We've walked the traps you're about to hit. Here's what to test.
Everyone can build a chatbot. The hard part is an assistant that only says what the business gave it, refuses what it isn't allowed to do, and proves every line.
It isn't. The model is a swappable part; the conversation's quality comes mostly from the engine and the business's content.
A good-enough chatbot, yes. Refusals before effects, exact practice wording, consent gates and a sentence-level audit take far longer.
Every fix becomes another branch, until the model barely gets a say. Code should decide what's true; the model should only mediate.
Months before anyone sees value. Molo starts with no integration and connects one journey at a time.
The audit, sentence by sentence.
Every sentence Molo sends carries its source. Anything the model says on its own is marked, and the client is told.
Your schedule shows home contents cover at your Bellville address.
From your policy scheduleItems you carry every day fall under Portable Possessions.
From your policy wordingThandi will call you tomorrow morning.
Handed to your adviser, with the whole conversationMost insurers treat this kind of loss the same way.
From this model's training [ ! ]Four rules, enforced in code.
- The model is the only voiceCode never writes a sentence to a client.
- Code refuses before it actsWhat's true and what's allowed are checked before anything happens.
- Your words are dataDisclosures, questions and refusals are used exactly as written.
- Every sentence is checkedAgainst what the business gave it, and the check is never rewritten.
Use your own AI provider and keys; Molo charges nothing for AI.
Molo-hosted, client-hosted or hybrid, confirmed with the project.
A POC runs apart from other clients, and all your data is removed at the end.
During the POC, try to make Molo say something your business never gave it.
Bring your own data and your hardest questions. You'll see the audit under every reply, and the [ ! ] wherever the model is on its own.