How to tell if an AI vendor is verifiable — cite sources, show confidence, admit "I don't know"

Every AI vendor sounds confident in the demo. That is exactly the problem. Confidence is cheap; being verifiable is not. When you are buying AI for a real business, the question that matters is not “how smart is it?” but “can I check its work?”

At BIXSO we hold to one principle above cleverness — Truth over Performance: cite sources, show confidence, and say “I don’t know” rather than over-claim. It is also the sharpest lens a buyer can use on any vendor, including us. Here is that lens as a checklist.

The three tests

1. Does it cite its sources? A verifiable system points to where an answer came from — a document, a record, a rule — not a fluent paragraph from nowhere. Ask a vendor: when your AI answers, can the user click through to the source? If the answer is “it just knows”, that is not knowledge, it is a guess dressed as an answer.

2. Does it show confidence — and act on it? Real systems know when they are unsure. Look for a confidence signal the system actually uses: high confidence acts, low confidence escalates to a human or asks a clarifying question. A vendor whose AI is equally certain about everything is hiding the uncertainty, not removing it.

3. Will it say “I don’t know”? The hardest and most important behaviour. A system that never refuses will confidently invent an answer when it is out of its depth — the failure mode that quietly destroys trust in regulated work. Test it with a question just outside its domain. If it bluffs, walk.

Beyond the answer: the governance tests

• Guardrails are enforced in code, not vibes. Ask where the “never do X” rules live. In a serious system they are gates in the pipeline, not a hopeful line in a prompt.
• Every call is logged and reviewable. You should be able to audit what the AI did and why. No log, no accountability.
• A human is in the loop where it counts. In high-stakes decisions, the system connects to a person rather than deciding alone. In health and legal work, that handoff is not a fallback — it is the law.
• Cost and usage are metered. A vendor who can’t tell you the per-call cost can’t run the system responsibly at scale.

Why this is our standard, not just our pitch

We build every BIXSO product on grounded cognition: the AI is grounded in a real map of the business, connects to a human when it should, and improves from reviewed feedback. Guardrails are code gates. AI calls run through one governed gateway that meters every call. We publish our own numbers — including the bad ones — because a vendor that only shows you wins is performing, not proving.

Use this checklist on every vendor you evaluate. Then use it on us. Book a consult — the first one is free — and bring the hard questions.