Hiring guide
AI agents, chatbots or RPA? What each one actually automates
By Samuel Marndi · Updated
The short answer
A chatbot answers questions. An agent takes actions on your systems and decides the order. RPA repeats a fixed sequence of clicks exactly, every time, and breaks when the screen moves. Those are three different purchases with three different failure modes, and the word "agent" is currently being used for all three.
Most things sold as AI agents are a chatbot holding one tool — it can look something up or create a ticket. That is genuinely useful and far cheaper than an agent, so the question worth asking a vendor is not "is it agentic" but: what can it change, and what happens when it gets one in twenty wrong?
The three, side by side
| Chatbot | AI agent | RPA | |
|---|---|---|---|
| What it does | Answers from a known body of content | Decides a sequence of steps and acts on systems | Repeats a fixed sequence exactly |
| Decides anything? | Which answer fits | Which steps to take, and in what order | Nothing |
| Changes anything? | Usually no | Yes — that is the point and the risk | Yes, within its script |
| Fails by | Answering confidently and wrongly | Taking several correct-looking steps to a wrong outcome | Stopping dead when a screen or field moves |
| Good at | Repetitive questions on real documentation | Multi-step work with judgement and messy inputs | High-volume identical work on stable systems |
| Bad at | Anything needing an action | Anything where a wrong action is expensive | Any variation at all |
| Supervision needed | Spot checks | A human approval step on anything consequential | Someone to fix it when it breaks |
| Indicative build | ₹30,000 – ₹4,00,000 | ₹4,00,000 – ₹15,00,000+ | Licence-led, per process |
The question that settles it: what can it change?
Forget the labels. Ask what the thing is permitted to alter, because that single answer tells you the architecture, the cost and the risk all at once.
If it can change nothing, it is a chatbot, whatever it is called, and it should be priced like one. If it can create a record, send a message or move money, it is an agent, and it needs the controls that go with that: a scoped permission, an audit trail, and a human approval on anything consequential. If it can only do exactly what it was recorded doing, it is RPA, and it will break the week somebody redesigns the screen.
- "It can look things up." Chatbot with a tool. Useful, cheap, low risk.
- "It can raise a ticket or book a slot." Light agent. Needs a confirmation step and an audit trail.
- "It can issue refunds, send invoices, email customers." A real agent, and the approval step is not optional. Anyone who tells you it is has not operated one.
- "It replicates what our team does in the portal." RPA, and ask what happens at the next vendor UI update.
- "It's fully autonomous." Ask who is accountable for its decisions. That question is rarely a technology problem and it stops more projects than feasibility does.
Where each one genuinely earns its money
| The work | The right tool | Why |
|---|---|---|
| Customers ask the same forty questions | Chatbot on real documentation | Known answers, no actions, cheap to run and easy to check |
| Enquiries arrive and need sorting and routing | Chatbot or light agent | Classification is what these models are reliably good at |
| Invoices arrive as PDFs and get re-keyed | Extraction, not an agent | One model call and a validation step. An agent here is over-engineering |
| A multi-step process with judgement and messy inputs | Agent, with approvals | The ordering genuinely varies, which is the one thing only an agent does |
| The same 400 records moved between two systems nightly | RPA or an integration | No judgement needed. An API integration beats both if one exists |
| Outbound phone calls at volume | Voice agent, and read the rules first | Regulated in India — see AI voice agents |
| A decision with legal or financial consequence | None of the three | Accountability cannot be delegated to a model, whatever the demo showed |
Why agents cost several times what a chatbot costs
It is not the model. The model is the cheap part and getting cheaper. The cost is in everything that has to exist around an agent before it is safe to let it act.
- Permissions, scoped narrowly. An agent with your admin credentials is not a feature, it is an incident waiting for a trigger.
- An audit trail of every action. What it did, why, on whose authority. Without this you cannot debug it and you cannot defend it.
- A rollback path. What undoes the wrong thing it just did, and how quickly.
- Evaluation harnesses. A set of cases you can re-run to know whether a model update made it worse. Nobody notices quiet degradation without this.
- Human approval gates on anything consequential, with a queue somebody actually works.
- The unhappy paths. Half the build is what happens when step three fails after steps one and two already changed something.
- Which is the honest argument for starting with a chatbot or extraction: none of that scaffolding is needed when the system cannot change anything.
What to ask a vendor selling you an agent
- What exactly can it change, and with whose permissions? A specific list, not a capability statement.
- Show me the audit log of a real run. Not the architecture diagram — the log.
- What happens when it gets one in twenty wrong? Every one of these is wrong sometimes. The answer should be a control, not a reassurance.
- Where is the human approval step? If the answer is "you won't need one", the vendor has not operated this in production.
- How will I know if a model update made it worse? The answer is an evaluation set. If there is no answer, there is no answer.
- What does it cost per month at ten times the volume? Usage scales with use, and pilots are priced at pilot volumes.
- Could extraction or an integration do this instead? A good vendor will say yes sometimes. One who never does is selling a category, not a solution.
Where to go next
If you have a list of candidate projects and no way to rank them, start with which AI projects are actually worth doing — a four-axis rubric, with eight common projects scored. It tends to put extraction first and the website chatbot well down the list, for reasons the article sets out.
If you already know what you want: AI agents has the published prices for each shape, AI chatbot and agent development covers the build, what an AI chatbot costs has the estimator and the monthly bill most quotes omit, and AI consulting is the engagement if the decision itself is the problem.
FAQ
Related questions
What is the difference between a chatbot and an AI agent?
A chatbot answers questions from a known body of content and usually changes nothing. An agent takes actions on your systems and decides the sequence itself. The test that cuts through the marketing: ask what the thing is permitted to change. If the answer is nothing, it is a chatbot whatever it is called, and it should cost what a chatbot costs.
Is an AI agent better than RPA?
Different, not better. RPA repeats a fixed sequence exactly and is excellent at high-volume identical work on stable systems — and it breaks the week somebody redesigns a screen. An agent handles variation and messy inputs because it decides the order, and it fails differently: several correct-looking steps to a wrong outcome. For 400 identical records moved nightly, RPA or a plain API integration beats an agent on cost and reliability.
How much does an AI agent cost in India?
A scripted FAQ bot is ₹30,000–₹75,000, a chat agent with one to three integrations ₹75,000–₹4,00,000, a voice agent ₹1,50,000–₹8,00,000, a retrieval agent over your own documents ₹6,00,000–₹15,00,000, and a multi-agent workflow system from ₹8,00,000. The gap between a chatbot and a real agent is not the model — it is the permissions, audit trail, rollback path, evaluation harness and approval gates an agent needs before it is safe to let it act.
Do I need an AI agent or just a chatbot?
Start by asking what the work requires the system to change. If customers ask the same forty questions, a chatbot on real documentation does it for a fraction of the cost and almost none of the risk. If invoices arrive as PDFs and get re-keyed, that is extraction, not an agent. You need an agent when the work is genuinely multi-step, the ordering varies, and the inputs are messy — and even then it needs an approval step on anything consequential.
Are fully autonomous AI agents safe for business use?
The limiting factor is accountability, not capability. Before autonomy, somebody has to answer who is responsible when it is wrong, and that is a governance question rather than a technical one. In practice the workable pattern is an agent that proposes and a human who approves anything with legal or financial consequence. Any vendor telling you the approval step is unnecessary has not operated one of these in production.
What should I ask a vendor selling an AI agent?
What exactly can it change and with whose permissions, as a specific list. Show me the audit log of a real run, not the diagram. What happens when it gets one in twenty wrong. Where is the human approval step. How will I know if a model update made it worse — the answer should be an evaluation set. What does it cost per month at ten times the volume. And could extraction or a plain integration do this instead?
Why are AI agents so much more expensive than chatbots?
Not the model, which is the cheap part. An agent that can act needs narrowly scoped permissions, an audit trail of every action, a rollback path, evaluation harnesses so you notice quiet degradation after a model update, human approval gates with a queue somebody works, and handling for the unhappy paths — what happens when step three fails after steps one and two already changed something. None of that scaffolding is needed when the system cannot change anything.
What is the most reliable AI automation for a small business?
Document and invoice extraction, almost every time. It is one model call plus a validation step, the input already exists in a folder or a mailbox, a human notices an error in seconds, and the hours it replaces are measurable. It is also invisible, which is why nobody proposes it — the website chatbot gets proposed instead, and it is rarely the better first project.