Tidio Lyro

Tidio Lyro: What The Agent Handles

• A published resolution rate: 67 percent on average, stated rather than implied.
• Grounded in your content — answers built from your knowledge base and guidance, not general knowledge.
• Smart Actions with MCP for order updates, lead qualification and product recommendations.
• Configurable handoff rules, with real-time monitoring and the ability to step in.
• No migration required — connects to Zendesk, Intercom, Salesforce and 20-plus platforms.

Tidio Lyro: An Agent Inside A Live Chat Product

Lyro is positioned as the AI agent customers want to talk to — resolving questions with AI that learns from your data and speaks like your best agents.

The context matters: this sits inside Tidio, an established live chat and helpdesk product, rather than arriving as a standalone agent looking for a queue to join.

That lineage produces a specific advantage. A support agent is only as good as its escape hatch, and a company that already built the live chat, the ticketing and the agent console has somewhere for a conversation to go when automation runs out.

Standalone agent products frequently have the harder half of the problem — the human handoff — as an integration rather than a native capability.

The stated positioning reinforces it: save on automation without switching platforms, no migration, connecting to your current stack. For most buyers that is the decisive practical point, because ripping out a working helpdesk to add automation is a project nobody wants to run.

A support agent is only as good as its escape hatch. Products built around an existing helpdesk have somewhere for a conversation to go when automation runs out.

Resolution Rate And How It Is Counted

The number stated is unusually specific: Lyro boosts your resolution rate to 67 percent on average, often higher.

Publishing an average rather than a best case is worth crediting, and it is also worth understanding precisely what such a figure measures.

Resolution rate is the share of conversations closed without a human. Its weakness across the whole category is that a conversation ending is not the same as a customer being helped — an abandoned chat and a solved problem both stop, and the naive metric counts them identically.

This is why resolution rates near 100 percent should be read as measurement artefacts rather than achievements.

Sixty-seven percent is credible precisely because it is modest. Roughly a third of a real support queue involves something unusual, emotional or genuinely complex, and any honest system escalates it. The number to ask a vendor for is satisfaction on resolved conversations alongside the rate, because those two together are the only pair that means anything. Enterprise platforms like Forethought report the same metric with the same caveats attached.

What Happens When The Agent Cannot Answer?

The escalation design is described as flexible handoff and escalation rules, with real-time conversation monitoring and the ability to jump in at any point — specifically to chat personally with high-value customers or discuss sensitive topics.

That framing is more sophisticated than a simple confidence threshold, and it reflects how support actually works.

Escalation is usually treated as a failure state: the agent could not cope, so a person takes over. But some conversations should never have gone to automation regardless of whether it could handle them — a complaint from your largest account, a bereavement, a legal threat, a safety issue.

Routing by customer value and topic rather than only by confidence is the difference between an automation policy and a support policy.

The other half is monitoring. Being able to watch conversations in real time and step in mid-exchange means automation is supervised rather than delegated, which is what makes support teams willing to trust it. The alert when a customer needs immediate human help is the mechanism that decides whether a frustrated customer waits two minutes or twenty, and that gap is where brand damage happens.

• Ask for satisfaction on resolved conversations, not just the resolution rate.
• Route by customer value and topic, not only by model confidence.
• Load knowledge from real transcripts, not from your public FAQ page.
• Test what it does when the knowledge base does not cover a question.
• Review the escalated conversations weekly and add what was missing.

Training On Your Own Help Content

The grounding claim is the architectural centre of the product: Lyro works only with the content and guidance you provide — always on script, always on brand — with strong guardrails meaning every response is accurate and helpful.

The first half of that is a real and important design decision. The second half is stronger than any grounding can guarantee, and it is worth separating them.

Restricting an agent to your own content genuinely does reduce fabrication, because the model is summarising retrieved material rather than generating from general knowledge. That is the correct architecture for customer service, where a confidently invented refund policy is worse than no answer at all.

What it cannot do is make every response accurate. Retrieval can surface the wrong passage, your documentation can be out of date, and a plausible answer assembled from two unrelated articles reads exactly like a correct one.

The practical consequence is that quality tracks your knowledge base, which is the part most companies underestimate. An agent grounded in a neglected help centre will confidently repeat two-year-old policies. The strongest preparation is loading real support transcripts rather than the FAQ page, because transcripts contain the answers your team actually gives. General-purpose builders such as Voiceflow give more control over that logic at the cost of more configuration.

Where Automation Costs More Than It Saves

The onboard, develop, promote framing is the most useful idea on the page: add Lyro and get resolutions from day one, review conversations and add missing knowledge, then assign more complex tasks as it proves itself.

Treating an agent like a new hire is exactly right, and it inverts how most deployments go wrong.

The common failure is switching automation on across the entire queue on day one, discovering it handles the hard third badly, and concluding the category does not work. Nobody would give a new support hire every ticket type in week one, and the same logic applies.

Starting narrow, reviewing what it got wrong and widening its scope deliberately produces both better performance and a support team that trusts it.

The place automation genuinely costs more than it saves is worth naming plainly: a customer with an urgent problem, no visible route to a person, and a system confidently answering a question they did not ask. That customer does not file a complaint — they leave, and the resolution rate looks excellent while it happens. Smart Actions, now with MCP support, are what move the agent from answering to doing, and doing is where real resolution comes from. Multi-channel commerce tools like Spur apply the same principle across messaging apps rather than website chat.

Deploying Support Automation Carefully

The measure of a deployment is not what it automates. It is what happens to the customers it could not help.

This order protects them.

1. Load knowledge from real transcripts before touching the FAQ page.
2. Start on one narrow, high-volume question type.
3. Set escalation by customer value and topic as well as confidence.
4. Monitor live conversations for the first fortnight rather than reading reports.
5. Read every escalated conversation and add the knowledge that was missing.
6. Widen scope only after satisfaction on resolved conversations holds up.

Tidio Lyro: What It Does And Where It Stops