The question everyone asks about an AI assistant is "what if it says something wrong to a customer?" The AI Trust Center exists so you never have to guess at the answer. It shows you how often the assistant resolves conversations on its own, and — before any knowledge change reaches a visitor — exactly how that change affects the answers it gives.
What is my AI resolution rate?
Paid feature — Basic and Pro.
When a chat ends, the assistant scores how it went: resolved, escalated to a human, unresolved, or abandoned — and records whether the AI or an operator handled it. Your AI resolution rate is the percentage of conversations the assistant resolved entirely on its own.
You'll find it on the AI Trust page and in Analytics, with the trend over time and the topics that most often end unresolved — which is a to-do list for your knowledge base, not just a chart.
Scoring starts the day the feature is enabled, so give it a week or two of conversations before reading much into the number.
What is a simulation?
Pro feature.
A simulation takes your recent real conversations and asks: if these same visitors asked these same questions today, what would the assistant say now?
Pick a sample size (10, 25, or 50 recent conversations) and run it. The assistant re-answers every visitor question using your current knowledge base and chatbot rules, and an AI judge compares each new answer with what was actually sent at the time. Each answer gets a verdict:
- Unchanged — same substance. The normal, boring, good result.
- Improved — the new answer is more complete or accurate.
- Degraded — the new answer lost something or got something wrong. Look at these first.
- Gained — a question that used to get no answer now gets one. Your knowledge base got better.
- Lost — a question that used to be answered now isn't. Usually a removed or disabled document — or an answer that originally came from a live tool, which simulations deliberately never call.
- Changed — different, but not clearly better or worse. Often just phrasing.
Every verdict shows both answers side by side and a one-line reason, so you can overrule the judge with your own eyes. Simulations never send anything to visitors and never change anything — they only report.
When should I run one?
After any knowledge change you care about: editing or removing documents, rewriting chatbot rules, disabling a source, a big questionnaire session. Run a simulation, scan the degraded and lost columns, fix what they point at, run again. It costs a few minutes and turns "I hope that edit was fine" into "I checked."
For single questions, the test console is still the fastest tool — simulations are for checking everything at once.
What are golden questions?
Pro feature.
Golden questions are your assistant's test suite: the questions it must always get right. Each one is a question plus the facts the answer has to contain — your opening hours, your refund window, the thing your industry regulator cares about.
They run with every simulation (or on their own with Run tests now), and each gets a pass or fail: paraphrasing is fine, but a missing or contradicted fact fails. You can add up to 50, and seed them straight from answer gaps — real questions your visitors already asked.
Write one for every fact you'd be embarrassed to see answered wrongly. That's the whole art of it.
Can it watch for problems automatically?
Pro feature.
Yes. Turn on Nightly checks in the AI Trust settings and every night a small simulation runs on its own — a sample of your recent conversations plus your golden questions. If answers degraded past your threshold or a golden test fails, you get an email naming the exact questions to look at. A quiet night sends nothing.
You can also email yourself the full trust report on demand — resolution rate, latest run results, test suite status — from the same settings card.
Does this cost extra or affect my visitors?
No to both. Simulations run in the background against a copy of the conversation, visitors never see them, and live tools are never triggered — so no bookings, lookups, or webhooks fire during a run. Simulation AI usage is metered like the rest of your plan and shown on each run, and runs are capped in size so they stay small.
Where do I find it?
AI Trust in the dashboard sidebar, under the insights section. The resolution rate card also appears in Analytics. Both are visible to workspace admins.