The four ways to get an AI agent built
Companies in Vietnam usually choose between four routes. Each one moves a different part of the risk to you.
A freelancer is the fastest way to start and usually the cheapest per hour. A studio or agency costs more but gives you a team, a contract and someone responsible when the original developer is away. A SaaS vendor sells you a product with an agent inside it: you configure, they build and host. An in-house hire gives you the most control and the slowest start.
Who carries what| Option | Good for | What you carry | Watch out for |
|---|
| Freelancer | A clear, small task with one system | Continuity, security review, hosting | No one to call when the person is busy or leaves |
|---|
| Studio or agency | Agents that touch several systems or customer data | Choosing a team that writes the code itself | Sales people promising what engineers have not checked |
|---|
| SaaS vendor | A standard job inside one tool, such as support tickets | Data stored with the vendor, limited custom logic | Per-message or per-seat fees that grow with volume |
|---|
| In-house hire | Companies that will run several agents for years | Recruiting, salaries, keeping skills current | Months before the first agent works |
|---|
What should decide your choice
Start from what the agent will be allowed to do, not from the price per hour. An agent that only drafts replies for a person to send is a low-risk project. An agent that creates invoices, changes order status or replies to customers on your Zalo OA without review needs careful engineering, logging and a clear owner.
Then look at your systems. If the agent must read from KiotViet, MISA, HubSpot or an internal database, the developer needs to work with those APIs or exports. Ask early whether they have connected to the specific system, and treat any answer as something to confirm on a technical call.
- Low risk: drafts, summaries, internal search — a freelancer or SaaS tool can work
- Medium risk: updates records a person reviews later — a studio with a test plan
- High risk: actions with money, contracts or personal data — a team with permissions design, logs and a rollback plan
What changes when the agent works in Vietnam
Your developer should handle Vietnamese and English in the same flow, because customers and staff often mix both. Local businesses often run customer contact through Zalo, so an agent that answers customers usually needs a Zalo Official Account connection, not only a website widget.
Accounting and electronic invoices (hóa đơn điện tử) follow local rules and local software. A developer from abroad may build a capable agent and still miss how your accountant expects documents to look. Ask who on their side has worked with Vietnamese invoices, contracts or bank transfers before.
Personal data of customers and employees falls under Vietnamese data protection rules. A serious developer will ask where data is stored and who can see it; the exact obligations you confirm with your lawyer.
Questions to ask on the first call
Good answers are specific and mention your systems by name. Vague answers about "AI magic" are a signal to keep looking.
- What exactly will the agent be allowed to change, and how do you stop it from doing more?
- Which steps need a person's approval before they happen?
- How will you test it on our real data before it goes live, and what counts as a pass?
- Where is our data stored, and which AI model provider receives it?
- What happens when the model gives a wrong or uncertain answer?
- Who owns the code, prompts and configuration after the project?
- Who answers when something breaks on a weekend, and how?
- How do we switch the agent off in one step if needed?
Red flags in proposals
Be careful with a proposal that promises a fixed share of work replaced before anyone has looked at your process. Nobody can know that without your data.
Other warning signs: a demo built only on public examples, no mention of logging, no plan for wrong answers, and a contract where the vendor keeps the code or your prompts. Also be careful if the developer wants full admin access to every system on day one; an agent needs the narrowest access that does its job.
A low-risk way to start
Run a short paid scoping phase before you commit to a full build. It should end with a written description of the task, the systems involved, the permissions the agent needs, a test set from your own data and a fixed quote. If the developer refuses to write this down, the build will be harder to control.
Our own version of this step is the AI audit: 2 weeks, fixed price $900, credited to the implementation if you continue. Any good team will offer something similar under a different name; compare what each one delivers at the end of it.
Whoever you hire, keep the admin passwords, API keys and the hosting account in your company's name.