AI agent, chatbot or automation: which one your business actually needs

Almost every software proposal in 2026 contains the word AI, and in most of them it is covering for one of three quite different things. A chatbot answers questions. An agent takes actions. An automation removes a repeated task without anyone talking to it at all. They cost different amounts, fail in different ways, and solve different problems. This article is the plain version of the distinction, written so you can tell which one you are being sold.

The three things, in one table

The useful question is not which is most advanced. It is how much authority the thing needs in order to be worth having. A chatbot needs none. An agent needs real authority, which is why it needs real logging. An automation needs narrow authority over a task you can describe in a sentence.

When a chatbot is the right answer

A chatbot earns its keep when you are answering the same twenty questions all day and the answers live in documents. A diagnostics centre asked the same thing fifty times a day: what are your timings, do you do this test, do I need to fast. None of those require an action. All of them have a document behind them.

The test for a chatbot is whether a correct answer ends the interaction. If the customer then needs something done, a chatbot is half a product, and the half you have built is the half that was already on your website.

When you actually need an agent

An agent is worth the extra engineering when the conversation has to end in a change. Booking an appointment. Checking a specific invoice's status. Recording a complaint against a real order. These need the model to call into your systems, which is a different class of software: it needs authentication, it needs permission boundaries, and it needs a record of what it did.

We built a multi-tenant voice agent platform where each tenant declares the tools its agent is allowed to invoke, and the platform executes them against that tenant's own endpoints. The interesting work was not the conversation. It was deciding what the agent may do, and keeping a transcript so an action can be traced afterwards. You can read how that was put together in the case study.

If it can change a record, it needs a log. If it cannot, it is a chatbot.

When the answer is automation, and nobody needs to talk to anything

This is the most commonly missed option, and usually the cheapest. A great deal of what gets proposed as AI is a scheduled job with a language model bolted to the front of it.

A pathology technician was typing roughly twenty numbers off an analyser slip into a report, around eighty times a day. The fix was not an AI assistant. It was middleware that listens to the analyser and files the result against the right sample. No conversation, no model, no prompt. The job simply stopped existing.

Ask the question this way: does a human need to be in this interaction at all? If not, you want automation, and you should be suspicious of anyone who proposes a chat interface for it.

How the three compare on cost

Rough shapes, not quotations, because every project differs.

• Automation has the lowest ongoing cost. Once written it runs on a server you already pay for, and a well-scoped one is days of work, not months.
• A chatbot has a low build cost and a per-question running cost. You pay the model each time someone asks something, so a popular chatbot is a recurring bill.
• An agent has the highest build cost, because the integrations and the permission model are the work, and a per-conversation running cost on top. Voice agents add telephony and speech on top of that.

The cost that surprises people is the running cost of conversation. A build is a one-time number you can plan around; a per-conversation charge scales with your success. We go through that arithmetic in what an AI voice agent costs to run in India.

A short decision procedure

1. Write down the task as one sentence, naming who does it today.
2. Ask whether a human needs to be in the interaction. If no, build an automation and stop here.
3. If yes, ask whether a correct answer ends it. If yes, a chatbot over your own documents is enough.
4. If the interaction must end in a change to a record, you need an agent, and you need to list the actions it may take before anyone writes a prompt.
5. Whatever you build, decide now what happens when it is wrong, and who sees that it was wrong.

The rule we apply to our own products

We add a language model where it removes typing or reading, not where it adds a chat box. Healthixio turns a doctor's voice note into a draft discharge summary, which a doctor then approves; the model removes the typing and a human keeps the authority. Every AI feature we ship has a plain fallback for when the model is unavailable, and a log of what it did.

That last part is not a compliance detail. An AI feature with no log is a feature you cannot debug, cannot defend and cannot improve.

Frequently asked questions

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

A chatbot answers questions and changes nothing. An AI agent can take actions in your systems during the conversation, such as booking an appointment or updating a record. That difference in authority is why an agent needs authentication, explicit permission boundaries and a transcript, while a chatbot mainly needs good source material.

Does my business need AI or just automation?

Ask whether a human needs to be in the interaction at all. If a task is repeated on a trigger or a schedule and nobody needs to converse with it, plain automation is cheaper to build, cheaper to run and far less likely to fail in an embarrassing way. Reserve AI for where language itself is the problem: reading documents, drafting text, or speaking to people.

How much does an AI agent cost to build in India?

The build cost sits mostly in integrations and the permission model rather than the model itself, so it scales with how many systems the agent must touch and how much authority it is given. The part buyers underestimate is the running cost, which is charged per conversation and therefore grows with usage. Ask any vendor for both numbers separately, and ask what the cost per conversation looks like at ten times your current volume.

Can an AI agent work in Hindi and English?

Yes. Current speech and language models handle Hindi, English and code-mixed Hinglish well enough for front-desk work, which is how most Indian customers actually speak. What matters more than the language list is testing with recordings of your own callers, because accent, background noise and domain vocabulary affect accuracy far more than the advertised language support.