I've been involved in or consulted on more than 30 enterprise AI projects in Vietnam over the past three years. About a third of them shipped. That's not unusual globally, but the failure modes here are distinct enough that I think they deserve a clear-eyed post.
This isn't a list of generic "AI project mistakes." It's a pattern I've seen repeatedly, across industries, from banks to logistics companies to government agencies.
The PoC trap
The most common failure mode isn't technical. It's organizational. A team builds a proof-of-concept that works well enough to impress leadership. Leadership approves a larger budget. The team now has to turn the PoC into a product — and discovers that the PoC was never designed to be a product.
PoCs in Vietnam AI projects typically share a few characteristics:
- Trained on curated, cleaned data that doesn't reflect production variance
- Evaluated on accuracy metrics, not on the downstream business outcome
- Deployed in a controlled environment without real user load
- Owned by one person who understands how it works
None of these are unique to Vietnam. But the combination of limited ML engineering talent and strong business pressure to show results quickly creates a pipeline where PoCs routinely make it to budget approval before anyone has asked the hard questions.
The data gap nobody admits
Vietnamese enterprise data is messier than people expect. Not because Vietnamese companies are less rigorous — the messiness is structural. Historical records are often partially digitized, in mixed formats, or split across legacy systems that don't talk to each other. OCR quality on Vietnamese text is improving but still lags English significantly. And labeled training data for Vietnamese-specific tasks is scarce.
The most common thing I hear from data teams: "We thought we had the data. Then we looked at it."
Projects fail because the data audit happens too late — after architecture decisions have been made, after models have been selected, sometimes after vendors have been hired. Doing the data audit first is non-negotiable.
The integration problem
Even when the model is good, integration kills projects. Enterprise systems in Vietnam — especially in banking, insurance, and government — are often built on legacy infrastructure that was never designed to consume ML predictions. Getting a model output into a real workflow requires API design, latency negotiation, fallback logic, and organizational alignment between the AI team and the systems team. This routinely takes longer than building the model.
The mistake is treating integration as an implementation detail. It's not. It's a design constraint that should shape what you build from day one.
The ownership vacuum
Who owns an AI system after it launches? In most Vietnamese enterprises I've worked with, the answer is unclear. The data science team built it, but they don't own the business outcome. The business team owns the outcome, but they don't understand the model. IT runs the infrastructure, but they aren't accountable for model performance.
When the model starts drifting — and it will — there's no one whose job it is to notice and respond. The system degrades quietly until someone notices the downstream metrics have moved.
How to close the gap
The projects that succeed share a few practices:
- Data audit before architecture. Before any model selection, spend two weeks looking at the actual data you'll train and run on. Not the clean version — the real version.
- Define "done" in business terms. What does success look like in the system you're replacing? An accuracy metric is not a success criterion. A reduction in processing time or error rate in a specific workflow is.
- Design for integration from day one. The model API, the latency budget, the fallback behavior — these should be specced before the first training run.
- Name an owner. Someone should be accountable for model performance in production, with the authority to trigger retraining or rollback. That person needs to understand enough about how the model works to make those calls.
None of this is exotic. But in my experience, the projects that do these things are the ones that ship.