AI Agent vs Chatbot: What Actually Separates Them

What each term actually means

Anthropic's engineering write-up Building effective agents draws the cleanest published line. Agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks". Workflows are "systems where LLMs and tools are orchestrated through predefined code paths". The distinction is not intelligence. It is who chooses the sequence.

Chatbot

A chatbot receives a message, produces a reply, and stops, with the sequence of what happens next determined by whatever the person types.

A modern AI chatbot is usually retrieval-augmented. It searches your content for passages relevant to the question, passes them to a language model with the question, and returns a generated answer grounded in what it retrieved. That is a genuinely capable system. It can handle phrasings nobody scripted, combine facts from two different pages, and answer in dozens of languages. It is still a chatbot, because the unit of work is one turn of conversation.

Chatbots do call software. Retrieval is a tool call, and so is writing a captured email address into a CRM or opening a support ticket on escalation. What makes them chatbots rather than agents is that those calls sit on a path somebody defined in advance. The model fills in the words, not the plan.

What you get

• Answers to questions your documentation already covers, available immediately and at any hour.
• Predictable cost, because one conversation is roughly one bounded set of model calls.
• Predictable blast radius. The worst case is a wrong sentence, not a wrong transaction.
• A short review loop. You read transcripts, find gaps, write the missing content.

Examples: matram.ai, Tidio Lyro, Most website support widgets, Most in-app help bots

AI agent

An AI agent is given a goal rather than a question, then chooses its own sequence of tool calls, observes the results, and keeps going until the goal is met or it gives up.

The loop is the definition. An agent takes an objective, decides what to do first, calls a tool, reads what came back, and uses that to decide the next step. Nobody wrote the branch it took, because the branch is chosen at runtime. That is precisely the property Anthropic's definition points at: the model maintains control over how the task gets accomplished.

In customer support that means the difference between explaining the refund policy and issuing the refund: checking the order, confirming eligibility, calling the payments API, writing the outcome back to the ticket, and telling the customer what it did. Each of those steps can fail, and the agent has to notice and adapt rather than following a script off a cliff.

What it costs you

• Real integration work. An agent with no write access to your systems is a chatbot with a longer sales deck.
• Permission design. You are deciding which actions a probabilistic system may take without a human in the loop.
• Harder evaluation. You cannot enumerate the paths in advance, so you test outcomes and audit traces instead of reviewing scripts.
• Variable cost. Multi-step reasoning burns far more tokens per conversation than a single grounded answer.

Examples: Salesforce Agentforce, Fin, Sierra, Decagon, Zendesk AI agents

Chatbot vs AI agent across the five axes that matter

Autonomy, tool use, memory, decision-making and oversight. Marketing copy tends to claim all five. A demo will only support some of them, and the third row is the one that settles most arguments.

Rows three and seven are the load-bearing ones. If a product cannot change anything outside the conversation, the rest of the agent vocabulary is decoration. If it can act but cannot notice a failed step and pick a different route, it is a workflow with a language model at the front, which is a legitimate and often preferable architecture, just not an agent.

When an AI agent is not the answer

Every page on this keyword has an incentive to tell you that you need the more expensive category. Here is the ramp running the other way, starting from the case where you need neither product.

Doing it by hand is genuinely fine

If the transactional work arriving each week fits in the gaps between other tasks, automating it is a project looking for a problem. Someone opens the order, changes the address, replies. It takes moments and it is right nearly every time, because a person who can see the whole account is a very good exception handler. Agent platforms are priced against volume you do not have yet, and the integration work costs considerably more than the clicking does.

Where the friction starts

Friction shows up as repetition long before it shows up as volume. The same small set of requests, several times a day, each one interrupting something else. This is the stage where a chatbot starts paying for itself, because most of that traffic turns out to be questions with documented answers rather than actions. It is not yet the stage where an agent pays for itself. The expensive part of an agent is the integration and the permission design, and neither of those gets cheaper because your inbox got busier.

Where it turns into a liability

The liability stage looks like backlog. Requests sit overnight, the customer chases, and whoever picks it up in the morning has no idea what was promised earlier in the thread. Now the manual route is costing you trust rather than hours, and the argument for automating the transaction itself becomes real. But notice what has to be true first. You need APIs for those transactions and somebody willing to own what a probabilistic system is allowed to do unsupervised. Without both, an agent moves the backlog somewhere less visible instead of clearing it.

The case that breaks it

What actually breaks the manual route is not the busy week. It is the request that has to touch several systems in a particular order, where finishing most of the sequence and missing the last step leaves an account in a state nobody can read afterwards. People skip steps when they are tired, and they rarely log the one they skipped. A system that retries, then records what it did and what it could not do, is genuinely better at that, and it is the one situation where the category itself is the thing you are buying rather than the tooling around it.

None of that is an argument against automation. It is an argument for buying the smallest thing that clears the problem actually in front of you, then reassessing when the shape of the work changes rather than when the vocabulary does.

Where the marketing gets it wrong, and the test that cuts through it

Gartner coined the term "agent washing" in a June 2025 press release for the practice of rebranding assistants, RPA and chatbots as agentic AI without the underlying capability. Its analysts estimated that of the thousands of vendors claiming agentic AI, roughly 130 were genuine.

That ratio matches what you find reading product pages. The word "agent" now attaches to any system that answers a question, and there is no penalty for using it, so everyone does. The useful response is not cynicism about the category, which is real, but a test you can run yourself.

The test: can it take an action beyond responding, and did it choose to?

Two parts, and both have to hold. Ask the vendor for a live demo, not a video, of a task that ends in something changing in a system that is not the chat window. Then ask who decided the order of the steps.

• "Show me it doing something, not saying something." A refund issued, a subscription downgraded, a delivery address changed, a ticket reassigned. If everything it produces is text, you are looking at a chatbot.
• "Who wrote the sequence?" If the answer is a flow builder, a canvas or a set of configured playbooks, that is Anthropic's "predefined code paths", which is a workflow. Workflows are fine. They are just not what the pricing usually implies.
• "Break it mid-task." Have the payments call fail, or the order lookup return nothing. An agent notices, tries something else or hands over cleanly with state intact. A dressed-up workflow either stalls or reports success.
• "What did it do, and can I see the trace?" Genuine agent products expose a step-by-step log of tool calls, because their own support teams cannot debug without one. Absence of that log is informative.
• "Which actions require a human approval?" A vendor that has actually shipped autonomy has thought hard about permission scopes. A vendor that has not will treat the question as a feature request.

Where matram.ai sits, plainly

matram.ai is an AI chatbot. It crawls your site, imports documents and connected sources, and generates answers cited to the page they came from, with a strict mode that restricts it to your approved content. It captures leads and escalates to a person through a shared inbox. Those are fixed paths that we wrote, not plans it invented. It does not process refunds, it does not decide its own multi-step sequences, and calling it an agent would be exactly the relabelling this section is about.

That is not a hedge about the roadmap. It is the honest category, and for most of the questions arriving in a support inbox the chatbot category is the one that actually fits the work. If your problem is genuinely transactional, the honest recommendation is a product built for it, and several of them are named on this page.

The pricing model tells you which one you are buying

Vocabulary is cheap, but pricing pages commit. Products that genuinely complete tasks tend to bill for completed tasks, and products sold on autonomy tend not to publish rates at all. All figures below are list prices published by the named vendor or its AWS Marketplace listing, read on 20 July 2026.

Two things stand out. First, outcome-based pricing is a real signal: a vendor only bills per resolution if it believes the system finishes the job, and Salesforce's June 2026 investor announcement of its agreement to acquire Fin for around $3.6bn stated that Fin resolves on average 76% of support volume end to end across 30,000+ customers. Second, cognigy.com/pricing and kore.ai/pricing both return 404, and ada.cx/pricing shows a consultation link rather than rates, which is why the only public numbers for those three come from a marketplace listing. If you are modelling any of this against your own volumes, the chatbot ROI calculator takes your inputs rather than a vendor's. For the helpdesk side of the same question, Intercom pricing and Zendesk chatbot pricing break down what the seat and resolution charges add up to.

What getting this decision wrong actually costs

The invoice is the smallest part of a wrong answer here, in either direction. The costs that hurt are the ones that never reach a finance system.

Buy agent-shaped software for work that was never transactional and the subscription is the visible loss. The quiet one is the quarter your team spends on an integration project, and the realisation somewhere near the end of it that what your customers needed was a better help centre. That time does not come back. Neither does the standing of whoever signed the purchase order, which matters more than people admit, because the next proposal to automate anything now starts from a deficit.

Undershooting costs you in a different currency, and it lands on customers rather than on a budget line. A system that keeps explaining the refund policy to somebody who has asked repeatedly for the refund itself teaches that person your chat window is where requests go to be restated. They stop using it. The volume moves to email, or to a review site, or nowhere at all, and none of those show up as a drop in a dashboard you are watching. The conversations that never happen are not logged anywhere.

Then there is the migration you now get to do twice. Content indexed, channels wired up, escalation rules argued over, a team finally trained on a tool you are about to replace. Habits form around whatever is in front of people, and they outlive the software: agents who learned to route around a system they did not trust will route around its replacement too. So ask what you would have to unpick a year from now if today's choice turns out to be the wrong one. Which of these two categories is cheaper to be wrong about?

Which one you actually need

The question is not which category is more advanced. It is whether the work you are automating ends in an answer or ends in a change to a record.

For most small and mid-sized businesses the honest answer is a chatbot now and a reassessment later, because the integration and governance work an agent needs is usually larger than the ticket volume justifies. That is the reasoning behind our small business chatbot guidance, and it is the same calculus for a SaaS support deployment where the questions are mostly about how the product works rather than about changing a record. If you are already inside a helpdesk and weighing its built-in agent against a standalone bot, the Zendesk alternatives comparison and the Intercom alternatives comparison cover that trade honestly, including the cases where staying put is correct.

The categories are converging, and the label will keep moving

Both sides are moving toward the middle, which is why the vocabulary will not settle any time soon.

Chatbots keep gaining tool use. Function calling is standard across every major model, so booking a slot or looking up an order status from within a chat is now ordinary rather than exotic. Agents, meanwhile, keep gaining retrieval and guardrails, because unconstrained autonomy turned out to be a liability and the fix is grounding the agent in approved content and gating its riskier actions. Push both trends far enough and they meet.

Vendors are redrawing the line as they go. Zendesk describes its AI agents as "autonomous systems designed to understand and autonomously resolve complex issues on any channel" that, "unlike traditional chatbots", "can reason across multi-step requests". Whether a given deployment clears that bar depends entirely on what it has been wired into, which is the general problem: the category name now describes ambition rather than architecture.

The market is consolidating around that ambition too. Salesforce reported Agentforce at "$1.2 billion in ARR in Q1 FY27, up 205% year-over-year" alongside its agreement to buy Fin, and the same investor materials are worth reading as evidence that support automation is now sold as outcomes rather than as software. Note the naming carefully, because it trips up a lot of coverage: the company is Fin, and the helpdesk product is still sold as Intercom with Fin as its AI agent. In the other direction, Salesloft's own site now states "We've transitioned from Drift to 1mind", so Drift is not a product you can buy today, whatever a two-year-old comparison post tells you.

What to do about a label that keeps moving

• Buy against a task list, not a category. Write down the ten things you want automated before you look at a single product page.
• Ask what it is permitted to change, and what happens when it changes the wrong thing. The answer separates the categories faster than any feature grid.
• Treat "agentic" in a headline as null information, and treat a published per-outcome price as a real commitment.
• Re-run the question in a year. The line between these categories has moved twice since 2023 and will move again.

None of that requires picking a side in the terminology argument. It requires knowing which side of it you are standing on, which is also the reason this page says out loud that matram.ai is a chatbot. For the neighbouring confusion between chatbots and platform-level conversational AI, the same test applies with a different word attached.

Frequently asked questions

Sources

• Anthropic, "Building effective agents" (19 December 2024) - The agent versus workflow definition: agents dynamically direct their own processes and tool usage, workflows run through predefined code paths.
• Gartner newsroom, press release 25 June 2025 - The term "agent washing" and the estimate that roughly 130 of thousands of vendors claiming agentic AI were genuine.
• Salesforce investor relations (15 June 2026) - The definitive agreement to acquire Fin, formerly known as Intercom, for approximately $3.6bn; Agentforce at $1.2bn ARR in Q1 FY27, up 205% year-over-year; Fin resolving on average 76% of support volume end to end across 30,000+ customers.
• Fin pricing page (accessed 20 July 2026) - $0.99 per outcome, one outcome billed per conversation, 50-outcome monthly minimum, $9.99 per qualification.
• AWS Marketplace vendor listings (accessed 20 July 2026) - Ada Strategic at $33,000 per 12 months including 60,000 conversations; NiCE Cognigy Basic 5K pm at $43,080 per 12 months including 60,000 conversations annually; NiCE Cognigy private SaaS enterprise licence at $1,000,000 per 12 months for up to 10 million conversations a year; Kore.ai per-session rates of $0.015 to $0.20 by module.
• Vendor pricing pages checked directly (accessed 20 July 2026) - cognigy.com/pricing and kore.ai/pricing returning 404, and ada.cx/pricing offering a consultation link rather than published rates, which is why the marketplace listing is the only public source for those three.
• Zendesk AI agents product page - Zendesk's own definition of AI agents as autonomous systems that, unlike traditional chatbots, can reason across multi-step requests.
• Salesloft - The statement that Salesloft has transitioned from Drift to 1mind, so Drift is not currently purchasable.