Starting cold takes twelve.
The third layer escalates ambiguity instead of guessing. When a thread looks potentially important but confidence is low, the owner gets a one-line summary and a direct question: respond now, delegate, or snooze? That prompt comes through whatever channel the owner watches most, whether team chat, SMS, or a digest email. Fewer decisions, full visibility.
Golden Horizons builds this as a scoped integration against the owner's existing Gmail or productivity suite account.
Use cases
- A managing partner at a seven-attorney firm gets 150 emails a day across three matters and firm admin. The bot surfaces four items needing a decision by noon, drafts three replies, and buries the rest. The partner spends 20 minutes on email instead of two hours.
- A consulting-firm founder leaves for a four-day conference. The bot triages the inbox in the background, escalates two time-sensitive client asks via team chat, and queues twelve drafted responses for review on return. Nothing falls through.
- An owner-operator keeps losing vendor follow-ups under client newsletters. The bot archives the newsletters, flags vendor threads URGENT, and drafts one reply template the CEO approves and clones across three open threads.
- A solo CPA hits tax season with 300 unread messages. The bot identifies eight client document requests, drafts follow-ups with specific document lists, and auto-archives IRS bulk mail and software renewal notices.
- A residential contractor gets vendor bids, client change orders, and supplier invoices in one inbox. The bot sorts them into labeled buckets, drafts approve-or-decline replies for the bids, and flags change orders that reference active job numbers.
- A regulated business BD lead loses a teaming-partner inquiry inside a 40-message thread. The bot surfaces it, summarizes the opportunity in one line, and drafts a reply proposing a call, sent inside the response window.
What’s included
- Fixed scope with written acceptance criteria before any build starts
- Customization layer for your brand voice and business rules
- Clean handover with documented runbook and live training
- Monthly ROI report for three months post-delivery
- Source code delivered to your GitHub on handover
What’s NOT included
- Third-party API subscription costs (billed to your accounts)
- Data migration from legacy systems
- Ongoing infrastructure costs after handover
How clients use this
Fixed-scope build with clean handover, documented ownership, and optional support for monitoring, maintenance, and minor changes.
Part of
Used in: Law Firms , Real Estate Agents , Construction Firms
Questions Inbox Zero Bot clients ask
Does the AI actually read the content of my emails, and who can see them?
Yes, the AI reads email content. That's how intent labeling works. The build routes your messages through a no-training, zero-retention API endpoint, so your email content never trains a model and isn't stored after the request runs. We use enterprise API tiers from AI processing or AI processing where that's a contractual term, not just a policy claim, and the signed data processing agreement is part of every engagement file. No one at Golden Horizons reads your email. The integration runs server-side with a scoped OAuth token tied to your account, with read and draft permissions only and no send authority until you approve. Want to see the exact API calls and data handling before committing? We walk through it in the audit.
What happens during the first two weeks when it's still learning? Will it miss something important?
The first two weeks are calibration, and we build a safety net for that window. During initial training the confidence threshold for escalation runs low, so the system escalates more than it needs to rather than less. Expect more team chat pings and digest items than you'll see once it's trained, and that's on purpose. As it learns your patterns, escalations drop on threads it has correctly tagged low-priority, and the draft quality tightens. Mislabeled something? Hit the one-click feedback button in the digest, and that signal feeds straight back into the labeling model. Most owners feel the system is running at the right confidence level by the end of week three. Nothing sends without your approval during calibration or after. That control never changes.
Does this work with Gmail, productivity suite, or both?
Both. The build integrates with Gmail via the productivity suite API and with productivity suite via document API. On Google, you authorize a scoped OAuth app registered to your Workspace domain. On Microsoft, same process through your identity provider. Hybrid setups, where some team members run one and some run the other, can be handled, but we scope that separately because the routing logic gets more complex. The draft interface surfaces in whichever channel you already watch: a digest email, a team chat message, or a sidebar in the inbox itself. We don't make you change how your inbox looks or works. The AI runs alongside it.
What if it escalates something as low-priority and I miss a critical email?
The core protection is calibration mode in the first two weeks, when the system escalates aggressively instead of conservatively. After calibration, two safeguards stay in place. Any sender you've flagged as a priority contact always triggers an escalation regardless of the AI's confidence, so that list is your hard override. And every thread labeled anything other than AUTO-DELETE stays in your inbox. Nothing gets deleted or moved out of reach without your explicit instruction. AUTO-DELETE only fires for categories you've pre-approved, like unsubscribe candidates, known marketing lists, and automated platform notifications. Uncertain about a category? We don't put it on AUTO-DELETE during the build. We label it NEWSLETTER and let you watch the volume before you decide. We narrow your decision surface without creating a new black box.
If the bot drafts a reply that's wrong, can I correct it and have that improve future drafts?
Yes. Every draft ships with an edit interface and a feedback flag. When you edit a draft before approving, whether you change the tone, fix a fact, or cut a sentence, those edits log as training signal. Over time the draft style tightens toward how you write now, not how you wrote six months ago. If the bot keeps missing something structural, say you always close client emails with a specific next-step format it drops, you flag it once and we add it as an explicit instruction in the system prompt. Ongoing support covers prompt updates like that as your communication style or business context shifts. Draft quality isn't frozen at build completion. It improves as long as you keep feeding it signal.