- Generative AI makes text, images or tables from an instruction, then stops. A person takes it from there.
- An AI agent takes one job, looks things up, opens your systems and finishes it, within the tools and permissions you grant.
- Agentic AI is a system that plans, splits the work and coordinates many steps and systems, inside limits and approval points people set.
- Get the first level right before you climb. Each level up needs more permissions and more people who own the outcome.
The three differ on who does the next step
Think of three new hires. The first drafts a good letter and leaves it on your desk. The second takes one task, say checking when an order will arrive, and answers the customer end to end. The third runs a team: takes a big goal, breaks it into pieces, hands them out, follows up, and brings you only what exceeds their authority. Those three are generative AI, an AI agent and agentic AI. Vendors draw the lines differently. This article uses the definitions development teams use when they design a system.
Generative AI makes something new from an instruction: text, an image, a table or code. The engine inside is an LLM.
An AI agent is software that takes one goal and decides for itself where to look, which system to call and how many steps to take until the job is done, within the tools and permissions you grant. The mechanism that lets it open a system is called tool calling.
Agentic AI is a way of designing the whole system so that AI plans, splits the work, coordinates several agents or systems and decides the next step on its own inside limits people set. The layer that coordinates everything is called orchestration. Agentic describes the system; an agent is one part of it.
Generative AI makes an answer and stops
It suits work a person already checks before use, such as drafting a reply to a customer on LINE OA (LINE Official Account, the business messaging channel most Thai companies sell through) for a staff member to send. The result is still a draft waiting for a human.
What you must give it is a clear instruction and correct company documents. The common method is RAG: the system searches your documents before it answers. Where it fails: the model answers confidently without data, such as quoting a delivery date when it has never seen a truck schedule, and it does not know today's stock or outstanding balance.
An AI agent takes one job and finishes it
An agent works in a loop: take the goal, pick a tool, act, look at the result, then decide whether to continue or stop and ask a person.
The first jobs Thai companies usually hand over are the ones where the question repeats but the answer needs a system opened. An agent answers an order-status question on LINE OA by finding the order number in the ERP and replying with the delivery date and tracking number. Or it reads a Thai-language invoice from email, matches it to the purchase order and queues a payment for someone to approve. Both open only a few systems and can be reversed if wrong.
What you need to add is an API into the ERP, CRM or accounting software, the narrowest permissions the job needs, a clear approval point, and real examples from your own work to test it with before it goes live. Where it fails: calling a tool with the wrong data, looping, holding more permissions than the job needs, and customer messages that secretly instruct the system, known as prompt injection. Read more in AI employee.
Agentic AI: many steps, many systems, people holding the limits
Agentic AI is a system that takes a large goal, such as "close this order from purchase to payment", breaks it into pieces, hands each to the agent or system that does it best, follows up, and escalates anything outside its limits to a person.
The examples are usually cross-department. Order to cash: one agent takes the order from LINE OA, one checks the credit limit, one books the truck slot, and a step issues the e-Tax Invoice (Thailand's electronic tax invoice), with any order over the limit waiting for human approval and every step written to an audit log showing what it did with which data.
The failures differ from the level below. One agent's mistake flows into the next with nobody watching, and when the result is wrong, nobody owns it because "the system did it". Customer personal data also flows through several systems. You must be able to say at which step it leaves the company and whether that meets PDPA, Thailand's personal data protection law. If it cannot leave, the option is a private LLM running inside your own environment. Read more in Multi-agent business platform.
| Generative AI | AI agent | Agentic AI | |
|---|---|---|---|
| What it does | Makes text, images, tables or code, then stops | Takes one job, calls systems, finishes it | Plans, splits the work, coordinates many agents and systems |
| What it needs from you | A clear instruction, correct documents | APIs, narrow permissions, approval points, real examples to test with | A process map, an audit log, a stop button, someone who owns the outcome |
| Where it fails | Answers confidently without data, does not know today's state | Wrong tool inputs, loops, permissions too wide, prompt injection | Errors cascade, costs climb, nobody owns the result |
| A Thai business example | Drafts a LINE OA reply for staff to send | Answers order status from the ERP, reads invoices and queues payments | Order to cash through credit check and truck booking |
| Where a plain chatbot sits | Here, if it only answers from documents | Becomes an agent the day it opens a system to act | One window into a larger system |
Hypothetical case
A building-materials wholesaler in Samut Prakan (name and details invented) takes orders from retailers over LINE OA, dozens of messages a day. The sales team answers the same questions all day and keys the orders into the system by hand. Stage one: the company has generative AI draft replies from the catalog, and staff read and send. "Is it in stock?" still needs someone to open the system, because the model does not know today's stock.
Stage two: the development team connects an agent to the stock and order systems so it answers stock questions and creates a draft purchase order for staff to confirm. The first failure: the agent confirmed a delivery date without knowing the trucks were full. The team added a truck-slot tool and a rule that delivery dates come only from the truck schedule. Stage three, once stage two is stable: the system coordinates the whole chain, credit check, truck booking and tax invoice, and any order over the credit limit waits for the sales manager.
The three-level ladder most businesses climb
Companies that get agents working usually take the same route, and each level has evidence to collect before the next.
The three-level ladder
- Level 1, answer accurately. The system answers from company documents; a person sends. Evidence: real questions from your own work tried before go-live, with a supervisor agreeing what counts as passing, and the share of answers people had to fix.
- Level 2, do one reversible job. The agent opens a few systems and does work that can be cancelled or corrected. Evidence: a record of every system call, and how many times people declined to approve.
- Level 3, coordinate several jobs. The system hands work between agents, with approval points wherever money is involved. Evidence: a complete audit log, a stop button that has been rehearsed, and the name of the person who owns the outcome.
Rule for moving up: climb when the number of human corrections keeps falling and someone has taken ownership. If the core systems have no API yet, put the data in one place first.
The owner's version, with finer levels of autonomy, is in Roadmap from chatbot to AI website and AI app.
When a chatbot is enough, and when agentic is too early
A plain chatbot is generative AI with a chat window, and it becomes an agent the day it opens a system to act. A chatbot is enough when the questions repeat, the answers live in documents you can maintain, and nothing has to happen in a system afterward. Move to an AI agent when something does have to happen in a system after the answer, every day, and it can be corrected if wrong.
Agentic AI is too early when the single source of truth is still scattered, the core systems have no API, nobody owns the process, or you cannot yet say how the system stops when it is wrong. How to choose the first job is in Where to start with AI.
Questions to ask a developer before you hire
Every vendor will say their system is agentic. These questions show what is actually inside. Ask for answers on a real screen rather than a slide.
- Which level of the ladder is the proposed system on, and why does our job need that level?
- Which systems can the agent call, and who grants the permissions?
- Where must a person approve first, and can we change that ourselves?
- When the system gets it wrong, can we stop it at once, and how is it reversed?
- Can we see afterward what the system did, with which data, and who approved?
- At which step does customer personal data leave the company, and how does that meet PDPA?
- If we change the AI model next year, what has to be rebuilt?
For the development team · Tool calling, orchestration and guardrails
Tool calling: the model requests a function; code on the system side validates the schema, the caller's permissions and idempotency before anything runs. Separate read tools from write tools, and route every write tool through a configurable approval policy. Orchestration: keep job state outside the model, set timeouts and retries that never repeat a completed transaction, and cap the number of loops per job. Guardrails: filter input and check for prompt injection before acting, check results against business rules after, and define the conditions that stop and call a person. The audit log stores input, tool calls, output, the approver and the prompt version.
How we build agents that work inside Thai companies
The rules an AI agent has to get right live in the heads of the people who do the job every day. DNA Maker starts by sitting with your sales, accounting and warehouse teams and writing down the process map, the decision points and where a person must approve, then says plainly which jobs should not be an agent yet.
The team that works with you is DNA Maker's own, with no subcontracting, and you see a demo every week. We build agents on LINE OA, the web or the app you already use, connect them to your ERP, CRM or accounting software, and put approval points at the steps you choose. When data cannot leave the company, we run a private LLM inside your own environment. Scope and price are fixed before work starts, and source code is delivered as scoped. Our 500+ projects include banks, hospitals and large Thai groups. Talk to us on the AI Agent & Chatbot page, or ask Dream on the chat page.
SOFTWARE ENGINEERING GLOSSARY
Software engineering glossary
| Term | What it is | A simple example | What executives should ask the development team |
|---|---|---|---|
| Generative AI | AI that makes text, images or code from an instruction, then stops | Drafts a customer reply for staff to send | Do the answers come from our documents or from the model's general memory? |
| AI agent | Software that takes one goal, uses tools and data, and finishes the job | An agent checks an order in the ERP and reports the delivery date | Which systems can this agent touch, and who granted that? |
| Agentic AI | A system in which AI plans, splits the work and coordinates several agents inside limits people set | An order-to-cash chain where a person approves anything over the credit limit | Who owns the outcome of the whole chain when it goes wrong? |
| Orchestration | The layer that coordinates many steps and systems and decides whether to continue, retry or hand over to a person | The system passes an order from credit check to truck booking | If a middle step fails, how are the completed steps reversed? |
| Audit log | A record of who or what did what, when, and with which data | A manager can see which truck schedule the agent used for a delivery date | Can we trace the whole chain afterward, and how long is it kept? |