AI assistants are now good at sounding right. That is the problem. A general chatbot writes fluent, confident text whether or not it knows the answer, and in regulated work a confident wrong answer is worse than no answer at all.
Why fluent is not the same as correct
Large language models generate the most plausible next words. Ask one about your company's leave policy, a clause in a supplier contract or the deadline on an audit finding, and it may produce an answer that reads perfectly and is simply invented: a policy number that does not exist, a clause that was never signed, a date that was never set. This is often called hallucination, and no amount of polish on the answer makes it safe.
Grounding, in plain terms
A grounded assistant does not answer from memory. It answers from your approved sources, and it shows you which ones. The flow looks like this:
- Ask. A person asks a question in their own words.
- Retrieve. The assistant searches your documents and records for the passages that actually address it.
- Answer from the sources. It writes the answer using only what it found.
- Cite. It shows the document, section or record each part of the answer came from, so anyone can check it.
- Escalate. When the sources do not cover the question, it says so and hands over to a person.
Knowing when not to answer
The most important behaviour of a trustworthy assistant is the one demos rarely show: declining. "I can't find that in our policies; I've passed your question to the HR team" is a good answer. It protects the person asking, the organization and the assistant's credibility.
Documents are not the only source
Many questions are about live records, not documents: which findings are overdue, what a customer owes, which contracts expire this quarter. A good assistant queries the system for these, with the same permissions as the person asking, rather than guessing. And when it takes an action on someone's behalf, that action is logged.
Six questions to ask any AI vendor
- Does every answer show its source, and can I open it?
- What happens when the answer is not in our documents?
- Does the assistant respect each user's permissions?
- Where do our documents, questions and logs live?
- Can we audit what it said, to whom, and why?
- Does it work as well in Arabic as in English?
How we build it. Every Moving Point platform includes KEIO, a copilot that answers from your own records in Arabic or English, by text or voice, and logs every action it takes. For customer-facing questions, Oris answers across chat, WhatsApp and voice, cites its source and hands over to your team when it should.
