Every business owner in Vietnam is currently being sold AI. The pitch is usually some version of "it will cut your costs" and the demonstration is usually impressive. What the demonstration leaves out is which jobs this technology is genuinely reliable at, which it is not, and what it costs to keep running after the person who sold it to you has moved on.

This is a straight answer to all three, written for a company of twenty to two hundred people rather than for a technology department.

What it is genuinely good at

Three categories, and they have something in common: a human can check the output quickly, and a mistake is cheap.

Getting structured data out of unstructured documents

Invoices, purchase orders, bills of lading, delivery notes, scanned contracts. A system that reads a supplier invoice and produces the vendor, date, line items and total as fields is the single most reliable use of this technology in a normal business, and it attacks the most common source of data-entry labour. It is not perfect — but it does not have to be, because the output lands in a screen where somebody confirms it in seconds instead of typing it in minutes.

First drafts

Product descriptions, a reply to a routine enquiry, an English version of a Vietnamese specification sheet. The output is a starting point that a person edits. For a company producing a lot of similar text, this changes the work from writing to editing, which is faster and less tiring.

The failure case is publishing the draft unedited. On an English-language site aimed at foreign buyers this is particularly costly, because unedited machine output has a recognisable flatness, and a buyer assessing whether you are a serious supplier will notice it.

Sorting and routing

Which department does this email belong to. Is this review a complaint. Is this enquiry a real buyer or a student asking for a sample. Classification is a well-understood problem and the technology is good at it. The useful part is not the accuracy — it is that it never gets bored at four in the afternoon.

What it is not good at

The honest list, which rarely appears in a sales deck.

Anything where being confidently wrong is expensive

These systems produce an answer in the same tone whether they are right or not. There is no reliable internal signal that says "I am guessing". For a draft email, fine. For a quotation sent to a customer, a dosage, a tax figure or a contractual commitment, the absence of that signal is the entire problem. Any process where a wrong answer costs real money needs a human decision in it, and designing that human out is where most AI projects become liabilities.

Replacing a process you have not defined

If three people in your company handle returns three different ways, automating returns means picking one of those ways — and now it is encoded in software and harder to argue with. The automation did not fix the inconsistency; it froze it. Define the process first, on paper, with the people who do it. This is unglamorous and it is where the actual value is.

Understanding your business

A model knows what is in the text you give it. It does not know that this particular customer always disputes the first invoice, or that the Tuesday shipment goes out early in December. That knowledge lives with your staff, and nothing currently on the market extracts it. Automation works at the edges of that knowledge, not in place of it.

Being cheaper than a person at a task you only do occasionally

Automating something that happens five times a month is almost never worth the build and maintenance. The economics need volume and repetition. If a task is rare, let a person do it.

Cost is the wrong first metric

"It will reduce headcount" is how this is sold and it is usually not what happens, at least not first. In a company that is growing, the realistic first wins are different and in some ways better:

  • Turnaround time. Enquiries answered in an hour instead of a day. For an exporter competing against suppliers in three other countries, response speed is a real differentiator and the buyer notices it immediately.
  • Consistency. The same quality of reply at 9am Monday and 5pm Friday, from whoever is on duty.
  • Capacity without hiring. Handling twice the enquiries with the same team is usually more valuable to a growing company than handling the current volume with fewer people.
  • Fewer transcription errors. A wrong digit in a purchase order costs more than the minute it took to type.

If a vendor leads with cost reduction and cannot talk concretely about any of the above, they are selling a number rather than a plan.

Where to start, specifically

The pattern that works is small, measurable and reversible.

Pick one task you can already count

It should be repetitive, high volume, low stakes individually, and currently measurable. "We process around 300 supplier invoices a month and it takes one person about 25 hours." That sentence contains a before, which means you will be able to tell whether anything improved. Without it, you will be relying on the vendor's assessment of their own work.

Run it alongside the current process

For the first few weeks, both. The automation produces its output, the person does the job as usual, and you compare. This is the only honest way to find the error rate, and it tells you where the human check needs to sit permanently.

Decide what happens when it is unsure

Every one of these systems will encounter something it cannot handle. The question is what it does then — and the only acceptable answer is that it flags it and a person looks. A system with no defined escape route will simply guess, and you will find out months later.

Then measure the same number again

Twenty-five hours became what? If the answer is "about the same, but fewer mistakes", that may still be worth it — but you should know which of the two you bought.

What it costs to keep running

The part that gets left out of the proposal:

  • Usage charges. Most of these services bill per document, per request or per volume of text. The bill scales with your business, so model it at next year's volume, not this month's.
  • The accuracy drifts. Your suppliers change their invoice layouts. A model provider updates a version. Output that was fine in March is subtly worse in September, and nobody notices unless somebody is checking.
  • Someone has to own it. A named person who looks at the flagged items and notices when the error rate moves. Without this, the automation quietly degrades and the staff quietly go back to doing it by hand.
  • Where your data goes. If supplier contracts or customer records are being sent to an external service, that is a decision with legal weight under Decree 13/2023/ND-CP and, if you handle EU customers' data, under GDPR. Ask where processing happens, how long anything is retained, and whether your data is used to train anything. Get the answer in writing.

Questions to ask anyone selling you AI

  • What is the error rate on my documents — not on your demonstration set?
  • What happens when the system is unsure, and who sees that?
  • Where is my data processed and stored, and is it used to train your models?
  • What does this cost per month at three times my current volume?
  • Who maintains it when the accuracy drops, and is that included?
  • If I stop using you in a year, what do I keep?
  • Which parts of my process would you tell me not to automate?

That final question is the one worth weighing most heavily. Anyone who has actually deployed this work has a list of things they would not touch. Anyone who says it can all be automated has not yet watched one of these systems fail in front of a customer.

If you have a repetitive task in mind and want a straight answer on whether it is worth automating, describe it to us. Sometimes the answer is that a better form or a fixed process gets you most of the way there, which is a cheaper conversation to have before the project than after it.