Most AI demos are built to impress. They answer everything, instantly and fluently. That’s exactly the problem when the answers matter.
When BRIO built an AI knowledge expert for our Litify practice, the most important behavior we designed wasn’t speed or breadth. It was the agent’s willingness to say “the sources don’t cover this.” In our experience, that one sentence is the difference between an AI tool your team trusts and one they quietly stop using.
A wrong answer doesn’t stay a wrong answer
Our consultants, analysts, and developers ask detailed Litify questions every day: how a document feature behaves, which fields a standard object ships with, what changed in a release. If an AI assistant answers one of those questions wrong, the mistake doesn’t stay in the chat window. It becomes a requirement, a design decision, or a promise to a client.
General-purpose AI assistants are fluent, but they aren’t reliable on this kind of detail. They’ll describe a setting that doesn’t exist or a behavior from the wrong release with complete confidence. And confidence is the problem: a plausible wrong answer is much harder to catch than an obvious one.
Three rules that make AI trustworthy
We built our Litify knowledge expert using retrieval-augmented generation (RAG), where the AI answers from a curated set of trusted sources instead of its general training. But RAG alone isn’t enough. What made it trustworthy were three rules we set before anyone relied on an answer.
1. Answer only from trusted sources
The agent reads three sources and nothing else: Litify’s product documentation, a complete map of how the Litify base package is actually built, and our own team’s documented fixes. It never falls back on general knowledge, and it doesn’t browse the open web, which avoids misinformation and removes a whole category of security risk.
2. Cite everything, and label what kind of source it is
Every answer cites its source, and says whether that source is official documentation, as-built metadata, or a BRIO workaround. The agent also separates what an article states from what it infers. An inference presented as a citation is the hardest error for a reader to catch, so we don’t allow it.
3. Treat "not found" as a feature
When the sources don’t answer a question, the agent says so. A known gap becomes something to test in the org. A confident guess becomes a commitment nobody can keep. We’d rather our team hear “I don’t know” a hundred times than act on one invented answer.
Governance comes first, not last
Most organizations add AI governance after something goes wrong. We did it the other way around. Before the agent answered its first real question, we defined what it may read, what it must cite, how it labels confidence, and what it must refuse to guess. We also verified every stage of the pipeline before building the next one, from article counts to the way content was split for search.
That order matters. Governance designed in from the start shapes how the whole system works. Governance added later is usually a filter bolted onto something that was never built to be accountable.
What this means for your firm
If you’re evaluating AI tools for legal work, or thinking about building your own, ask the same questions we asked:
- Where do the answers come from? A tool that can’t tell you its sources can’t be verified.
- Does it cite, every time? Citations turn an AI answer into something a person can check in seconds.
- What happens when it doesn’t know? Test this deliberately. Ask questions the tool shouldn’t be able to answer and see what it does.
- Who decided the rules? Look for governance decisions that were made on purpose, not defaults nobody reviewed.
The goal isn’t AI that knows everything. It’s AI that knows what it knows, shows its work, and is honest about the rest. That’s the kind of AI your team can build real work on.
Want to talk about putting trustworthy AI to work at your firm? Learn more about our AI consulting or read what the ABA’s guidance on generative AI means for your firm.
Frequently asked questions
What is retrieval-augmented generation (RAG)?
Retrieval-augmented generation is an approach where an AI system first retrieves relevant passages from a curated set of trusted sources, then writes its answer from those passages instead of relying only on what it learned in training.
Why does it matter if an AI tool says "I don’t know"?
A tool that admits gaps lets people verify or test what’s missing. A tool that always answers can present wrong information with confidence, and those errors are much harder to catch.
How can a law firm test whether an AI tool is trustworthy?
Ask where its answers come from, check whether it cites sources every time, and deliberately ask questions it shouldn’t be able to answer to see whether it admits the gap or makes something up.
