Agent vs chatbot
Outgrown chatbots?
Meet an agent that knows your business.
Not a script. Not a decision tree. An agent grounded in your own content, with your team one step away.
A rule-based chatbot answers the questions someone predicted, in the words they predicted. Polycarp reads your website, answers the question a visitor actually asked, shows where the answer came from — and when it doesn't know, it says so and offers a person. Try both below.
Same question, two answers.
Pick a business, then ask both the same question. The left side is a rule-based chatbot as they are usually built — simulated here, connected to nothing. The right side is a live Polycarp agent that has read that business's website; the sample businesses are fictional, each with its own voice.
A decent rule-based bot: intents with synonyms, correct canned answers, menus, "did you mean", a fallback to a person. Simulated here; connected to nothing.
The real agent, answering from polycarp.yohanun.com. Every answer shows its sources; when it has none, it offers a person.
What an agent does that a chatbot can't.
The honest version. Where a chatbot's approach is fine, we say so.
| Polycarp | Rule-based chatbot | |
|---|---|---|
| How it answers | Understands the question as the visitor put it, answers from your website's content, and shows the pages the answer came from. | Matches keywords to a menu of scripted replies. Anything phrased differently is "I didn't understand". |
| When it doesn't know | Says so, and offers your team — every time. The offer is guaranteed by the platform's own grounding verdict, not left to the model's habits. | A dead end, or the nearest canned reply whether or not it fits. |
| Human handoff | The visitor's words and the whole conversation go to your team by email. A person can take the conversation over; the agent goes quiet and returns when they are done. | Often missing. When present, the visitor usually starts over with someone new. |
| Setup | One form: your website and your email. Polycarp reads the site, proposes the agent's look from your brand, and gives you one line to embed. | Someone designs every flow and writes every reply, then maintains them. |
| Documents | Upload a PDF or Word file — a price list, a policy, a leaflet — and the agent answers from it too, citing the document and the page. Nothing has to be on the website. | Only what someone typed into a flow. A document is a link at best. |
| Keeping up | When a page changes, refresh it (or re-read the whole site) and the agent answers from the new version. Questions it couldn't answer come back grouped by topic, with your team's replies drafted into the knowledge base for you to approve. | Every change is a manual edit to a flow. |
| Taking action | Connect your shop or your own system and it can look up an order and refund it. Your rules are checked in code, not by the model; above the limit you set, a person approves. Only ever for the signed-in customer's own account. | Can link to a form. Anything more is custom development against each system. |
| Showing its work | Open any answer in your inbox to see the passages it used, the page each came from, and whether each still stands. | You can see which flow fired. |
| Guardrails | Rate limits per visitor and per address, a daily conversation ceiling set by your plan, a relevance gate on off-topic traffic, and instruction-override attempts tested on every agent. | A script can't be talked into anything — and can't answer anything it wasn't given. |
| Your data | Each agent lives in its own walled tenant. Export everything as one zip, or delete the lot, from your own account. Hosted in the EU. | Varies by vendor. |
| Where it works | Your website, where the question is asked. Not email, phone or social — we don't claim channels we don't have. Automations tell Slack, email or your own system when something needs you. | Usually web chat as well. |
| Testing | Every agent starts paused, with a test bench that is the real chat window talking only to you. Polycarp itself is checked on every change against ratified question sets run through the real chat surface across seven failure classes — grounded, refusal, isolation, follow-up, injection, register, handoff. | Clicking through the flows by hand. |
What we don't claim.
Polycarp is new. We don't publish a resolution rate yet, because we don't have one worth your trust: the number depends entirely on how "resolved" is defined, and we would rather show you the definition with the figure. What we can show today is the agent itself, live, above — and the question sets every change to Polycarp must pass before it ships. Read more in the FAQ.
Put your own website behind it.
Your agent reads your site while you watch. On Starter your first month is free, and we'll set it up with you if you'd rather.
Start your free month Talk to us