Monday morning, 340 unread messages in the queue, and Priya has two weeks to prove the pilot was worth the licence fee. Her goal isn’t a showpiece. It’s one bot, answering one question well, so that order-status tickets stop swallowing her team’s mornings.
That’s the frame for this guide. Not a feature tour, but the order of operations worth following when you sit down to build something real in LivePerson AI and want it live before your stakeholders lose interest. Everything here assumes you have access to the platform and a queue of conversations to work with.
If you want the bigger picture first, it helps to understand how LivePerson AI is making customer conversations smarter at scale. Then come back and start small.
Before you open Conversation Builder, make three decisions
Skipping this step is the single most common reason pilots stall. LivePerson’s tooling is flexible enough that you can spend a month building something nobody needed.
1. Pick one intent that is high volume and low risk
Order status. Appointment rescheduling. Password resets. Store opening hours. Anything where a wrong answer costs a follow-up message, not a refund. Leave cancellations, billing disputes, and anything a customer might screenshot for a lawyer in version one.
2. Write down the number that means “working”
Pick two metrics and pin them to a date. A realistic first-month pair for a mid-size retail support team: containment of 35 to 45 percent on the chosen intent, and CSAT within 0.3 points of your human baseline. Anything vaguer and you’ll be arguing about feelings in week four.
3. Name the person who owns it after launch
Bots rot. Intents drift, product names change, a competitor launches a new pricing page and suddenly everyone asks about it. Thirty minutes a day, one named owner, written into their job description. Not a committee.
Step 1: Read 200 real conversations before you design anything
Pull transcripts from the last 90 days and tag them by hand. It’s tedious and it’s the highest-value hour you’ll spend. Customers don’t phrase things the way your help centre does. Nobody writes “I would like to check my shipment status.” They write “where’s my stuff” and “my package has been sitting in Ohio since Tuesday.”
A typical retail export of 200 chats might break down like this:
- Order tracking (71 chats): “where is it”, “tracking hasn’t updated”, “says delivered but nothing arrived”
- Returns (38 chats): “how do I send this back”, “wrong size”, “refund timing”
- Account access (29 chats): password resets, locked accounts, email changes
- Product questions (34 chats): sizing, materials, stock
- Everything else (28 chats): a genuine long tail you should not try to automate
Build the intent around the actual language, then load the canonical answers into the Knowledge Base so your agent-facing tools and the bot pull from the same source. Note the long tail. That 14 percent is where human agents earn their keep, and it’s also your best source of ideas for the next intent.
Step 2: Build the fallback before the happy path
Counterintuitive, but do it first. The fallback is what happens when the bot has no idea what someone means, and most first attempts handle it badly with a polite dead end: “Sorry, I didn’t understand that. Please rephrase.” That reply, repeated twice, is where containment rates go to die.
Set a two-strike rule. The first miss gets a clarifying question with a bit of help: “I can pull that up. Is this about an order you’ve already placed, or something you’re thinking of buying?” The second miss triggers a handoff with no further prompting. Some teams add a third attempt with quick-reply buttons, but two strikes plus a fast route to a human beats three strikes and a frustrated customer.
Step 3: Write replies that sound like your best agent on their best day
Read your draft responses out loud. If you wouldn’t say it to someone standing at a counter, rewrite it.
Weak: “Your request has been received. An agent will assist you shortly.”
Better: “Got it. I can see the order, it’s with the courier and due Thursday. Want me to text you if that changes?”
Three rules that do most of the work. Confirm what you understood before answering, so a wrong match becomes obvious in the transcript. Keep replies under about 40 words where possible, because long walls of text get skimmed. And never apologise twice in a row; one acknowledgment then a next step.
Step 4: Wire up the handoff while it’s still ugly
Connect the human route on day one, before you optimise a single response. In LivePerson this means thinking about which skill or queue the conversation lands in, whether the agent sees the full transcript, and whether context carries across so the customer never repeats an order number.
Set escalation triggers explicitly. The ones that catch most of the real cases:
- Two failed intent matches in a row
- Sentiment drop detected mid-conversation
- Keywords like “cancel”, “complaint”, “manager”, or “refund”
- High-value order or account flags passed in from your CRM
- Any conversation running past a set number of turns without resolution
Then test the handoff with a colleague who knows nothing about the build, and watch whether they can pick up the thread within ten seconds.
Step 5: Read transcripts for 30 minutes a day for two weeks
This is the unglamorous part that separates a bot that improves from one that quietly degrades. Log every mismatch in a running list. You’ll find patterns fast: a phrase the intent doesn’t catch, a response that answers a slightly different question, a quick reply nobody taps.
One team working on order tracking moved intent match rate from 62 percent to 89 percent in eleven days doing exactly this, mostly by adding eight customer phrasings to a single intent and shortening two responses. Containment followed, from 41 percent to 68 percent, with the human queue absorbing the hard cases more happily because the easy ones had gone.
Where teams stall in week three
Momentum drops when the pilot works, because success creates the next problem: scope. Someone senior asks why the bot can’t handle refunds yet, and the roadmap balloons.
Two habits help. Keep a visible list of intents ranked by volume and difficulty, so expansion is a decision rather than a demand. And spend an hour a week looking outside your own category, because the ideas that move the needle often come from adjacent fields. The sessions on how AI can engineer nature’s comeback at Disrupt 2026 are a good example of a topic with nothing to do with retail support that will still change how you think about modelling complex, messy systems.
Whether you’re running a two-person support team or a contact centre with twelve queues, the fix is the same: ship the next intent only when the current one holds above your target for two consecutive weeks.
What to measure once the bot is boring
Containment gets all the attention, but it’s easy to game and easy to misread. A bot can contain a conversation by frustrating someone into leaving. Watch these instead once the pilot settles:
- Repeat contact rate: did the same customer come back within 48 hours with the same question? This is the honest score.
- Containment by intent: an overall figure hides a strong performer covering for a weak one.
- CSAT split between bot-only and bot-to-human journeys: if the second is much higher, your handoff timing is off.
- Time to first meaningful answer: not handle time, but how long until the customer learns something useful.
Repeat contact rate is the number to put in front of your leadership, because it’s the one that tracks whether customers actually got what they came for. A bot that closes 70 percent of conversations but sends a third of those customers back the next day hasn’t solved anything. A bot that contains 45 percent and leaves those people genuinely done has bought your human team the space to handle the conversations that need them, which was the entire point.