It’s 11:40pm on a Thursday. A customer in Denmark wants to return a pair of boots she bought three weeks ago, and your inbox is asleep. That gap — the questions that arrive after hours, and the ones your two-person team never gets to at 2pm either — is exactly what Tidio Lyro was built for.
Lyro is Tidio’s AI agent: software that reads an incoming message, works out whether it can answer, and either replies on the spot or hands the conversation to a human with the context attached. It’s not a decision-tree chatbot dressed up with a nicer interface. It generates answers.
What Tidio Lyro Actually Does
Lyro lives inside the Tidio platform, so it sits alongside your inbox rather than replacing it. The core mechanics are straightforward:
- Trained on your content. Help centre articles, FAQ pages, product descriptions, PDFs and past conversation logs get indexed, and Lyro answers from those sources.
- Pulls live order data. Connect Shopify, WooCommerce or similar and it can look up where a parcel is instead of pasting a generic shipping policy.
- Replies in the customer’s language. Someone writing in Portuguese gets a Portuguese answer, without you translating anything.
- Escalates with a transcript. When it hands over, the human agent sees the whole exchange, not a blank reply box.
- Billed per resolution. You pay for conversations Lyro closes without a human, not for seats.
It is not the old Tidio chatbot
The older Tidio flows followed a script: press 1 for returns, press 2 for shipping. Useful, limited, and painful to maintain once you have forty branches. Lyro handles the question you never wrote a flow for — “do these run narrow if I’ve got wide feet?” — because it’s reasoning over your content rather than following a path. The trade-off is that you can’t preview every possible answer, so testing matters more than configuration.
Setup Takes an Afternoon, Tuning Takes Weeks
Most small teams can get a working version live in a few hours. The first pass looks like this:
- Connect your store platform before anything else, so order lookups work from day one.
- Feed it your 10 or 15 strongest help articles rather than all 200. Thin, outdated pages are worse than no pages, because Lyro will confidently repeat whatever wrong thing they say.
- Write the tone instructions by hand: “Warm but brief. No emoji. Never promise a specific delivery date. Never apologise on behalf of the courier.” Vague instructions produce bland, over-apologetic replies.
- Set escalation rules explicitly. Anything involving chargebacks, refunds above £100, or a second message from the same customer should route straight to a person.
- Test it against last month’s real transcripts, not questions you invented. The awkward ones live in your inbox history.
Then leave it alone for a week and read the transcripts properly. This is the unglamorous part, and it’s where most rollouts either succeed or quietly get switched off. There’s a useful hands-on look at how the AI agent performs in day-to-day use if you want a second opinion on what that first week feels like.
Where Lyro Shines, and Where It Doesn’t
The sweet spot
Order status, delivery windows, return windows, sizing, stock checks, password resets, “do you ship to Norway”. High-volume, repetitive, low-stakes. In a store doing 2,000 conversations a month, that bucket is often more than half the queue, and most of it arrives outside working hours.
Where humans still win
A customer who writes four paragraphs about a damaged delivery that ruined a birthday isn’t asking for an FAQ answer. Neither is someone disputing a charge, or someone with three separate problems in one message. Lyro tends to answer the first identifiable question and miss the rest, which reads as dismissive.
Tidio’s marketing quotes deflection figures north of 60%. Treat that as a ceiling, not a baseline. Stores with tight, well-written help content often land somewhere between 30% and 50% resolution. Stores with a neglected knowledge base land much lower, and a small number discover Lyro answering fluently and incorrectly, which is worse than no answer at all.
The Pricing Maths That Matters
Lyro is an add-on, not part of every Tidio plan. As of writing, entry pricing sits near $39 per month for a pool of around 50 AI-resolved conversations, with larger buckets above that. Check the current numbers yourself, since these tiers move.
The comparison that matters isn’t Lyro versus nothing. It’s Lyro versus the marginal hour of a support agent. If a human-handled chat costs you roughly £5 to £8 in time, and an AI resolution costs well under £1, the maths works — but only if the resolution rate holds up. If six in ten Lyro conversations get escalated anyway, you’re paying for the AI and the agent, and you’ve added a step to the customer’s journey. Watch the escalation rate before you scale the plan up.
Four Numbers to Check Every Monday
- Resolution rate. Conversations Lyro closed with no human touching them.
- Escalation reasons. Not the count — the reasons. Ten escalations about one product page means the page needs rewriting.
- Repeat contact within 48 hours. The metric people skip. A chat marked resolved that sparks a follow-up the next day was never resolved.
- CSAT split. Satisfaction for AI-resolved chats versus human-resolved. If the gap is wide, your tone instructions need work.
Tone Instructions Are the Whole Game
Lyro writes the way you tell it to write, and the instructions field gets less attention than it deserves. Short, concrete rules beat long paragraphs of personality notes. “Keep replies under 60 words. Ask one question at a time. If you don’t know, say so and offer to pass this to a colleague” will outperform three sentences about being friendly and helpful.
The same principle applies to any conversational AI: the quality of the output tracks the quality of the brief. If you’ve ever spent time learning how to shape a conversational assistant’s replies with better prompts, you already know the drill. Give examples of good and bad responses inside the instructions, and Lyro will match the good ones far more often.
A Realistic 30-Day Rollout
Week one. Connect your store, load a dozen help articles, write tone rules, set three hard escalation triggers. Run Lyro in draft mode only — it suggests replies, your team sends them. You learn what it gets wrong without any customer seeing it.
Week two. Turn it live for a narrow slice: shipping and order status first, nothing else. Read every transcript at the end of the day.
Week three. Add returns and sizing. Rewrite any help article that produced a shaky answer. By now you’ll have a list of phrases that trip it up, and you can add those to the escalation rules.
Week four. Widen coverage, then compare CSAT for AI-resolved versus human-resolved chats. If they’re close, raise the resolution cap on your plan. If they’re not, the gap is almost always an information problem, not a model problem.
The teams that get the most out of Tidio Lyro tend to be the ones who treated it as a new colleague with a lot of enthusiasm and no context — someone who needs a decent handbook, clear boundaries about what to pass on, and a manager who reads their work for the first month. Skip that, and you’ll get an agent that answers fast, confidently, and sometimes completely wrong, which is a harder problem to fix than an unanswered message at 11:40pm.