Golden Horizons
CAPABILITY · OPS & BACK-OFFICE
Inbox Zero Bot
Email triage and drafted replies — inbox cleared, nothing important missed.
$4,500–$6,500 build · $1,500–3,000/mo
Talk to us about a Inbox Zero Bot build →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 continuous background process. Owner approves before anything sends.
Most owners who come to us about email aren't complaining about volume. They're complaining about the specific, gnawing fear that something important is sitting in there and they haven't seen it yet. They're right to worry. An inbox with 200 unread messages doesn't just mean wasted time — it means a client waiting on a contract, a vendor who needed a decision three days ago, and a referral that went cold while the newsletter noise piled up on top.
Generic filters don't fix this. Moving everything from a sender you don't recognize to a "review later" folder just relocates the anxiety. What's missing is context: is this email from the attorney's client about an upcoming deposition, or is it a firm newsletter from that same law firm? The subject line doesn't tell you. The sender alone doesn't tell you. You need something that understands what actually matters to this owner, in this business, on this day.
The build works in three layers. First, an AI reads every incoming message and assigns an intent label: URGENT-respond, CALENDAR, INFO, NEWSLETTER, or AUTO-DELETE. The labeling logic is trained on the owner's past behavior — how they've responded to similar senders, which thread types they forward to staff, which ones they delete without opening. Over the first two weeks, it learns what the owner actually treats as urgent versus what just looks urgent on the surface.
Second, for any message labeled URGENT-respond or CALENDAR, the system drafts a reply in the owner's voice — pulled from their past outbound email patterns — and surfaces it in a daily digest or a Slack alert, depending on preference. The owner reads the draft, edits if needed, and approves. Nothing sends without that approval. This isn't about removing the owner from the loop; it's about eliminating the blank-page problem. Replying to a 70%-finished draft that already uses your tone takes two minutes. Starting from scratch takes twelve.
Third, the system escalates anything ambiguous rather than guessing. If a thread gets flagged as potentially important but the confidence is low, the owner gets a one-line summary and a direct question: "Respond now, delegate, or snooze?" The escalation comes through whatever channel the owner monitors most — Slack, SMS, or a dedicated daily digest email. The goal is fewer decisions, not zero visibility.
Golden Horizons builds this as a scoped integration against the owner's existing Gmail or Microsoft 365 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 requiring a decision by noon, drafts three replies, and buries the rest. Partner spends 20 minutes on email instead of two hours.
- A founder at a consulting firm leaves for a four-day conference. Bot triages the inbox in the background, escalates two time-sensitive client asks via Slack, and queues twelve drafted responses for review on return. Nothing falls through.
- A busy CEO at an owner-operated services company keeps missing vendor follow-ups buried under client newsletters. Bot auto-labels and archives the newsletters, flags vendor threads URGENT, drafts a single reply template the CEO approves and clones across three open threads.
- A solo CPA at tax season with 300 unread messages. Bot identifies eight client document requests, drafts follow-up emails with specific document lists, and auto-archives IRS bulk mail and software renewal notices.
- A residential contractor owner gets subcontractor bids, client change order requests, and supplier invoices all in the same inbox. Bot separates them into labeled buckets, drafts approval or decline replies for the bids, and flags change orders that reference active job numbers.
- A government contractor BD lead misses a teaming partner inquiry buried in a thread with 40 messages. Bot surfaces the thread, summarizes the opportunity in one line, and drafts a reply that acknowledges the inquiry and proposes a call — sent within 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
Retainer
Monthly retainer covers monitoring, prompt tuning, config refinement, and minor integration additions. Range: $1,500–3,000/mo.
How clients use this
Fixed-scope build with clean handover, then an optional monthly retainer covering maintenance, monitoring, and minor changes. Most clients move to retainer within 60 days of delivery.
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, which means your email content isn't used to train any model and isn't stored after the request is processed. We use enterprise API tiers from OpenAI or Anthropic where this is a contractual term, not just a policy claim. 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 — read and draft permissions only, no send authority until you approve. If you 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 in a safety net for that period. During initial training, the confidence threshold for escalation is set low — the system escalates more than it needs to rather than less. You'll get more Slack pings and digest items than you will once it's trained, and that's intentional. As it learns your patterns, you'll start seeing fewer escalations on threads it's correctly identified as low-priority, and the draft quality tightens. If it mislabels something, there's a one-click feedback button in the digest — mark it wrong, and that signal feeds directly back into the labeling model. Most owners feel like 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 doesn't change.
Does this work with Gmail, Microsoft 365, or both?
Both. The build integrates with Gmail via the Google Workspace API and with Microsoft 365 via Microsoft Graph. If you're on Google, you'll authorize a scoped OAuth app registered to your Workspace domain. If you're on Microsoft, same process through Entra ID. Hybrid setups — where some team members are on one and some on the other — can be handled, but we scope that separately because the routing logic is more complex. The draft interface surfaces in whichever channel you already monitor: a digest email, a Slack message, or a sidebar in the inbox itself, depending on what fits your workflow. We don't require you to 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 — the system escalates aggressively, not conservatively, while it's learning. After calibration, two safeguards remain. First, any sender you've explicitly flagged as a priority contact always generates an escalation, regardless of the AI's confidence — that list is your hard-override layer. Second, every thread that gets labeled anything other than AUTO-DELETE is still in your inbox. Nothing is deleted or moved out of reach without your explicit instruction. AUTO-DELETE only fires for categories you've pre-approved — unsubscribe candidates, known marketing lists, automated platform notifications. If you're uncertain about a category, we don't put it on AUTO-DELETE during the build; we put it on NEWSLETTER and let you see the volume before deciding. The goal is to narrow your decision surface, not to create 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 comes with an edit interface and a feedback flag. When you edit a draft before approving — changing the tone, correcting a fact, removing a sentence — those edits are logged as training signal. Over time, the draft style tightens toward how you actually write, not how you wrote six months ago. If the bot consistently misses something structural — say, you always close client emails with a specific next-step format it keeps omitting — you flag it once and we add it as an explicit instruction in the system prompt. The retainer covers prompt updates like that as your communication style or business context shifts. Draft quality isn't fixed at build completion; it improves as long as you're giving it signal.