A client asked me last month why their shiny new chatbot had not cut a single hour off their accounts payable team. It answered questions about invoices beautifully. It just could not process one. That gap, between talking about the work and doing it, is the whole difference between a chatbot and an agent.
People mix the terms up constantly, and vendors are happy to let them. Calling a chatbot an "AI agent" sounds better on a pricing page. But if you are deciding what to build or buy, the distinction is the thing that decides whether the project saves you money or just adds another window for staff to ignore. This post sits inside our broader series on agentic AI, where I go deeper on what an agent is under the hood.
What people mean by each word
A chatbot is a conversation. You type, it replies. Modern ones are powered by large language models, so the replies are fluent and often genuinely useful. But the loop is closed: text in, text out. A chatbot can tell you how to reset your password. It cannot reset it for you.
An agent is built around a goal instead of a reply. You give it an objective and some tools, and it works out the steps, calls those tools, checks the result, and keeps going until the task is done or it hits something it cannot handle. The conversation is optional. Plenty of the agents we run at PTAS never talk to anyone; they sit on a queue of documents and quietly clear it.
The five differences that actually matter
Strip away the marketing and it comes down to five things. Each one is the gap that tripped up that AP team I mentioned.
1. A chatbot responds; an agent acts
This is the big one. A chatbot produces words. An agent produces outcomes. Ask a chatbot to onboard a new vendor and it will write you a lovely checklist. Ask an agent and it will read the vendor's documents, create the record, flag the missing tax form, and email the vendor to chase it.
2. Memory and state
Most chatbots forget. Each session starts fresh, and anything from yesterday is gone unless someone built a workaround. An agent holds state because it has to: it is partway through a task, it knows which invoices it already matched this morning, and it picks up where it left off. The continuity lives in the work, not the conversation.
3. Tools and access
A chatbot, by default, can only read and write text. An agent is wired into the systems where the work lives. It can query your ERP, write to a database, call an API, move a file. The model is still the brain, but the tools are the hands, and without hands you are only ever describing the job.
What tool use looks like in practice
Tool use is the part people underestimate. It is not one clever prompt. It is the model deciding, mid task, that it needs to look something up, choosing the right tool, reading what comes back, and adjusting. That decide-then-act loop is what makes an agent feel less like a search box and more like a junior colleague who gets things done.
A quick example from a document workflow
An invoice lands with no purchase order number. A chatbot would explain what a PO is. An agent opens the vendor master, finds the open PO by matching the amount and date, attaches it, and queues the invoice for approval. If it cannot find a confident match, it routes the invoice to a person with a short note on what it tried. This is exactly where our agentic document extraction earns its place, because reading the document is only step one.
4. Autonomy and decisions
A chatbot waits for the next prompt. An agent decides what the next step should be. That autonomy is the source of all the value and most of the risk. An agent that can act without asking is faster, and an agent that acts wrongly without asking is a liability. Getting the leash length right is most of the engineering.
5. What happens when it is unsure
Ask a chatbot something it does not know and it tends to guess, confidently. A well-built agent knows its own confidence and stops when it drops. It hands the case to a human with context instead of forcing a wrong answer through. Knowing when to ask for help is not a small feature. It is the line between something you can trust in production and something you switch off after a month.
A side-by-side, in plain numbers
When teams swap a question-answering chatbot for an agent on real document work, the numbers move because the agent removes touch time rather than just informing it. Here is the range we tend to see across finance and operations rollouts in the first 90 days.
Ranges PTAS observes in the field, not published industry benchmarks. Your numbers depend on volume and how messy your documents actually are.
Where each one fits
Neither is better in the abstract. A chatbot is the right call for some jobs and a waste of an agent for others. A rough way to place a task across four buckets.
Answer questions
Support FAQs, policy lookups, guiding a user to the right page. The output is information, and that is enough.
Draft and explain
Summarise a document, draft a reply, explain a clause. A person still reviews and acts on what comes back.
Finish a task
Read invoices, match them, post to the ERP, chase the gaps. The output is completed work, not advice about it.
Run a workflow
Onboarding, KYC, claims intake from end to end, with the agent escalating only the calls a human needs to make.
The mistake we see most
The pattern that wastes the most money has nothing to do with the model and everything to do with picking the wrong shape for the job.
Buying a chatbot when the job needs an agent. This is the AP team from the top of the post. They wanted invoices processed and bought something that could only talk about invoices. The demo looked great. Six weeks later the team was still keying everything by hand and answering the bot's questions instead of the other way round.
Building an agent when a chatbot would do. The opposite happens too. If all you need is to answer "where is my order," wiring an autonomous agent into five systems is expensive over-engineering. Match the tool to the task. Most failed projects I see are a mismatch, not a bad model.
So which do you need?
Ask one question: does the work end with information, or with an action? If a person reads the answer and then does the real task themselves, a chatbot is fine and probably cheaper. If you want the task itself done, reading the documents, updating the systems, handling the exceptions, then you need an agent, and a chatbot will only ever be a polite distraction.
For most finance and operations teams drowning in repetitive document work, the honest answer is an agent. The fastest way to know for sure is to put a real stack of your own documents through one and look at how much work actually leaves people's desks.
Common questions
What is the difference between an AI agent and a chatbot?
A chatbot holds a conversation and answers questions. An AI agent has a goal and can take action to reach it: read a document, call a system, update a record, and escalate the cases it cannot resolve. The chatbot talks; the agent does the work and reports back.
Is ChatGPT an agent or a chatbot?
On its own, a model like ChatGPT is a chatbot: you ask, it answers. It becomes an agent when you give it tools, memory, and the freedom to decide which steps to take, so it can act on the world instead of only describing it.
Do I need an AI agent, or is a chatbot enough?
If the job is answering questions or pointing someone to information, a chatbot is enough. If the job is finishing a task end to end, such as reading invoices, matching them, and routing exceptions, you need an agent. A chatbot can tell a customer their invoice status; an agent can process the invoice.