Step 1. Choose one task, not a department
Start with a single task that repeats often, has a clear input and a result someone can check. "Read incoming supplier invoices and prepare entries for the accountant" is a task. "Automate finance" is not.
Write down who does the task today, how long it takes, which systems they open and what they type. This description becomes the specification for the agent and the yardstick for testing.
Step 2. Collect real examples
Gather a set of real cases from your own work: emails, PDFs, chat messages, records, together with the correct result for each. Include the awkward ones: scanned documents, mixed Vietnamese and English, missing fields.
These examples do two jobs. They show the developer what the agent must handle, and later they become the test set. Remove or mask personal data you do not need for the task before sharing it.
Step 3. Map systems and permissions
List every system the task touches and what the agent must do in each: read, create, update or send. Then give it the narrowest access that covers that list.
Check how each system can be reached. Some local accounting, POS and CRM tools offer an API, others only file exports or nothing. Whether a given system has a usable API is confirmed on the audit, system by system.
- Read-only where possible
- A separate account for the agent, never a person's login
- No permission to delete records
- Spending, sending documents to customers and contract changes always need a person's approval
Step 4. Design the steps and the limits
Break the task into steps the agent performs, and decide for each one whether the agent acts alone, drafts for approval or stops and asks a person. Write explicit rules for what the agent must never do.
Define what happens when the agent is unsure: if a document is unreadable or two records conflict, it should hand the case to a person with a short note, not guess. This rule matters more than the choice of model.
Step 5. Test on your own cases
Run the agent on the examples from step 2 before it touches live data. Compare its output with the correct results and agree in advance what counts as a pass for each type of case.
Test in a copy of your systems or in draft mode, where the agent's actions are prepared but not executed. Repeat the tests after every change of prompts, rules or model; a change that fixes one case can break another.
Step 6. Plan the hand-off to people
Decide who receives the cases the agent cannot handle, in which tool, and how fast they need to react. For a support agent this may be a staff member on Zalo; for a document agent, the accountant's task list.
Tell the staff what changes for them. People who know which cases the agent sends them, and why, catch errors earlier.
Step 7. Launch narrow and monitor
Start with a part of the volume, one branch or one type of document, and widen it when the logs show stable results. Keep a log of every action: input, decision, output, and who approved it.
Review the logs regularly in the first weeks and then on a schedule. Models, your prices and your processes change; an agent that worked at launch needs someone who owns it and can switch it off in one step.
Personal data in prompts, logs and model provider requests falls under Vietnamese data protection rules; where data may be stored and processed is something to confirm with your lawyer.