Three ways to do this
A scripted chatbot, an in-house build on an LLM API, or a trained agent. They fail in different places, and the difference shows up in maintenance rather than in the demo.
| Scripted chatbot | Build it yourself | OU Chat | |
|---|---|---|---|
| Time to a working answer | Days. Someone has to write every branch of every flow. | Weeks. Retrieval, chunking, embeddings, evals, and a UI. | Minutes. Paste a URL and it trains on what is already there. |
| What happens off-script | Dead end. "I did not understand that" and a contact form. | Depends entirely on the retrieval you built. | Answers from your content, or says it does not know and escalates. |
| Keeping it current | Manual. Every price change is a flow edit. | You own the re-index pipeline and its failures. | Re-crawls on a schedule. Catalog syncs from Shopify or WooCommerce. |
| Can you trust the answer | Yes, because it only says what you typed. | Only as far as your own evals go. | Every answer cites the page it came from, so it is checkable. |
| Handover to a person | Usually a form. The context is lost. | Another system to build and connect. | Built in. The operator gets the full transcript and the visitor identity. |
| Channels | Usually the website only. | One integration per channel, each one yours to maintain. | Website, WhatsApp, Messenger, Instagram, Telegram, LinkedIn, voice, SMS, email. |
| What it costs | Low licence, high staff time. | Engineering time, plus model spend you have to govern yourself. | From $29/mo, $5/mo for verified non-profits. |
When the other two are the right answer
- A scripted bot, when the path must never vary
Regulated disclosures, a fixed intake sequence, anything where an unexpected phrasing is a compliance problem. A script does exactly what you typed, and sometimes that is the requirement.
- Build it yourself, when the agent is the product
If the retrieval behaviour is your differentiator, own it. Buying it means inheriting someone else's opinions about chunking and ranking.
- Us, when it is a job rather than a project
You need visitors answered accurately, escalation that carries context, and leads captured, and you would rather not run an ML pipeline to get it.
Paste your URL and ask it the questions your customers actually ask. Seven days without a card, and a full refund on any plan if it is not right.
Build your agent →