Short answer: where should a business start with AI agents?
Start with one specific, repetitive workflow that has accessible data and a clear owner, not with a tool. Give the AI agent the minimum permissions it needs, put a human handover at every risky step, measure results in a pilot, and only then scale.
Rather than a “top N tools” list, this guide offers a vendor-neutral decision framework that helps leadership and engineering answer two questions together: what to do first, and how to keep it safe. For the basic concept, see what is an AI agent.
AI agent vs chatbot vs RPA: when do you actually need an agent?
In its guide “Building effective agents”, Anthropic distinguishes workflows, where language models and tools are orchestrated through predefined code paths, from agents, where the model dynamically directs its own process and tool use. The same guide recommends finding the simplest solution possible and only adding complexity when needed, because agentic systems often trade latency and cost for better task performance.
The practical consequence: not every job needs an AI agent. A process with fixed steps is better served by a workflow or RPA; an AI agent fits when inputs vary, the number of steps is hard to predict and the result can still be verified. A detailed comparison table of AI agents, chatbots, RPA and workflows is in What is an AI agent.
5 readiness questions before deploying an AI agent
Before discussing platforms or models, answer these five questions for each candidate workflow. If two or more answers are “not yet”, that workflow should not be your first AI agent project.
- Is the process documented? Can someone write the steps, exceptions and “done” criteria on one page? If not, map the workflow first.
- Who owns it? There must be one person with authority who is accountable for results and reviews the cases the AI agent escalates.
- Where is the data, and can it be reached through an API or another secure connection? Agents cannot work well with data scattered across chats and personal files.
- What does a mistake cost? Separate read-and-suggest work (low risk) from sending to customers, paying or deleting data (high risk).
- How will you measure it? Pick two or three metrics such as handling time, rework rate and handover rate, and record today's baseline.
Choosing the first workflow
The first workflow should be small enough to pilot safely but repetitive enough to show a difference. A simple scoring approach looks at four factors: frequency, how much reading and interpretation it needs, data readiness and the cost of errors. Prefer high frequency, real comprehension work, ready data and low risk.
If your first workflow sits in a back-office function, see the domain-specific guidance in AI in accounting and finance and AI in recruitment and HR.
Illustrative examples of good starting workflows
- Classifying and summarising incoming customer requests and proposing the handling team; staff approve before routing.
- Matching incoming invoices against purchase orders and flagging differences for accounting to check.
- Drafting a weekly report from internal system data; the team lead edits and approves.
- Answering internal policy questions from approved documents, with citations.
Leave for later
Irreversible decisions, actions that directly affect other people's rights, and processing of sensitive personal data should wait until the organisation has operating experience, oversight and complete logs.
Permissions and human handover
OWASP lists “Excessive Agency” in its 2025 Top 10 for LLM applications (LLM06): damaging actions can be performed in response to unexpected, ambiguous or manipulated model outputs. The usual root cause is functionality, permissions or autonomy beyond what the task needs.
OWASP's recommended controls are concrete and apply directly to a first project:
- Expose only the tools the agent actually needs; an agent that summarises email needs read access, not send or delete.
- Avoid open-ended tools such as arbitrary shell commands or fetching any URL; use narrow, purpose-built tools instead.
- Enforce authorisation in downstream systems rather than letting the model decide whether an action is allowed.
- Require human approval before high-impact actions.
- Log and monitor tool activity, and rate-limit actions so anomalies are caught early.
Designing the handover
Every workflow should define when the agent stops and hands over: low confidence, missing data, a customer asking for a person, or an action in the high-risk group. The receiving person needs full context: the input, the steps taken and the reason for handover. See human-in-the-loop and access control and audit logs for AI agents.
Preparing data for AI agents
An AI agent is only as good as the data it can reach. For a first project the goal is not a company-wide data platform, but making sure the data this workflow needs is standardised, owned and access-controlled.
- List the data sources the workflow uses: which system, who owns it, how often it changes.
- Define the authoritative version of policies, internal price lists and templates; remove outdated copies from what the agent can read.
- Tag personal and sensitive data; include it only with a lawful basis and when necessary for the purpose.
- Use a test environment with sample or masked data before running on real data.
- Agree where agent logs are stored and for how long.
The 2026 legal context: Vietnam's AI Law and Personal Data Protection Law
The Law on Artificial Intelligence No. 134/2025/QH15 was passed by the National Assembly on 10 December 2025 and has 8 chapters and 35 articles covering research, development, provision, deployment and use of AI systems in Vietnam. Under Article 34 it takes effect on 1 March 2026. Chapter II (Articles 9 to 15) covers risk-based classification and management of AI systems. Article 35 gives AI systems already in operation 18 months to comply in healthcare, education and finance, and 12 months in other sectors. Decree No. 142/2026/ND-CP (issued 30 April 2026, effective 1 May 2026) details the law: AI systems fall into high, medium and low risk levels, and providers must classify a system before putting it into use. Details: Vietnam AI Law 2025.
The Law on Personal Data Protection No. 91/2025/QH15 was issued on 26 June 2025 and took effect on 1 January 2026. Decree No. 356/2025/ND-CP (effective the same day) details the law and replaces Decree 13/2023/ND-CP. The law establishes core data-subject rights including the rights to be informed, to consent, to access, to correct and to request deletion. When an AI agent reads or writes customer or employee data, the business must identify its processing basis and how it honours these rights. See: Vietnam Personal Data Protection Law 2025.
In practice, NIST's AI Risk Management Framework (US, voluntary) organises AI risk work into four functions: Govern, Map, Measure and Manage. NIST notes that after the Map function an organisation should know enough about a system's impacts to make an initial go/no-go decision on design, development or deployment. That is a sensible way to tie an AI agent rollout to risk management; it does not replace obligations under Vietnamese law.
From pilot to scale: a roadmap
The roadmap below describes stages and exit criteria rather than fixed dates, because readiness differs between organisations.
| Stage | Goal | Main work | Exit criteria |
|---|---|---|---|
| 0. Prepare | Pick the right problem | Answer the 5 readiness questions, map the workflow, record baseline metrics, review personal data | Owner, metrics and data scope approved |
| 1. Supervised pilot | Prove value | Agent reads and suggests only; a person approves every output; full logging | Metrics beat baseline; errors classified with fixes |
| 2. Conditional automation | Reduce manual review | Allow low-risk actions; high-risk actions still need approval; monitoring and rate limits | Stable handover and rework rates |
| 3. Scale | Replicate to other workflows | Standardise permission templates, logging and evaluation; orchestrate multiple agents only when needed | Governance and periodic review in place |
| 4. Operate | Sustain quality | Watch for drift, refresh data, reassess risk when the process or law changes | Repeat the cycle |
Where EYC can help
EYC is an AI and technology company in Ho Chi Minh City working on AI agents, automation, software and systems integration. EYC is building toward an operating model in which AI agents perform most repeatable work under human direction, so the principles in this guide are the ones we apply internally. If you need help assessing a first workflow or designing permissions, see AI agents and automation or integration and consulting.
Frequently asked questions
How is an AI agent different from a chatbot?
A chatbot mainly answers conversations. An AI agent can plan and call tools to act on systems, such as looking up records, creating tickets or updating data, within the permissions it is granted.
Should a small business deploy AI agents?
It can, if it has a repetitive workflow, accessible data and a clear owner. If the steps are fixed, a simple automated workflow is usually cheaper and more predictable than an AI agent.
What data preparation does an AI agent deployment need?
For a first project, prepare only the data the workflow uses: a list of data sources and their owners, authoritative versions of documents, and clear tagging of personal and sensitive data. Add a test environment with sample or masked data, and agree where agent logs are stored and for how long.
When does Vietnam's 2025 AI Law take effect?
Law No. 134/2025/QH15 takes effect on 1 March 2026 under Article 34. AI systems already operating have 18 months (healthcare, education, finance) or 12 months (other sectors) to comply under Article 35.
How do you stop an AI agent from doing unintended things?
Following OWASP's guidance on Excessive Agency, grant least privilege and use narrow, purpose-built tools. Check permissions in the target system, require human approval for high-impact actions, and log and rate-limit tool use so unusual behaviour is caught early.