A common pattern in AI customer support automation pitches is connecting a language model to a static FAQ document, embedding a chat widget, and promising instant cost reduction. In production, FAQ bots break down the moment a buyer asks about live stock, district-level delivery, or changing an active booking.
That is why the public case study from Ho Chi Minh City self-storage operator MyStorage and its investor Emerging Markets Investment Advisers (EMIA) stands out. Summarized on MyStorage’s blog on 4 November 2026 (updated 12 November 2026, covering an October 2026 report), it gives rare operational numbers from a Vietnamese business running an AI enquiry system in production.
To be clear: MyStorage is not an Insonl client. What follows is an engineering teardown of their public data—separating what the numbers prove, what they leave unmeasured, and what Vietnamese-language realities operations leaders must plan for before building something similar.
What MyStorage built—and what the published numbers say
Self-storage carries an immediate pre-sales bottleneck: buyers do not think in cubic meters (CBM). As MyStorage’s write-up explains, customers describe belongings as everyday items—a sofa, 20 cartons, two bicycles, or the contents of a two-bedroom apartment. Sales advisers previously exchanged repeated messages to estimate volume, check vacant units, quote rates, and follow up on bookings across languages.
Rather than deploying an isolated FAQ bot, three in-house MyStorage engineers built an assistant called STOW wired directly into the warehouse management system (WMS), live vacant-unit inventory, and customer care workflow. According to the case study, STOW supports four languages, translates item lists into unit sizes, quotes prices, follows up on unfinished bookings, edits and updates bookings directly in the system, and hands conversations to staff for negotiations, complex cases, or explicit human requests.
For Q2 2026, MyStorage and EMIA reported four core metrics:
| Metric | Reported figure (Q2 2026) | What it proves | What it does not show |
|---|---|---|---|
| Median first response time | 3m40s → 42 seconds | Direct system lookups cut queue wait time sharply | Breakdown by channel, language, or after-hours traffic |
| Outgoing replies automated | 56% automated | Automation settles near half; 44% stays human | Error rate, escalation quality, or re-contact rate |
| Staff hours saved | 89 hours/month (~1,784 replies × ~3 min) | Auditable arithmetic grounded in real reply volume | Net savings after engineering upkeep and QA review |
| Order conversion rate | ~4% → ~7% (vs. manual year prior) | Conversion rose year-over-year alongside rollout | Controlled attribution separate from fourfold storage capacity growth |
What MyStorage’s numbers actually prove
Three engineering conclusions hold up cleanly.
1. Live operational integration beats static FAQ bots
STOW delivers speed because it queries real vacant units and writes booking updates straight into the WMS. In one documented interaction from the case study, a customer enquired about a 16 m³ unit, balked at cost, and STOW reviewed the item list to recommend a 3 m³ to 5 m³ unit at a closer location—then updated the booking in seconds. Without read/write access to inventory, pricing rules, and your CRM, an AI assistant can only collect contact details and tell the buyer to wait.
2. Hybrid human handoff is the real steady state
Even with three in-house engineers building and maintaining STOW, the assistant generates 56% of outgoing replies—leaving 44% to human advisers. That split is healthy, not a shortfall. Routine sizing, availability checks, and follow-ups fit software; B2B accounts, site surveys, negotiation, and edge cases belong with people.
3. The labor math is transparent enough to benchmark
MyStorage publishes the underlying formula: roughly 1,784 automated outbound messages per month at an average saving of about 3 minutes per message yields roughly 89 hours saved per month. Any operations director can apply the same calculation to their own monthly message count before scoping a build.
What the published numbers do not show
Before using this case to justify an AI support budget in Vietnam, examine four gaps in the published report.
1. Conversion lift is confounded by business expansion
The ~4% to ~7% conversion increase compares Q2 2026 against fully manual interactions the previous year—not a simultaneous A/B holdout. Over that same window, MyStorage’s post notes that EMIA’s ASEAN Frontier Markets Fund invested growth capital in October 2025, after which MyStorage expanded storage capacity fourfold to seven self-storage facilities and one pickup-and-delivery facility by late 2026. More locations, wider unit availability, and shifts in marketing mix all move conversion independently of response time, and the case study does not isolate what share of the 3-percentage-point change came from faster replies versus business growth.
2. Engineering and operating cost is undisclosed
Three in-house engineers built STOW, wired it to the WMS, and continue extending it with photo-based sizing and after-sales automation. Neither engineering cost, per-conversation inference cost, nor project payback is published, so outside readers cannot determine net financial return from the 89-hour monthly labor saving alone.
3. Failure rates and bad quotes are unreported
The write-up highlights two successful customer conversations but publishes no error rate for misestimated volumes, incorrect quotes, failed handovers, or customer complaints. In commercial support, one hallucinated discount or broken booking update costs more trust than rapid replies recover.
4. Channel and language performance are not split
According to VNG’s Q3 2026 earnings release (31 October 2026), Zalo reached 82.6 million monthly active users and 2.2 billion daily messages, showing the scale of chat platforms Vietnamese businesses commonly support alongside web chat, phone, and other messaging apps. MyStorage’s website exposes Zalo, WhatsApp, phone, and web booking, and STOW supports four languages, yet the summary does not disclose enquiry share by channel, name the four languages, or split automation performance across channels or languages.
Vietnamese-language engineering realities: lessons from Review Nguoi Yeu
English demos hide how messy live Vietnamese customer text behaves. Insonl’s own documented AI build—Review Nguoi Yeu, a Vietnamese user-generated relationship-review platform built with Next.js and a modular Nest.js backend—forced us to deal directly with unfiltered Vietnamese submissions.
As documented in our moderation write-up, every submitted post passes through a pre-publication AI screening pipeline checking for toxicity and hate speech, personally identifiable information (full names, addresses, phone numbers, and identifying details), and safety threats, alongside AI summarization of longer posts. Because automated models inherently carry false-positive and false-negative risk when reading nuance and intent, the system is designed for pre-publication AI screening while allowing review of flagged content and ongoing model tuning rather than assuming automated checks catch every edge case.
Applying language models to Vietnamese customer enquiries introduces four operational hazards that teams must test explicitly:
- Unaccented text (không dấu) and mobile shorthand. Customers on Zalo routinely drop diacritics and use abbreviations (“ko”, “dc”, “ib”, “bn”). Prompts and catalog search tested only on clean Vietnamese sentences degrade sharply on real chat logs.
- Regional phrasing and Vietnamese–English code-switching. Buyers mix Northern and Southern vocabulary with English SKUs, model codes, and technical terms in the same message.
- PII and screenshot-heavy threads. Vietnamese buyers routinely paste phone numbers, ward-and-district addresses, and bank-transfer screenshots directly into chat. Any generative AI system calling external model APIs requires explicit PII handling, access controls, and log retention rules—the same safety discipline required on Review Nguoi Yeu.
- Deterministic grounding for prices and stock. Never let a model quote prices or policies from memory. Ground every number in live tool calls to your ERP or CRM, and escalate immediately whenever a buyer asks for custom terms.
When enquiry automation does not pay off in Vietnam
Skip customer-facing AI chat when any of these conditions hold:
- Low monthly enquiry volume. If saving a few minutes on roughly half of outgoing replies frees too few staff hours to offset integration and maintenance costs, keep replies human.
- No structured source of truth. If prices, stock, and customer history live in personal Zalo accounts and ad-hoc spreadsheets rather than a CRM or ERP, build the core operational system first.
- High-negotiation B2B deals. When every enquiry requires bespoke scoping or custom contract terms, customer-facing bots add friction. Build internal search tools for your human reps instead.
Audit real chat transcripts before committing engineering time
Before building an AI assistant for Vietnamese customer enquiries, pull a batch of real historical transcripts across Zalo, web chat, and messaging channels, measure where staff lose time looking up answers, and verify what APIs your inventory and CRM expose.
Insonl runs a free AI workflow audit from our studio in Ho Chi Minh City. We inspect your enquiry workflow and underlying enterprise systems, then deliver a written workflow map, honest recommendations (including where not to use AI), and a fixed-scope estimate. Follow-on Map, Build, or Hold work is optional.
Sources
- MyStorage & EMIA (4 Nov 2026, updated 12 Nov 2026), MyStorage And EMIA Case Study: STOW AI Assistant Answers In 42 Seconds: https://mystorage.vn/mystorage-emia-case-study-stow-ai-assistant/
- VNG Corporation (31 Oct 2026), Thông cáo báo chí — Kết quả kinh doanh Quý 3/2026 (VNG Q3 2026 Earnings Press Release): https://vng.com.vn/news/press-release/vng-fs-q3-2026.html
- Insonl, Free AI workflow audit
- Insonl, Review Nguoi Yeu case notes
- Insonl, AI Moderation for Vietnamese UGC on Review Nguoi Yeu
- Insonl, Enterprise Solutions (ERP, CRM, BI)
