Tidio’s Lyro resolves 67% of conversations on average — what that number is actually measuring
Tidio puts Lyro’s average resolution rate at 67%, which is the figure that separates an AI agent from a live chat widget. But a resolution rate is mostly a property of your support content, not of the model — and that determines whether you land above or well below the average.
Table of contents
Every AI support tool now leads with a resolution rate, and most of them are impossible to compare. Tidio's is at least specific: it puts Lyro, its AI agent, at an average resolution rate of 67% — roughly two in three conversations handled without a human stepping in.
That number is the entire argument for an AI agent over a plain live chat widget. It is also the number most likely to be misread, because a resolution rate is far less a property of the model than of what you feed it.
What Lyro is, and how it differs from a chat flow
Tidio ships two distinct things, and the difference matters more than the branding suggests.
Flows are rule-based. You draw the paths: if the visitor clicks this, ask that; if the answer is X, route to a human. Predictable, auditable, and completely blind to any question you did not anticipate.
Lyro is the AI agent. Rather than following a decision tree, it interprets the question and composes an answer from your own support content — help centre articles, FAQs, whatever you point it at. Tidio has also extended it with Lyro Actions, so it can carry out tasks such as checking an order or triggering a workflow rather than only replying with text.
The distinction we drew in AI agents vs chatbots applies exactly: a flow answers the questions you predicted, an agent attempts the ones you did not.
What actually counts as a "resolution"
Here is the part worth slowing down on, because it is where vendor figures across the whole category get slippery.
A resolution is generally counted when a conversation ends without being escalated to a human. That sounds reasonable until you notice what else produces that outcome:
- The customer got a correct, complete answer. (A genuine resolution.)
- The customer got a plausible answer and left, then emailed you a day later. (Counted, deferred.)
- The customer gave up. (Counted, and worse than no bot.)
No vendor's headline metric distinguishes these three cleanly, and that is not unique to Tidio — it is how the metric is defined industry-wide. So treat 67% as a ceiling on how much work is being deflected, not a floor on how well it is being done, and pair it with something that measures the other half: post-chat satisfaction, or the rate at which "resolved" conversations reappear as tickets within 48 hours.
If you want a harder number to plan against, our guide to calculating cost per chatbot resolution sets out how to turn a deflection percentage into an actual figure per conversation.
Your resolution rate is decided before you install anything
The single most useful thing to understand: an AI agent cannot resolve what your documentation does not contain.
Lyro answers from your support content. If your help centre covers twelve articles and your customers ask about forty topics, no amount of model quality closes that gap — the agent will either escalate or improvise, and improvising is the failure mode you least want. Companies that land above the average almost always have a well-maintained help centre already; companies that land far below it usually discover their documentation was the problem all along.
Which gives you a genuinely useful pre-purchase exercise, and it costs nothing:
- Pull your last 100 support conversations.
- Tag each with the topic.
- For each topic, check whether a public article answers it completely.
The share where the answer is yes is roughly your realistic ceiling. If that comes out at 30%, fix the documentation first — you will get more from a fortnight of writing than from any tool. Our guide on connecting a chatbot to your knowledge base covers how the content needs to be structured to be usable by an agent, which is not the same as being readable by a human.
Testing it without committing
Tidio gives every account 50 Lyro conversations free, which is enough for a real test rather than a demo. Beyond that, Lyro is metered in conversation packages separate from the seat pricing, so check the current tiers on Tidio's pricing page before budgeting — the units are conversations, not messages, and the distinction changes the cost meaningfully.
Use the free allowance properly:
- Point it at your real help centre, not a curated subset. The test is worthless on content you tidied up for the occasion.
- Feed it your ten most awkward real questions, including the ones with conditional answers ("depends on your plan", "depends on your country"). This is where agents break.
- Check the escalation path, not just the answers. What a well-configured agent does when it does not know is more important than its hit rate — a clean handover beats a confident guess every time.
- Read the failures. The escalated conversations are a documentation to-do list.
That third point is where most deployments go wrong; several of the patterns in 10 chatbot mistakes that make customers leave are really just badly handled handovers.
Who it fits
Lyro makes most sense for small and mid-sized teams with high-volume, repetitive questions and documentation that already exists — e-commerce order status, shipping and returns, plan and billing questions. That is the shape of workload where a two-thirds deflection rate translates into real hours saved.
It fits poorly where every enquiry is bespoke, where answers depend on account state the agent cannot reach, or where the support conversation is the sales conversation. In those cases the ceiling is low regardless of tooling.
Bottom line
67% is a credible figure for what a well-fed AI agent achieves, and it is a real gap over a scripted widget. But it is an average produced by companies whose support content was ready for it.
Do the 100-conversation audit before you buy anything. It tells you what your own number will look like, and it is the same work that would raise it. Then use the free conversation allowance on your hardest questions rather than your easiest — the easy ones were never the point.
Sources
