Building an AI Executive Assistant That Runs a 50-Person Company
A Perth-based commercial services company needed to scale operations without scaling headcount. We built an AI system that handles email management, financial operations, HR workflows, property management, and client communications — running autonomously, 24 hours a day.
60+
Automated workflows
6
AI sub-agents
80+
Hours saved / month
24/7
Uptime
The Problem
The company had grown from a small operation to managing over 50 staff across multiple service sites, several Airbnb properties, and a complex roster of client contracts. The managing director was spending upwards of 20 hours per week on administrative work that followed the same patterns every day: sorting emails, chasing invoices, updating schedules, coordinating staff, managing property bookings.
The existing toolset — Microsoft 365, Xero for accounting, ServiceM8 for job management — worked well individually but had no connective tissue between them. Information flowed through manual copy-paste operations. A client email requesting a schedule change meant updating ServiceM8, notifying the relevant staff, adjusting the Xero quote, and sending a confirmation — four separate systems, four manual steps, every time.
During peak periods, the backlog of administrative tasks would cascade: late invoices led to late payments, which led to time spent chasing them, which meant less time for business development. The company needed operational leverage — not more staff, but better systems.
What We Built
We designed and deployed a comprehensive AI operations layer — an autonomous system that sits between the company's existing tools and handles the orchestration that humans were doing manually. The architecture uses N8N as the workflow engine, Claude as the reasoning layer, Supabase as the memory and state management system, and Microsoft Graph, Xero, and ServiceM8 APIs as the integration points.
Email Intelligence Agent
Monitors three Microsoft 365 mailboxes continuously. Classifies every inbound email by intent (new enquiry, schedule change, complaint, invoice query, spam). Routes actionable items to the correct sub-agent for processing. Drafts context-aware replies using conversation history and client records. The system handles approximately 200 emails per day without human intervention — escalating only edge cases that fall below a confidence threshold.
Finance Sub-Agent
Connected directly to Xero via API. Generates invoices automatically when jobs are completed in ServiceM8. Reconciles incoming payments against outstanding invoices. Flags overdue accounts and sends tiered follow-up reminders — polite at 7 days, firm at 14, escalation alert at 30. Handles payroll data preparation and expense categorisation. Before this agent, invoice generation alone took 3–4 hours per week.
HR & Staff Coordination Agent
Manages the staff onboarding pipeline from application to first shift. Collects and validates documents (Working with Children checks, police clearances, visa status). Coordinates roster changes and notifies affected staff via SMS (Twilio integration). Tracks leave requests and calculates coverage requirements. Handles candidate screening for new applications using structured evaluation criteria.
Airbnb Property Management Agent
Fully integrated with the Airbnb API. Monitors bookings, coordinates cleaning schedules between guest checkouts and check-ins, manages guest communications (check-in instructions, local guides, issue resolution), and handles review responses. This agent manages multiple properties and ensures zero-gap turnarounds — when a guest checks out at 10am and the next checks in at 3pm, the cleaning crew is automatically dispatched and confirmed.
Operations Agent
The backbone of daily coordination. Monitors ServiceM8 job statuses, alerts management to overdue or at-risk jobs, tracks quality inspection results, and generates daily operational summaries. Integrates with the communications agent to ensure clients receive proactive updates about their service schedule.
Research Agent
Handles ad-hoc information gathering requests. When the managing director needs competitor pricing, compliance requirements for a new contract, or background on a prospective client — the research agent searches, synthesises, and delivers structured briefs. Saves approximately 5–8 hours per month of manual research time.
The Memory Architecture
What separates this from a collection of simple automations is the shared memory layer. Every agent reads from and writes to a centralised Supabase database with over 170 tables — including episodic memory (what happened), semantic memory (what things mean), procedural memory (how to do things), and entity memory (who people are and their relationships).
When a client emails about an issue, the email agent doesn't just classify the email — it pulls the full client history, recent job records, outstanding invoices, and any previous complaints. The response draft includes context that would take a human 10 minutes to gather. This memory architecture means the system gets more effective over time as it accumulates institutional knowledge.
Technology Stack
Results
80+ hours per month recovered — administrative time that the managing director now spends on business development and client relationships instead of email triage and invoice chasing.
Invoice processing time cut by 90% — from 3–4 hours per week of manual Xero data entry to fully automated invoice generation triggered by job completion events.
Email response time under 5 minutes — for routine enquiries that previously sat in an inbox for hours. Complex emails are drafted for human review within minutes.
Zero missed Airbnb turnarounds — automated coordination between bookings, cleaning crews, and guest communications eliminated the scheduling gaps that previously caused guest complaints.
Staff onboarding pipeline reduced from weeks to days — automated document collection, validation, and follow-ups mean new hires move through the pipeline without manual chasing.
Key Lessons
Start with the highest-pain workflow, not the easiest one. The instinct is to automate simple tasks first. We started with email management — the most complex but also the biggest time sink. The ROI from that single agent justified the entire project within the first month.
Memory makes the difference between automation and intelligence. Without the shared memory layer, each workflow would be stateless — unable to reference previous interactions or learn patterns. With it, the system contextualises every action using the full history of the business.
Build for graduated autonomy.Every agent started in “draft mode” — suggesting actions for human approval. As confidence and accuracy were proven, autonomy was gradually increased. Today, approximately 85% of actions execute without human review.
Frequently Asked Questions
How long did it take to build this AI executive assistant?
+
What did this AI executive assistant system cost to build?
+
Does the AI replace human staff?
+
What tools does the AI executive assistant integrate with?
+
Can this pattern work for businesses outside commercial services?
+
Want an AI operations layer for your business?
Every system is different. Book a call and tell us what's eating your time — we'll tell you exactly what we can automate and what it would cost.
Book a Free Call →