Guardrails, by design
The most important thingour AI does is say"I don't know."
Zee is the voice assistant that answers questions about property contracts and reports. Most of the engineering behind it wasn't spent teaching it to answer. It was spent teaching it where to stop.
Today, Zee responds to what you tell or ask it. Direct document ingestion is on the roadmap — and the same guardrails will apply whether the source is the file itself or the conversation around it.
The problem with fluency
A confident guess reads exactly like knowledge.
General-purpose AI is trained to be helpful, and being helpful is measured by producing an answer. That incentive is fine when the cost of being wrong is a bad recipe. It is not fine when the subject is a contract of sale, a title, and a settlement date.
A property transaction is not a general knowledge question. It is a specific matter, in a specific jurisdiction, under a specific edition of a specific form, with annexures that change what the front page appears to say. The plausible answer and the correct answer diverge constantly — and only one of them is supported by the facts at hand.
So we built Zee around a different objective. Not "always answer." Be right, or be explicit that you aren't sure — whether the source is the document itself or what someone has shared in the call.
Optimised to always produce an answer. Confidence is the product, so uncertainty gets smoothed over — and the reader can't tell where knowledge ended and pattern-matching began.
Optimised to be right or silent. Uncertainty is surfaced as a finding, not hidden as fluency. The boundary of what the conversation supports is visible on the page.
Four hard rules
What Zee is not
allowed to do.
These aren't tone-of-voice preferences. They are constraints in the system itself.
It never invents a clause
Zee only speaks to what has been shared in the conversation — whether that is a clause read out, a term described, or a question about the contract or report. If something isn't there, the answer is that it isn't there, not a plausible-sounding version of what usually appears.
It cites, or it stays quiet
Every substantive answer traces back to the specific information the caller has provided. Where there is nothing to point at, Zee says so instead of filling the gap with fluent prose.
It refuses to give advice
Zee explains what a clause does and what to watch for. It does not tell anyone whether to sign, what to negotiate, or what the commercial call should be. That is the conveyancer's judgement, and Zee names it as such.
It hands back when it should
Ambiguity, missing information, unusual facts, anything turning on circumstances outside the conversation — Zee flags the difference between a quick explanation and something that genuinely needs a professional's own read.
Why the restraint is the feature
In property, the cost of a wrong answer is someone's house.
A hallucinated easement sends a buyer into a renegotiation over something that was never on the title.
A missed special condition is exactly the pattern that shows up in reported negligence decisions against conveyancers.
A confident wrong answer is worse than no answer — because nobody goes back to check it.
The professionals we build for don't want an assistant with opinions. They want one that listens carefully, shows its working, and knows the exact point where their judgement takes over. That handoff is the whole design.
Trust in this category isn't earned by how much an AI can say. It's earned by what it declines to say.
Zetld is a technology company, not a law firm or conveyancer. Zee explains documents based on the information provided in conversation; it does not provide legal advice, and using it does not create a solicitor-client relationship.
Built to be right,
not to be impressive.
See how the same discipline shows up in a full read.