Case Study: How an Elevator Maintenance Company Saves 2,400 Hours Annually with an AI Agent

1. Pain — Manual dispatch turns the service desk into a bottleneck

At 8 a.m. on Monday, Nam, director of an elevator maintenance company managing 3,000 units in Ho Chi Minh City, receives 17 fault reports at once. Customers say an elevator has stopped, receptionists omit the asset code, and dispatchers must open spreadsheets, check SLA contracts, review service history, and call technicians one by one.

This rice-powered manual process consumes 10–20 minutes per request. If a dispatcher mishears an address or misses a trapped-passenger warning, the job goes to the wrong team. When a key employee is absent or burns out under the workload, the entire workflow effectively loses power.

Many companies respond by adding chat groups, forms, and data-entry staff. The result is a patched-together system, duplicate records, and mounting technical debt. Reports may look healthy while actual response times remain unchanged, producing vanity KPIs rather than operational improvement.

2. Agitate — 2,400 hours spent on work that creates no revenue

For 3,000 elevators, assume each unit generates four requests per year. The company must process 12,000 tickets. If re-entering information, checking the SLA, and finding a technician consumes 12 minutes per ticket, 2,400 working hours disappear every year.

At a fully loaded labor cost of VND 80,000 per hour, that equals roughly VND 192 million, or more than one full-time employee. It excludes SLA penalties, technicians traveling to the wrong location, repeated customer calls, and revenue lost because the existing team cannot accept more maintenance contracts.

Buying a quick chatbot only moves the bottleneck. Customers may receive faster messages, but dispatch remains blocked. That is half-finished optimization and wasted spending on a beautifully presented project that never fixes the operating constraint.

3. Solve — Deploy a controlled AI Agent in three steps

• Step 1 — Normalize incoming requests: The AI Agent receives speech-to-text calls, Zalo messages, emails, and forms around the clock. It extracts the elevator ID, address, symptoms, and contact person. SLA decisions must come from contract data rather than model guesswork.

Step 2 — Classify and dispatch: The Agent checks maintenance history, urgency, technician skills, location, and shift availability. A minimum triage function can be implemented as follows:

def triage(ticket):
    text = ticket.get('description', '').lower()
    trapped = any(x in text for x in ['kẹt người', 'trapped'])
    severity = 'critical' if trapped else 'normal'
    return {
        'asset_id': ticket.get('asset_id'),
        'severity': severity,
        'sla_minutes': 15 if trapped else 120,
        'requires_human_approval': trapped
    }

• Step 3 — Control risk and measure ROI: HimiTek places OpenClaw Gatekeeper between the Reasoner and Actuator. The stack uses 9router v0.4.66 with LiteLLM dual-instance failover, automatic API-key rotation, rate limiting, and hard budget caps such as USD 5 per month for each virtual key. Shell commands, data modifications, and other elevated tools remain locked unless explicitly whitelisted or approved. This separation prevents prompt injection from hijacking the entire VPS and stops runaway Agent loops from generating uncontrolled costs.

The Agent sends customer confirmations, updates estimated arrival times, reminds technicians, and prepares completion reports. Under the stated assumptions, automating 12 minutes per ticket saves approximately 2,400 hours annually, shortens response time by 35–50%, and lets the current team support more customers without immediately hiring additional dispatch staff.

4. CTA — Prove the savings before expanding

Do not begin with a large transformation program. Select one service area, one technician group, and 300–500 tickets for a 30-day pilot. The measurement checklist should include ticket creation time, dispatch time, SLA breach rate, repeated calls, AI cost, and staff hours saved.

HimiTek can build the pilot from existing data, restrict execution through Gatekeeper, and compare results before and after deployment. The required outcome is concrete: faster responses, less manual entry, fewer SLA penalties, and more elevators managed by each dispatcher.