What Is a Chatbot — Types, Costs and How to Choose in 2026

The same question arrives from inside the company every single day. A morning in the administration office disappears into questions from Thai staff about how to file a form. A salesperson spends the day tied up fielding first-contact enquiries from customers. Somewhere in that load sits the reason so many companies are now asking, in plain terms, what a chatbot actually is and whether one would help. The complication is that a chatbot in 2026 is barely the same product as the scripted bots of a few years ago. The word now covers such a wide range of technology that two people in the same meeting can use it to mean completely different things. This article starts with the definition and the underlying mechanics, works through the difference between rule-based, generative AI and RAG chatbots, looks at the market numbers and the situation in Thailand, sets out where these systems earn their keep at a Japanese-owned factory, and finishes with a way to work out which type fits your own operation.

What a chatbot is — the definition and the basic mechanics

Start with the word itself. Leave it undefined and every participant in the discussion will picture something different, and the project will be forced back to the drawing board at the requirements stage.

The definition covers any software that answers automatically in a conversational format

A chatbot is any software that takes a question posed as text or speech and returns an answer, based either on logic prepared in advance or on a trained model. The word is a contraction of chat and robot, and the same thing is sometimes called an automated conversation program or a conversational agent.

The point worth holding onto is that the word describes the format the interaction takes, not the technology underneath it. A simple program that does nothing but present a set of predefined options and a system where a large language model composes an original answer are both called chatbots, because from the user’s side both amount to asking a question and getting something back. That breadth is the single biggest reason internal discussions go around in circles.

When a board member says the company should introduce a chatbot, what they have in mind may be a free-flowing conversation with generative AI. The IT manager in the same room may be picturing a scripted menu built on the existing FAQ. Those two things can differ by two orders of magnitude in cost, timeline and risk. Simply establishing in the first meeting which type everyone is talking about removes a great deal of later rework.

Three layers make the mechanics easy to reason about

Without going anywhere near the technical detail, the inside of a chatbot can be understood as three layers. The breakdown holds regardless of which type you eventually choose.

• The interpretation layer. This takes what the user typed or the option they selected and works out what the question is about. A rule-based chatbot handles this with keyword matching or menu branching, while an AI chatbot has a language model read the meaning.
• The answer layer. Having identified the intent, this layer decides what to send back. Whether the system returns a fixed passage written in advance, searches internal documents and summarises them, or has a model generate prose is decided here.
• The escalation and logging layer. When the system cannot answer, or should not answer, this layer hands the conversation to a human. It also records what was said so the exchange can be used for improvement later.

The third layer is the one routinely overlooked when a purchase is being evaluated. Vendor demonstrations show off the first two layers almost exclusively, but whether a deployment produces any business effect is decided by how unanswered questions are captured, who they are passed to, and how that record accumulates. A chatbot with no design for this reaches the point of being quietly unused within a few months of going live.

Treating it as first-line intake keeps the decision from drifting

Put in one sentence, a chatbot’s job is to take the enquiries that a person used to handle, and hand over to the machine the ones that recur and have a settled answer. It is not a device for answering every question that anyone might ask.

Once that premise is fixed, the measures of success follow naturally. They are the share of total enquiries resolved without reaching a person, and how much context came across when a person did take over. Fluency and how natural the wording sounds, by contrast, produce no operational benefit on their own.

Types of chatbot — rule-based, generative AI and RAG

As of 2026 the types a company can realistically choose between come down to three. They are not better and worse versions of each other. The accurate way to think about it is that each is suited to a different kind of enquiry.

Rule-based chatbots have clear boundaries and behave predictably

This type runs on a prepared table of questions and answers, or on a scripted set of branching options the user works through. Because it never leaves the paths that were designed for it, you retain complete control over what it says.

The operational numbers sit at an easily understood level. According to a Resumly.ai article comparing chatbot types, a rule-based bot runs at roughly USD 50 to 150 a month on the assumption that it is hosted on a small virtual machine. Citing a 2023 Gartner study, the same article reports that 71% of organisations using rule-based bots keep their handoff rate to humans at 5% or below, which it frames as a high degree of predictability. In exchange for accepting only questions inside a defined boundary, the system resolves reliably within it.

That characteristic is why regulated industries such as finance and healthcare, where a wrong answer translates directly into risk, continue to choose rule-based systems. The same logic applies in manufacturing to anything like work rules or application procedures, where the answer is fixed and it would be a problem for the system to paraphrase it on its own initiative.

The weaknesses are a poor tolerance for phrasings nobody anticipated, and script maintenance that stays a manual job indefinitely. Every new way of asking the same question means another pattern to add, and the whole thing goes stale fast when the person who maintains it moves to another role.

LLM chatbots are flexible but escalate far more often

This type passes the enquiry straight to a large language model and has it generate an answer in prose. It copes with phrasings nobody planned for, and it needs almost no scripting work. The early experience is dramatically better than with the alternatives, which is also why it performs so well in internal demonstrations.

The picture on running cost and stability is a different one. The same Resumly.ai comparison puts an LLM chatbot at roughly USD 300 to 600 a month including API charges, assuming usage of 10,000 tokens per day, which is several times the rule-based level. Citing the same 2023 Gartner study, it also reports that 58% of LLM bots have a handoff rate of 20% or higher, and attributes the main cause to hallucination — plausible-sounding answers that are wrong.

Read from an operational standpoint, a handoff rate of 20% or higher carries real weight. It means two or more enquiries in every ten come back to a human anyway, and when they do, the user has often already been shown an incorrect answer. Correcting a mistake is frequently more work than answering from scratch.

RAG grounds the answers in your own documents

RAG, or retrieval-augmented generation, searches your own documents and databases before generating an answer, then has the language model compose its response using those search results as the basis. Because the raw material is documents that actually exist inside your company rather than the general knowledge held in the model, it suppresses the hallucination problem of a pure LLM setup while making it possible to answer questions specific to your organisation.

Applied to a manufacturing site, that difference is decisive. Questions like who to contact first when the line stops, or what the inspection interval is for a particular machine, cannot be answered correctly by a general-purpose language model no matter how capable it is. The answer exists nowhere except in internal documents. Only with RAG in the picture do these questions come inside a chatbot’s scope at all.

RAG is not a universal solution, though. If the documents it searches are not in order, it will answer confidently and incorrectly on the basis of a superseded revision of a procedure. Deciding which document is authoritative is an unavoidable piece of work before deployment, and it is not unusual for that clean-up to be heavier than the build itself.

Setting the character of the three types side by side gives the following picture.

Both the costs and the handoff rates are indicative figures drawn from the Resumly.ai comparison article, and the two handoff figures are numbers that article cites from a 2023 Gartner study. Actual cost moves a long way with user numbers, the number of integrations and real token consumption, so treat the table as a way of grasping the relative gap between types rather than as a price quote. How the individual line items stack up in a real estimate is broken down in the cost structure of a chatbot.

Hybrid architectures are becoming the mainstream in 2026

What follows naturally from all of that is an architecture that does not commit to a single type. Among corporate deployments in 2026, the hybrid pattern is becoming the mainstream — rule-based logic sets the boundaries of what the system will respond to, and an LLM or RAG handles everything inside those boundaries.

Concretely, it is a three-stage arrangement. Areas where a wrong answer is unacceptable, such as work rules and application procedures, are pinned down with rule-based branching. Free-form questions that fall outside those areas are passed to RAG. If RAG still finds no answer, the conversation goes to a person. Another way to describe it is deciding, area by area, where you want the predictability of a rule-based system and where you want the flexibility of an LLM.

The advantage of this arrangement is that high-risk questions never reach the generative AI in the first place. For a question like whether a particular machine may be shut down, or whether a particular chemical may be disposed of, the path where a model reasons its way to an answer is blocked from the outset. Meanwhile the small talk and the endless variation in phrasing get absorbed on the inside, so from the user’s point of view the conversation still feels natural.

Which type suits which enquiry, and how to sort your own enquiry mix between them, is dealt with case by case in comparing chatbot types and choosing between them.

The 2026 chatbot market and the situation in Thailand

Market figures are useful material when an internal proposal has to explain why now. How you cite them matters, though.

Market size estimates vary considerably between research firms

Several research firms publish different estimates of the chatbot market. Treating any one of them as the definitive number is a mistake worth avoiding.

A report from The Business Research Company, distributed by Research and Markets in February 2026, puts the market at USD 10.25 billion in 2025 growing to USD 13.28 billion in 2026, an annual growth rate of 29.5%. The same report forecasts USD 37.53 billion by 2030.

Mordor Intelligence, in a report current as of 30 July 2026, puts the 2026 market at USD 11.45 billion and projects USD 32.45 billion by 2031 at a compound annual growth rate of 23.15%. That report has Asia-Pacific growing at 24.71% a year, ahead of the global average, and identifies North America as the largest region with a 38.72% market share as of 2025.

A gap that wide for the same year exists because each firm draws the boundary of what counts as the chatbot market around a different set of products. When you cite either figure in an internal document, name the source and the publication date, and avoid presenting it as a single settled value. Both sources agree on the direction of travel, and for the purposes of a business case that is enough.

Sentiment in Thailand’s digital industry is improving

There is also a useful indicator for conditions inside Thailand. Nation Thailand reported on 12 August 2026 that Thailand’s digital industry sentiment index reached 48.4 points in the second quarter of 2026, up from 44.5 points the previous quarter. By sector, software came in at 51.1, digital services at 50.6 and telecommunications at 50.0, all three reaching or exceeding the 50-point mark that is treated as the threshold for favourable conditions. Five sectors were surveyed, and those three were the ones that reached that level. Rising AI demand is identified as the driver.

One thing needs stating clearly here. This index is not a statistic about chatbots. It measures sentiment across Thailand’s digital industry as a whole, and it does not directly measure demand for chatbots specifically. Present it internally as a chatbot market indicator and you will not be able to defend it when somebody later asks where the number came from.

That said, the shape of the result is informative. With the overall index at 48.4 and still below 50, the fact that only software, digital services and telecommunications among five sectors have climbed back to 50 points or above suggests that appetite for systems investment in Thailand is concentrated in particular areas. From the perspective of a Japanese-owned company here, it signals a period when proposals from local vendors will be plentiful and comparative material will be easy to gather.

Where chatbots earn their keep at Japanese-owned factories in Thailand

From here the discussion moves onto the factory floor. Once you add the conditions of a Japanese-owned plant in Thailand, the priorities come out differently than they would in other sectors.

Internal enquiry intake is where a chatbot pays off most readily

The first candidate to examine is internal, not customer-facing. Enquiries that land on administration, HR and IT satisfy all three conditions that make a chatbot effective — there are many of them, the content repeats, and the answer already exists in a document.

The questions that come up constantly at Japanese-owned factories in Thailand are things like how to check remaining annual leave, social security procedures, how to read a payslip, requesting a uniform or a locker, and the process for borrowing IT equipment. In every case the answer is written down in a regulation or a procedure manual, and yet it is faster to ask the person directly, so it gets handled verbally, and the responsible staff member’s day is shaved away in small pieces.

How to move this area over to the machine side is set out concretely in automating internal helpdesk enquiries.

Triaging first-contact enquiries from customers and suppliers

On the outward-facing side, the targets are routine communications — checking order status, asking about delivery dates, requesting that a specification or drawing be sent again. These are usually handled individually by sales staff, and they are the textbook example of work that stops dead when one person is out of the office.

Outward-facing deployment is harder than internal deployment, for three reasons. You cannot control how the other party phrases things, a wrong answer bears directly on a commercial relationship, and the content of the answer depends on live data in your core systems. The moment you want to return order or inventory information, integration with the production management system becomes a precondition, and the project stops being about a chatbot on its own. The design considerations on the customer-facing side are covered in detail in customer-facing chatbots.

Multilingual capability is worth more at a Thai site than almost anywhere

At a Japanese-owned factory in Thailand, Japanese expatriates, Thai staff and in some cases employees from Myanmar or Cambodia share the same workplace. It is not unusual for a company to run a three-layer structure in which reporting to head office in Japan is in Japanese, instructions on the floor are in Thai, and the screens of the internal systems are in English.

In that environment, the work of maintaining the same guidance in several languages is itself a permanent overhead. A chatbot with generative AI in the mix can produce responses in multiple languages from a single source text, which reduces the effort of rebuilding translated versions every time a regulation is revised.

The caveat is equally clear. Wording that touches work rules or safety cannot tolerate the variability of machine translation. The hybrid approach applies here too. Give the passages that carry legal or safety weight a fixed, human-approved version in each language, and let generation handle only the ordinary guidance around them.

Four questions that reveal what fits your operation

Choosing a type is not a judgement about which technology is superior. It follows from the nature of the work you are targeting. Answering four questions narrows the field considerably.

• Is the answer to the enquiry written down in a document? If it is, RAG becomes a candidate. If it is not, creating that document is the work that has to happen first.
• If a wrong answer is given, can the damage be undone with a correction? If it cannot, fix the boundary with rule-based logic.
• How many languages do the enquiries arrive in? Two or more and the value of generative AI rises sharply.
• Does the answer require live data from a core system? If it does, the real subject is the integration design with the production management system, and the chatbot is only one part of the whole.

The fourth is the one most often missed. When you set out to build a chatbot that tells people delivery dates, the hard part is not the conversation. It is the mechanism that reads order data safely and returns it only to the people entitled to see it. Misjudge that and you end up having spent all that time selecting a chatbot that never goes into service.

Decide how you will measure the effect before you start

We would recommend settling on what counts as success before deployment rather than after. Three measures work well in practice.

• Self-service resolution rate. The share of enquiries that finished without being handed to a person. The difference in handoff rates between types shows up directly here.
• Time to first response. Read it against what the same enquiry took when a person handled it.
• Quality of the handoff. When a case does reach a person, record whether they had to start the questioning again from scratch.

How natural the answers sound and how satisfied users say they are come after those three. Chase satisfaction from day one and the engineering effort goes into refinements that never touch the operational result.

How the rollout runs, and a view on cost

The skeleton of the process is narrowing down the target work, taking stock of the documents, running a pilot on a small scope, then widening the scope. Skip that order and start with a company-wide launch, and you will discover afterwards that the documents supposed to underpin the answers were never assembled, at which point the system you built has to be rebuilt along with them.

On cost, as set out above, running costs differ substantially by type. The indicative figures — roughly USD 50 to 150 a month for a rule-based system, roughly USD 300 to 600 a month for an LLM system including API charges — cover the operating phase only. Initial build, integration with existing systems and the effort of getting documents in order all have to be budgeted separately. In a real approval process it is these line items, the ones that do not show up prominently on a quotation, that determine the total.

The sequencing and the way to think about cost when multilingual operation is a given are set out concretely in the cost and rollout of a chatbot deployment.

FAQ

What is the difference between a chatbot and an AI chatbot?

Chatbot is the broad term for any software that responds automatically in a conversational format. AI chatbot refers to the subset that uses machine learning or a large language model to interpret the input and produce the response. A rule-based system that runs purely on branching written in advance is still a chatbot, but it would not normally be called an AI chatbot. Operationally, the Resumly.ai comparison article puts rule-based systems at roughly USD 50 to 150 a month with 71% of the organisations using them keeping handoff rates at or below 5%, against roughly USD 300 to 600 a month for LLM systems with 58% of bots at a handoff rate of 20% or higher. The choice is between flexibility and predictability.

How much does it cost to deploy a chatbot?

As an indication for the operating phase, a rule-based system runs at roughly USD 50 to 150 a month and an LLM system at roughly USD 300 to 600 a month including API charges. That is the cost of the response layer alone. The real total is set by the initial build, integration with the production management and HR systems already in place, and the effort of preparing the documents that will underpin the answers. Document preparation in particular rarely appears on a quotation, is hard to estimate, and tends to stay on the customer’s side of the line as an internal burden. When you compare bids, do not stop at the monthly cost of the response layer — line the proposals up on the number of integrations and on where the responsibility for document preparation sits.

How do you prevent hallucination in a generative AI chatbot?

There is no way to eliminate it entirely at present. The practical response has two stages. First, combine RAG so that answers are grounded in your own documents rather than in what the model happens to remember. Second, pin down the areas where a wrong answer is unacceptable with rule-based branching so those questions never reach the generative AI at all. The reason hybrid architectures are becoming the mainstream in corporate deployments in 2026 is that they implement exactly this two-stage defence as a structure. On top of that, always provide a route to a human for the cases the system cannot answer, and design it so the system is never forced to produce an answer.

Summary

• A chatbot is any software that responds automatically in a conversational format, and the word describes the format rather than the technology inside, so internal discussions have to establish first which type everyone means
• The mechanics are easiest to reason about as three layers — interpreting the input, deciding the answer, and escalating and logging — and operational success is decided by the third layer, which is the one that gets overlooked
• Rule-based systems run at roughly USD 50 to 150 a month with 71% of the organisations using them holding handoff rates at or below 5%, a level of predictability that suits regulated industries such as finance and healthcare
• LLM systems run at roughly USD 300 to 600 a month including API charges, with 58% of bots showing a handoff rate of 20% or higher, and the main cause is hallucination
• Adding RAG suppresses hallucination and makes answers grounded in your own documents possible, and as of 2026 the mainstream is a hybrid in which rule-based logic sets the boundary and an LLM or RAG works inside it
• Market size estimates vary between research firms, with The Business Research Company putting 2026 at USD 13.28 billion and Mordor Intelligence at USD 11.45 billion, so neither should be cited as a settled figure
• The indicative costs and handoff rates come from the Resumly.ai comparison article, with the handoff figures originating in a 2023 Gartner study, so carry the source and the year across whenever you copy them into an internal document
• Thailand’s digital industry sentiment index improved to 48.4 points in the second quarter of 2026, but that measures the digital industry as a whole rather than chatbots specifically
• At Japanese-owned factories in Thailand, first-line intake of internal enquiries is where a chatbot pays off most readily, and the multilingual working environment raises the value of generative AI further still
• The choice of type narrows down on four points — whether the answer sits in a document, whether a wrong answer can be undone with a correction, how many languages enquiries arrive in, and whether live core system data is required

Chatbots are one of the areas where picking a product before deciding the target work guarantees a rebuild. We are happy to talk at the exploratory stage — which enquiries are worth moving to the machine side, or how far a new system would need to connect to the production management and HR platforms you already run. We also get a lot of questions about operating in an environment where Japanese, Thai and English are all in use at a Thai site, and those conversations work best when they start from your actual workflow. Tell us what your enquiry load looks like today and we will map out together how much of it can move and in what order. Get in touch through our Contact Us page.

References

• Chatbot Market Report – Research and Markets / The Business Research Company
• Global Chatbot Market Size and Share Analysis – Mordor Intelligence
• Difference Between Rule-Based Chatbots and LLM Chatbots – Resumly.ai
• Thailand’s digital industry sentiment index improves in Q2 2026 – Nation Thailand
• Comparison and selection guide for AI chatbots – CloudCIRCUS izanai